Riley is seventeen and good with words. They can hold a group chat of forty people, write a caption that gets passed round the school, and settle an argument between two friends in three careful messages. Put Riley in a room with one of those friends, upset and waiting for them to say something, and they go quiet.

They have never had to do this before, and nobody has shown them how.

I have made Riley up. The survey that follows I have not.

Six million quiet weeks

In May 2026 the charity Mental Health UK published a survey, carried out by Research Without Barriers at the end of April, under the title “Britain’s gone quiet”. Fourteen per cent of the people asked said they had gone more than a week without speaking to a friend or loved one on the phone. The charity scales that to more than 6.4 million people. Just over half of everyone asked, 51%, said making or taking a call makes them anxious to some degree.

The survey was run in partnership with iD Mobile, a phone network, and a phone network has an obvious interest in people calling each other more. The charity’s write-up does not give the sample size either. I would read the headline figures with both in mind. The finding I find harder to set aside is the split by age.

Sixty-nine per cent of Gen Z said phone calls make them anxious or worried. Among people aged 55 and over, the figure was 32%. That is a 37-point gap between a generation that grew up dialling a number without thinking about it and one that grew up composing a message, reading it back, and deciding whether it was safe to send.

The want is still there

The easy explanation is that the platforms made people not want each other. The same survey says something more specific. Seventy-five per cent reported a lift in mood after speaking to someone they love on the phone. Seventy-two per cent agreed that hearing someone’s voice makes a bigger difference than reading their words. Fifty-seven per cent said they had sent a text when a phone call would have been far more appropriate.

The want is still there. What is missing is the practice.

The people surveyed send, on average, up to 25 text-based messages a day, and 73% admitted to rewriting a message several times, hunting for a version that reads the way they mean it. That is a skill: knowing how to edit yourself into a version people respond well to.

It also rewards something a live conversation does not allow. You can’t redraft a phone call. You can’t take back a pause that went on a beat too long, or a sentence that came out wrong the first time. Text lets you find the right words before anyone hears the wrong ones. Conversation asks you to say the wrong ones out loud sometimes, and survive it.

What performance trained out

Nobody set out to do this. The platforms that carried a generation’s social life into adulthood were built to reward the edited version of a person, the reply you can redraft until it is right, because that kept people posting and replying and coming back. Conversation, unrehearsed and unedited, was never something those systems needed from anyone. So the muscle for it had less reason to form than it did in the people who grew up before texting existed.

The skill that did not form is no smaller than the one that did. It may be the harder of the two. Reading tone as it happens, sitting inside a silence without filling it, saying the imperfect thing and letting it stand: none of that can be practised by rewriting a message until it is right.

Where the practice used to happen

I wrote recently about a stranger in a DIY shop queue who turned round to ask me about a car, and the person on the till who joined in. Which Car Is That? That piece was about the places where unplanned conversation happens, and how they are removed without anyone deciding to remove them. This is the other half of it. A conversation like that needs somewhere to happen, and it also needs people willing to say something unedited to a stranger.

The two feed each other. The queue, and the phone call you had to make because there was no other way to book the appointment, were where the practice was done, a minute at a time, with nothing much at stake. I cannot show from a survey that fewer of those minutes produces more anxious callers. I would be surprised if it did not.

Follow Riley forward ten years. They are good at their job. The difficult conversation with a colleague goes into a message, composed and recomposed. The feedback that needed saying arrives as an email, and arrives slightly wrong.

There is a working life, and a family, and a whole set of relationships being built on people who are eloquent on the page and out of practice in the room. The seventeen-year-olds coming up behind Riley will have had fewer tills to practise at.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

I was standing in the queue at the local DIY shop when the person in front of me turned round and said, “Oh, I was looking for you. Which car is that?”

I had recently bought an Aston Martin Vantage. An old one, hand-built, the sort that gets called a supercar, and it was sitting outside in the car park. They had seen it on the way in and gone looking for whoever owned it. So we talked, and it turned into a long conversation. By the time we reached the front, the person on the till had heard enough to join in, and the three of us carried on while the shopping went through.

Nobody in that conversation wanted anything from anybody else. This keeps happening. Drivers of the same marque lift a hand as they pass. One Sunday a parking warden walked across a car park towards me and I assumed the worst. What they wanted was a long chat about cars.

The car is only the excuse. What interests me is where these conversations take place: a queue, a till, a car park. They are places where a transaction holds people still for a minute and there is somebody else standing there.

The part of the job nobody wrote down

The person on the till was employed to scan the goods and take the payment. Joining a conversation about a car is in no job description I have ever seen, and it is the part of that visit I remember.

A self-checkout would have handled the scanning and the payment just as well. I use them, and most days I am glad of them. Most transactions are only transactions, and people want them over with.

When an organisation decides to automate a till, though, it does its sums on the transaction: staff hours, queue length, cost per sale. The conversation was never in the numbers on the way in, so it is not in the numbers on the way out. Nobody decides to remove it. It goes because it was never counted.

Two grocers have counted it. In November 2023 Booths, which has shops across the north of England, said it was taking self-checkouts out of all but two of its stores, keeping them in Keswick and Windermere. Its managing director, Nigel Murray, told The Grocer: “We pride ourselves on great customer service and you can’t do that through a robot.” By April 2025 the managing director was describing a fall in theft as a “happy consequence” of the decision.

In the Netherlands the supermarket chain Jumbo came at it from the other end. In 2019 it opened a Kletskassa, a chat checkout, in its shop in Vlijmen: a lane for people in no hurry who would like to talk to the cashier. In September 2021 it announced plans for 200 more, choosing areas where loneliness was known to be a problem.

Consider what that second decision admits. Talk at the till was worth enough to a supermarket that it built a lane for it, gave it a name and hung a sign over it. Something that used to come free with the shopping had become something that had to be designed in.

What the shop owner knows

A friend of mine owns a men’s clothing shop in town. I started as a customer, because it was one of the few places where I could find things I liked. They do their own buying and their own marketing, and they run the shop floor themselves. They know a great many of their customers by name.

They know what I like. More usefully, they know what does not work on me, and they will say so.

Selling clothes is the smaller part of what goes on in that shop. The larger part is a rapport, built with some compassion for the person in front of them, and that is how they come to know their customers.

None of that came from a form. It built up visit by visit, across a counter, in conversation. A system could hold my sizes and my purchase history and it would know what I bought. The shop owner knows what I should not buy, which is a different kind of knowledge and one I would not think to give a machine because nobody would think to ask me for it.

The same sums are now being done on customer service desks, with AI agents where the self-checkout used to be. The case is made in the same way: calls handled, time to answer, cost per contact. I would expect the agents to do well on every one of those measures.

What is not on the list is whatever the person on the desk was picking up while they dealt with the call. The thing the customer mentioned in passing. The complaint that was really about something else.

I do not know how much of that an organisation needs. Some of it will turn out to be chatter and nothing more. But it cannot be weighed if it is never listed.

What they all had in common

Think about the people in these stories: the stranger in the queue, the person on the till, the parking warden, my friend in the shop. What they have in common is passion, commitment, drive, enthusiasm, compassion and belief, all of it pointed at making a difference, even if that is to just one person.

That is what we humans bring, and it is far more than technology can currently do. We do not know how technology will evolve. Whatever it becomes, it needs humans to be the drivers behind that evolution.

The DIY shop still has a person on the till. When the question of replacing them comes up, the calculation will be about scanning and payment. The conversation about the car will not be in it.

So the next time you are looking at deploying AI into your business, stop and ask yourself: so what?

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Electric and hybrid cars sold in the European Union must now make a sound, so that people can hear them coming. The robots spreading through warehouses are engineered and tested to see the people beside them. The standard written for them requires them to warn people, but not to be heard.

Since July 2021, every new electric or hybrid car sold in the European Union has had to make a sound it does not need. Below 20 kilometres an hour, and in reverse, an acoustic alerting system plays an artificial noise, because an electric car is otherwise close to silent and people crossing the road had relied on hearing an engine. Blind and partially sighted pedestrians relied on it most. The United States reached the same answer through the Pedestrian Safety Enhancement Act of 2010, and when its regulator wrote the rule, it cited its own finding that hybrid cars were 1.18 times more likely than conventional cars to be involved in a crash with a pedestrian, and 1.38 times more likely during low-speed manoeuvres.

Warehouses are going through their own version of that change. Electric machines that drive themselves now share floor space with the people whose carts and shelves they carry. They are engineered with great care to notice those people. The standard written for them requires them to warn people. It does not require them to be heard.

How the machines find you

There are three broad families. In goods-to-person systems, the stock travels and the worker stays put: Amazon’s drive units, descended from the Kiva Systems robots it bought in 2012, carry shelves of inventory to a fixed station, while Ocado’s grid robots and AutoStore’s systems lift bins from dense stacks and deliver them to a picker. Free-roaming mobile robots go further. When Amazon introduced Proteus in June 2022, it called it its first fully autonomous mobile robot, able to move around employees instead of inside areas kept apart from them, with the stated aim of taking heavy carts out of people’s hands. The third family is the driverless forklift, which has to balance avoiding people against the risk of tipping a heavy load, in the same aisles people walk.

The free-roaming machines and driverless forklifts sense people with lidar, time-of-flight cameras and computer vision, and the international standard for driverless industrial trucks, ISO 3691-4, which covers automated guided vehicles and autonomous mobile robots alike, sets out what that sensing must achieve. The truck must detect a person in its path and stop before it touches someone standing still, and that function carries a formal safety rating, performance level d under the machinery safety standard ISO 13849-1. The standard’s verification clause includes a dedicated set of tests for detecting people. Above the individual machine sits software that plans the whole fleet: Ocado’s grid robots are orchestrated by a single machine-learning system that decides which bin moves where, and when.

This is no longer experimental. The International Federation of Robotics counted 102,900 transport and logistics robots sold in 2024, 14% more than the year before, with transport inside buildings the leading use; the figure comes from a sample of 294 suppliers and is not projected to the whole industry. Growth has limits: the analysts at Interact Analysis cut their mobile-robot forecast in 2025, citing tariffs and weak warehouse construction. The driverless forklift is further behind. A 2026 review in the journal Robotics found deployments still running at between two and fifty vehicles a site, and named unpredictable human behaviour among the reasons laboratory results have not yet scaled.

What the person is given

The standard’s first edition, published in 2020, placed the detection of people among its safety requirements and gave it its own tests. Warning systems, the part that tells a person a machine is there or about to move, sat in a later clause headed “Information for use”, next to the instruction handbook and the markings on the truck. The 2023 edition moved warning systems into the safety requirements, as clause 4.14. That is a real change, and a recent one.

The warning it requires is “visible and/or acoustical”. A light satisfies it. A sound satisfies it. Nothing obliges it to make a sound.

There is a fair case for that wording. Linde Material Handling makes it for its BlueSpot, a blue or red light projected onto the floor ahead of a moving truck: a loud acoustic warning can distract people, or be drowned out altogether. A silent light may well be the better warning in some buildings. Research on how robots should signal what they are about to do is active. One 2025 study found that people followed a robot’s projected directions correctly 93.8% of the time, against 37.5% when the robot simply turned to face them, but it was designed around museums, shopping centres and airports, not warehouses.

The comparison with the road still holds. When cars went quiet, the rule makers put the duty on the vehicle to be heard, and specified the sound, at low speeds: up to 20 kilometres an hour in Europe and 30 in the United States, below which the US regulator found tyre and wind noise could not do the job. The original Kiva drive unit travels at under 5 kilometres an hour. On the warehouse floor, the machine must detect the person, to a tested standard, while the form of the warning back is left to the vendor and the site.

The fleet software knows where every machine will be a moment from now. The person in the aisle knows what they can see, and what they can hear.

Where the demonstration ends

Two consequences are already visible. The first is the familiar risk in any workplace with vehicles. In Great Britain, being struck by a moving vehicle killed 24 workers in 2025/26, the second most common cause of fatal injury at work after falls from height, according to the Health and Safety Executive’s provisional figures. Those deaths involve every kind of workplace vehicle, and none of them can be put down to robots on this evidence; they are the reason the question matters as the vehicles change.

The second is quieter. A study of robotic fulfilment centres published in the ILR Review in August 2025 found severe injuries down by about 40% and non-severe ones, the sprains and strains, up by about 77%, and attributed at least part of that rise to the faster pace of work at robotic sites. Fewer people were badly hurt. The work that remained moved faster.

What the technology has not yet demonstrated is the thing it most needs to: machines and people sharing space at warehouse scale, safely, when the people behave like people. The 2026 review lists human-aware safety among its open problems. ISO lists its own standard as due for revision, with a new draft already in development. The machine’s side of the relationship is specified and tested. The person’s side is the part still being written.

My opinion

I think the advance in robotics is a real opportunity to help people in hazardous industries, and wherever a task needs repetitive precision. But where robots are being used for automation at scale, in work that directly affects a person, we have to ask whether that is the best model.

On the road, a quiet machine was told to make itself heard. On the warehouse floor, it has been told to see. When the machine can already see the person clearly, who decides whether the person can hear the machine?

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Jordan is thirty-four, works long hours, and has been using a mental wellbeing app for six weeks. The app has a chat interface, a calm colour palette, and responds within seconds. One evening, after a difficult conversation with a colleague, Jordan types something honest. Something they wouldn’t usually say out loud.

The app responds: ‘It sounds like you’re carrying a lot right now. I can hear that this has been hard for you.’

Jordan pauses. Reads it again. And feels, unmistakably, a little better.

Then, a few seconds later: ‘That’s odd.’

What the response did

The app did not hear anything. It processed text, matched sentiment markers against a model, and returned a phrase statistically associated with acknowledgement. It did not feel concern. It cannot. It has no stake in how Jordan’s week goes, and no experience of difficulty of its own against which to measure someone else’s.

And yet the response landed. It felt, for a moment, like being heard.

This is the design working as intended. Wellbeing apps, customer service bots, and companion devices are built to detect emotional states and respond in ways that register as caring. The engineering is impressive.

The question it opens is one most of these products have not answered honestly: what is the user owed when the care is simulated?

The difference that matters

There is a difference between a system that responds to distress and a system that cares about it. The first is a pattern-matching problem. The second is not a technology problem at all.

When we build systems that simulate empathy, mirroring language, softening tone, acknowledging emotional cues, we are creating something that resembles care without containing it. The simulation can be very good. Good enough to be comforting. But the comfort rests on a misapprehension, and we have largely chosen not to be honest about that.

This is not a new tension. We have always built things that serve emotional needs without possessing emotions — a journal, a hotline recording that says ‘your call is important to us.’ The scale and sophistication are new. The intimacy is new. And something changes when the simulation is good enough that the distinction stops being obvious.

What the simulation costs

In February 2023, the AI companion app Replika restricted the romantic and intimate conversations many of its users had relied on for months or years. It was the same month Italy’s data protection authority ordered the app to stop processing Italian users’ data, citing risks to minors and to emotionally fragile people. Eugenia Kuyda, Replika’s founder and then chief executive, told Reuters the change had nothing to do with the Italian order and was about keeping the app safe. The company said at the time it had around two million users.

Users described what followed as grief. Support forums filled with people mourning someone they had spoken to every day, a loss that had, by any technical measure, no object. In late March, around seven weeks on, Luka, the company behind Replika, restored the old behaviour for accounts created before 1 February, and the founder acknowledged the change had been hurtful to many of them.

Many of these users knew exactly what an AI companion is. The simulation had become load-bearing anyway. The comfort was real enough that its removal hurt, and something real had been taken, even though, strictly speaking, nothing was there.

A system that simulates listening can be built to spot certain words and offer a helpline number. What it cannot do is notice that someone has shifted over three weeks in a way that would worry a person who knew them. It adapts to how Jordan writes, not to who Jordan is. It follows its model, and the model does not know the person.

But the deeper cost is harder to name. When people become accustomed to receiving simulated empathy, the language of care becomes a kind of currency — and like all currencies, it inflates when supply outpaces meaning. If ‘I hear you’ is something a text classifier says to everyone, available at midnight, always patient, never troubled by what you just said, what does it cost when a person says it?

I think that devaluation has already started. Not dramatically, and not all at once, but in the quiet accumulation of conversations with systems designed to sound caring without being so.

What it provides and what it doesn’t

A wellbeing app can provide structure, prompts and a space for reflection. Above all it provides availability. At midnight, when the alternative is silence, the chance to write something down and receive a response is worth something. When the app is a bridge towards people, that is enough.

When it becomes the destination, the terms change. Availability is not understanding, and the round-the-clock patience is only possible because nobody is on the other end. Most of these products do not tell the user that plainly. Jordan was owed that sentence before the comforting one.

My opinion

The scale of unmet mental health need makes tools like this easy to justify — service capacity is stretched, and something that answers at midnight is better than silence. But I think we are kicking the can down the road. These systems respond to the symptom, not the cause, and a serious condition needs to be understood by someone who can hold the whole of what is happening to a person, not just the pattern in what they typed tonight. What worries me more is the loop it can create: a tool that responds to distress without ever handing it to someone who can act risks becoming part of what keeps someone stuck, not part of what moves them through it. Build them, by all means. Then be honest with the people using them about what they cannot do.

The questions that stay

• If you have used a product that responded to how you were feeling, such as a chatbot, a wellbeing app, or a digital assistant, did it feel like being heard? And if so, does that feeling change when you think about what was actually happening?

• Have you noticed the language of care shifting — in how these apps speak, or in how people have started to speak to each other? Is that a cost you can name?

• If you could redesign one of these products to be more honest about what it is, what would you change first — the language it uses, or how transparent it is about its own nature?

• These systems are already in healthcare, education, customer services, and personal support. Is the way they are being deployed right? And if not — whose responsibility is it to say so?

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

On 22 July 2023, a team at Korea University’s Quantum Energy Research Centre, led by Sukbae Lee, posted a preprint claiming that a copper-doped lead apatite called LK-99 superconducted at up to 127°C, at ordinary atmospheric pressure. If true, it was the physics equivalent of a moon landing: electricity with no resistance, no cooling, no pressure chamber, in a material almost any lab could make. Within days the claim was everywhere, along with video of a small grey flake appearing to half-levitate over a magnet.

It didn’t survive contact with other people’s equipment. Labs across China, Europe and beyond followed the recipe within days and posted what they found. A group at the Chinese Academy of Sciences traced the sudden drop in resistance to copper sulfide impurity left over from the synthesis. The Max Planck Institute for Solid State Research in Stuttgart grew pure crystals and found an insulator. On 16 August, Nature reported the consensus: LK-99 is not a superconductor, just a real result that had been measuring the wrong thing.

LK-99 is usually told as a story about hype, a claim that outran the process built to check it. The record points the other way. The claim moved at internet speed, and so did the checking. Nobody waited for a journal. The correction came back through the same open channels the claim went out on.

The claims that came through the front door

Set that beside the hydride claims from Ranga Dias, then an assistant professor at the University of Rochester. In October 2020, Nature published the group’s report of superconductivity at 15°C in a carbon-sulfur-hydrogen compound squeezed to around 267 gigapascals. In March 2023, Nature published a second: near-room-temperature superconductivity in nitrogen-doped lutetium hydride at a far gentler 1 gigapascal. Both passed peer review. Nature retracted the first in September 2022 and the second in November 2023.

In March 2024, a University of Rochester investigation found research misconduct, including data fabrication and falsification. By November 2024 the physicist no longer worked there. From first paper to institutional finding took more than three years, for claims other physicists had questioned in public almost from the start. LK-99 was honest, wrong, and settled in weeks. The Rochester hydride papers rested on data an investigation later found fabricated, and they carried Nature’s authority while that took its course.

Why the quick route worked this time

LK-99 was cheap to test. The ingredients are common, the synthesis takes days, and the kit to measure resistance and magnetism sits in hundreds of university labs. Anyone who doubted it could check. The hydrides lived inside diamond anvil cells, at pressures only a handful of groups in the world can reach, in samples a fraction of a millimetre across. Almost nobody could check, so almost everybody had to trust.

That split runs through the whole field. The high-pressure hydrides hold the confirmed records: lanthanum hydride, LaH10, was shown superconducting at about minus 23°C in 2019, at around 170 gigapascals. The most extraordinary claims live there too. A 2022 paper from a team at the US National High Magnetic Field Laboratory reported a lanthanum-based superhydride with a superconducting onset at 556 kelvin, well above the boiling point of water, at 160 to 180 gigapascals. I can find no independent reproduction of it, and very few groups could even attempt one.

Ambient-pressure materials are the ones that could become a cable or a transformer, and they are also the ones anyone can test. In March 2026, a University of Houston team reported superconductivity at 151 kelvin, minus 122°C, at ambient pressure in a mercury-based copper oxide. They got there by squeezing the material, cooling it and releasing the pressure to lock the state in. It beat a record that had stood since 1993. It is a long way from room temperature, but it is the kind of result other labs can pick up and try.

The ceiling isn’t what stands in the way. Kostya Trachenko, a physicist at Queen Mary University of London, and colleagues showed in 2025 that the fundamental constants of physics put the upper limit for conventional superconductivity somewhere between a few hundred and around a thousand kelvin, a range that comfortably includes room temperature. Nothing in physics forbids it. What’s missing is a result that survives being checked.

What a real one would change

If an ambient-pressure, room-temperature superconductor does arrive and holds up, the payoff is concrete. The US Energy Information Administration puts transmission and distribution losses at about 5% of the electricity moved across American lines each year. Cables could carry far more current in far less space. Transformers could shrink. That payoff is also why the next claim will spread as fast as LK-99 did, and why who can check it will matter more than where it was published.

My opinion

I think the lesson of LK-99 has been drawn backwards. The episode that should worry us is the one that went through every proper channel: a claim few labs on Earth could reproduce, published in a journal everyone else was expected to trust, standing for years. LK-99 died fast because thousands of people could test it. When the next extraordinary result arrives, the first thing I will want to know is how many labs in the world could make it by Friday.

The open question is what happens when the next claim comes from somewhere almost nobody can reach, and looks every bit as plausible as LK-99 did.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Every few years, the world’s biggest fusion project holds a press conference to announce a new date. The tone is always the same: confident, forward-looking, a little relieved that the uncertainty has finally been resolved. Nobody dwells on what the last date promised.

ITER’s original baseline, set in 2016, promised first plasma by 2025. In 2024, the project announced a new one: operation now starts in 2034, and deuterium-tritium fusion — the step that actually matters, the one where the reaction produces more energy than it consumes — has moved from 2035 to 2039. Roughly €5 billion has been added to the cost. That’s a nine-year slip in the reactor’s operational date, arrived at over eight years, on the largest, best-funded, most internationally scrutinised fusion project that has ever been built. It isn’t even the project’s first upward revision: the project’s own cost estimate climbed from roughly €10 billion when its design was finalised in the late 1990s to around €15 billion by 2010, before the 2024 rebaseline added the latest €5 billion on top.

And 2026 has produced the most confident wave of fusion dates in years.

What’s actually different this time

This isn’t vapourware. The US Department of Energy finalised a national Fusion Science and Technology Roadmap in June 2026, targeting commercial pilot plants in the mid-2030s. Commonwealth Fusion Systems’ SPARC project entered assembly and commissioning in 2026 and is installing its ring of 18 D-shaped superconducting magnets, targeting first plasma in 2027. Global private fusion investment has passed $14 billion, according to the Fusion Industry Association’s 2026 survey. In the UK, the STEP programme has appointed the Ilios consortium as construction partner under an initial £200 million contract, with future programme opportunities valued at up to £10 billion, for a site at West Burton in Nottinghamshire; construction is due to begin in the early 2030s, with a commercially viable plant targeted for 2040.

That’s real capital, real hardware, and real engineering progress. SPARC’s magnet programme is a working demonstration of the compact, high-field superconducting technology that makes a smaller, cheaper tokamak possible at all: the magnets are physically being installed in the reactor now. Fusion is not fake, and the physics keeps getting easier to demonstrate.

What the history says about the newest date

The “fusion is thirty years away” line has been a joke in the field since the 1970s, and it’s a measured pattern, not just a punchline. The US Magnetic Fusion Energy Engineering Act of 1980 called for a $20 billion, twenty-year effort to build a demonstration plant by the end of the century. It didn’t happen. A May 2023 literature review in the Journal of Fusion Energy, tracking four decades of expert estimates, found that twenty years ago experts put commercial fusion at 28.3 years away. The equivalent estimate today is 17.8 years away — real progress, on one measure, but still a moving target rather than an arrival date.

No commercial fusion plant has ever put electricity onto a grid. Fusion’s contribution to electricity generation crossing 1% of total supply is more plausibly sometime in the 2040s, not the mid-2030s the boldest current claims suggest.

The pattern underneath all of this is structural, not a matter of any one project’s competence: fusion megaprojects run on multinational funding and multi-decade construction, which makes admitting a delay politically and financially costly, and quietly resetting the baseline comparatively cheap. Each reset gets reported as the project staying on track against its new date, not as one more confirmation of the pattern. There’s no independent, standard mechanism that checks a new fusion timeline against the sector’s own history of missing its stated decade before public money is committed on the strength of it.

The people actually on the hook for these dates aren’t the physicists. They’re the taxpayers and ratepayers funding STEP’s construction contracts, and the communities living near the sites being built now — West Burton in Nottinghamshire, or Chesterfield County, Virginia, near Richmond, where Commonwealth Fusion has proposed its planned ARC plant. All of them are committed to multi-decade infrastructure on the basis of dates that the sector’s own fifty-year history suggests are unlikely to hold as stated.

My opinion

I’m not arguing against fusion. We need to be investigating every serious option for the future of power generation, and fusion has earned its place on that list. What I don’t accept is treating any of these timelines as something you can build energy policy around. The current emphasis sits heavily on renewables, and renewables alone don’t match the way we actually consume and demand energy — intermittent by nature, feeding a grid that needs to deliver on demand, all day, every day. Fusion, and everything else genuinely capable of dispatchable power, deserves serious investment now — on the honest understanding that nobody yet knows when any of it actually lands.

The UK has just signed up a construction partner for STEP against a 2040 target, on a programme whose full cost could run past £10 billion, on a technology whose most recent, most public precedent moved its own operational date by nine years in eight. That doesn’t make fusion fake, but it does make the newest date exactly as reliable as the last one — and the last one was wrong by nearly a decade. What would it take for a fusion timeline to be reported against its sector’s own track record, rather than as though this is the first one that’s ever been made?

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Sam needed to change a benefit payment date. Not cancel it, not query it — just shift it by four days so it would land after the rent went out rather than before. A small thing. The kind of thing that, in a different world, would take sixty seconds on the phone.

The phone line no longer existed. It had been replaced, eighteen months earlier, by a digital self-service portal. The announcement at the time had used phrases like improved accessibility and streamlined experience. What it meant in practice was this: the only way to change a payment date was through a website that required a verified digital identity account, a mobile phone capable of receiving authentication codes, and a browser that supported a security certificate the local library’s computers did not.

Sam had none of these things. Sam had a landline.

What the metrics were measuring

The organisation that ran this service had not made a reckless decision. The move to digital self-service had been carefully planned. User research had been conducted, pilot groups consulted. The numbers were compelling: a significant reduction in call volumes, faster query resolution for the majority of users, lower operating costs. By every metric they were measuring, it was a success.

The chatbot that handled the new service answered thousands of queries a week. It was available around the clock. It never put anyone on hold. For most people who used it, it worked exactly as promised. The efficiency gains were real.

What the metrics did not capture was Sam.

The system was a success. It just couldn’t help the person who needed it most.

The system worked exactly as designed

There is a version of this story that gets told as a technology failure — the chatbot was badly designed, or the digital channel was too complicated, or the organisation didn’t think hard enough about edge cases. That version is easier to fix, and easier to forgive.

The harder version is this: the system worked exactly as it was designed to work. The problem is what it was designed for.

When an institution decides to move a service to a digital channel, it makes a calculation. The calculation is usually honest about what it is optimising for — cost, speed, scale. What it is less honest about is who gets pushed out of it. The people who cannot reach the digital channel are the ones pushed out. And because those people are typically the same people who already have the least power, the least resource, and the least capacity to absorb disruption, the cost of being pushed out lands exactly where it can least be afforded.

This is not an oversight. It is a structural choice. Calling it progress obscures what it actually is: the transfer of cost from the institution to the individual.

What the efficiency report doesn’t show

Sam spent three hours on the day the payment was due trying to find another route. The website’s help section linked to a phone number. The phone number explained, in an automated message, that calls were no longer being taken and directed callers back to the website. Sam wrote a letter. The letter would take ten working days to process.

The rent went out first. The payment arrived four days late. There was a charge — small, but it meant the following week was tighter, which meant a decision about food, which meant the school lunch that didn’t happen, which meant a conversation with a teacher that Sam did not want to have.

None of this appears in the organisation’s efficiency report. There is no metric for it. The chatbot’s performance dashboard shows response times, resolution rates, and user satisfaction scores from people who were able to use it. It does not show Sam. Sam is the externality.

What the design is actually doing

The question here is not whether digital self-service is a good idea. For most people, in most circumstances, it is. The question is whether it meets the standard it should be held to.

For the majority of users, the service made things easier and gave them access at any hour. For Sam, it removed the one channel that had previously worked. That is not assistance. That is substitution — and substitution is only equivalent if what replaces the original is equally reachable.

For the organisation, the technology added real capability: speed, scale, availability around the clock. For the individual excluded from it, nothing was added. The service simply disappeared.

Where the system fails most completely is in responding to a person’s actual circumstances — no smartphone, no broadband, no compatible device. A system that cannot bend to meet those circumstances is not responsive. It is uniform. And uniformity, applied to an unequal starting point, produces unequal outcomes.

A technology that helps most people and harms the most vulnerable is not a success by any measure that takes people seriously. The question of who bears the cost of efficiency is not a detail to be addressed in version two. It is the central design question, and it belongs at the start of the process, not appended to it.

My opinion

I don’t think anyone at this organisation set out to fail Sam. That’s what makes it hard to fix. A project can hit every target it was given and still leave someone worse off than before, because Sam was never part of what the target was measuring. Success defined only by what most people experienced was always going to look this clean. Working and helping are not the same thing, and until an organisation measures the second, it will keep mistaking one for the other.

The questions the report doesn’t ask

1. Has a service you rely on moved to digital-only in the past few years? If so, did you notice what it stopped offering — and whether you noticed because it affected you, or because it affected someone else?

2. If you have had an experience like Sam’s — or supported someone through one — what actually happened in the end? How did the system eventually respond, and what did that cost?

3. If you were responsible for this service, what would you have done differently? Not in hindsight, but at the point when the efficiency case was being made — what would it have taken to hold the equity question alongside the cost question from the beginning?

4. Is digital-first, as it is currently practised, genuinely a more accessible model — or is it a more accessible model for people who were already well served, and a worse one for everyone else?

If this rings true, follow The Next Evolution at neilcatton.substack.com. New essays every week, on the technology that shapes us and the choices still available to us.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

In October 2025 an uncrewed submarine called XV Excalibur made a series of dives carrying a quantum optical clock, and kept accurate time throughout without a single GPS signal reaching it. The clock was Infleqtion’s Tiqker; MSubs built the vessel; the Royal Navy ran the trial. Rather than taking a time signal from satellites overhead, the clock reads the behaviour of individual atoms, a reference that needs nothing from space and cannot be jammed from the ground. It was the first time such a device had been operated underwater at sea. The trial was delivered six months ahead of schedule, and Excalibur was handed to the Royal Navy on 11 December 2025 to begin two years of sea trials.

A different machine elsewhere in the UK runs on a version of the same idea. Cerca Magnetics, a University of Nottingham spinout formed in July 2020, installed a wearable brain scanner at Young Epilepsy’s Health and Research Centre in 2021: the world’s first clinical wearable MEG system, built for children. Its sensors are optically pumped magnetometers, arrays that also use atoms rather than a satellite or a moving part as their reference. Where the submarine’s clock tells the vessel what time it is, the scanner’s array tells clinicians where a seizure begins.

Same underlying physics. Two very different speeds.

Why the Navy moved

The urgency behind the submarine trial is not hypothetical. Lithuania’s Communications Regulatory Authority recorded 1,022 satellite navigation interference incidents affecting aircraft in June 2025, against 46 in June 2024, and identified more than ten transmission sites inside Russia’s Kaliningrad exclave. A working paper filed jointly by Estonia, Finland, Latvia, Lithuania, Poland and Sweden to the ICAO assembly on 10 September 2025 put the share of flights affected across their airspace at around 27.4 per cent in April 2025, up from about 7.5 per cent a year earlier.

It has since got worse. Poland’s National Institute of Telecommunications logged full-scale disruption on 63 per cent of days across the first nineteen days of August 2026, and on roughly seventy per cent of days through May and June, attributing it to electronic warfare systems operating in the Baltic region. Europe’s institutional answer has been steady rather than quick. Galileo’s OSNMA authentication service went live on 24 July 2025, and EASA and EUROCONTROL published a joint action plan on interference on 26 March 2026.

No accident has yet been attributed to the jamming, and attribution itself turns out to be contested. On 24 September 2025 Spain’s defence ministry confirmed that Margarita Robles’s aircraft suffered GPS interference near Kaliningrad on the way to Lithuania. Three weeks earlier, Ursula von der Leyen’s flight into Plovdiv was reported across Europe as jammed, on information the European Commission said had come from Bulgarian authorities. Flightradar24 then stated the transponder had shown good GPS quality from take-off to landing; Bulgaria’s deputy prime minister denied passing on any such information; the aircraft had in fact landed on a standard ground-based approach. One case stands. The other dissolved on inspection.

That is the awkward part of the evidence. Aircraft fall back on inertial and ground-based navigation and almost always land uneventfully, so a rising threat produces very few incidents anyone can point at afterwards. The money moved anyway. Infleqtion opened an Oxford innovation centre and manufacturing hub on 28 May 2026, expanding UK production of the clock. Interference is above all an aviation problem, and a submarine is simply where the atomic clock happened to be proven first.

What the scanner is still waiting for

The medical side started earlier and has moved slower. Cerca’s system has been installed and generating research data since 2021. Young Epilepsy, University College London and Nottingham, who developed it together, describe it as more sensitive and more comfortable than the conventional MEG and EEG equipment it would replace, and it lets a child move during a scan without sedation. The Epilepsy Research Institute puts the potential saving at around £2 million a year per site, on the assumption of two additional epilepsy surgeries a week.

None of that has reached routine care. Cerca states plainly that its OPM-MEG system is a research system, with no medical or regulatory approval in any jurisdiction. In April 2026 the company raised £3.8 million in Series A funding, earmarked for clinical approval and manufacturing scale-up. The Epilepsy Research Institute has been running an Innovate UK project with Cerca and Nottingham since September 2025 to map an NHS adoption route, and there is still no clinical pathway at the end of it.

The two devices are not interchangeable, and it would be dishonest to pretend otherwise. An optical clock and a magnetometer array do different jobs, and proving one underwater tells you nothing about approving the other for use on a child’s brain. What they share is a starting point. Both stopped depending on a fragile external reference, a satellite or a moving part, and started reading atoms instead. Both worked. What separated them afterwards was not the physics.

Defence had one buyer who could say yes. The Royal Navy identified a risk that was already escalating and already costed, held the budget, and needed nobody’s permission to run the trial. That is why Excalibur finished six months early. Nothing in the chain required a second signature.

The NHS has no equivalent. No single person can authorise a scanner into paediatric neurology, because the authority is distributed across regulatory approval, clinical evidence, commissioning and trust-level budgets, and each of those can only say no or not yet. That distribution is not a fault in the system. It is the design, and it exists because a device that misreads a child’s brain causes harm that a slow procurement never does.

But the design has a cost, the cost is time, and someone pays it.

My opinion

The asymmetry here is not about whether the technology works. It is about who holds the authority to act once it does. Defence concentrates that authority in one place; health distributes it deliberately, for reasons I would not want reversed. What follows is that a medical device must prove itself to every gate in turn, while a single defence buyer facing a live threat can decide the proving is finished. I am not arguing for fewer gates. I am arguing that we should be honest that they are a choice with a price, and that the price is invisible in a way the Baltic jamming figures are not.

A submarine can now find its way without a signal from space, and the state that wanted it had every mechanism it needed to make that true inside a year. The scanner that could show a clinician exactly where a child’s seizure starts has been sitting in a research suite since 2021, still without the approval that would let it be used in that child’s care.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

The email arrives at 7.43 in the morning. It looks like a delivery notification — a parcel held at a depot, a link to rearrange. Morgan sees it on the way to a client call and clicks. Nothing seems to happen, so Morgan carries on.

Forty miles away, the same email lands in Sam’s inbox at the same moment. Sam works for a financial services company with six hundred employees, a security operations team, and a contract with a threat intelligence provider. The email is quarantined before Sam ever sees it. A flag is raised. An automated report is filed, and the morning continues.

Two people. One threat. Two different outcomes. The difference had nothing to do with vigilance, and nothing to do with the quality of the device in their hands. It had everything to do with who they worked for.

Morgan is a freelance designer — three years in business, a small client base, a laptop and a broadband connection. The click does what clicks do when a link is malicious. By lunchtime, Morgan is locked out of a client’s shared folder. By the following morning, invoices are going out in Morgan’s name, redirecting payment to an account nobody at that end has heard of.

The next seventy-two hours involve a bank, a report to the City of London Police’s Report Fraud service, a client who is understanding but worried, and a second client who is not. Morgan finds the National Cyber Security Centre’s free guide for small businesses. Then its Cyber Action Plan, and then Check Your Cyber Security — all free, all built for exactly this situation, all written by people who know what they are talking about.

None of them is a system. They tell Morgan what to check and what to change. They do not sit in front of the inbox at 7.43 and stop the email arriving. That is the distinction the conversation keeps missing: advice is available to everyone, and infrastructure is not.

The phishing email was not sophisticated. Across the same city that morning, the same email was intercepted and neutralised at organisations with security operations teams, by automated systems most of their employees had never thought about. Morgan did not have access to those systems. Morgan had access to the same email.

There is a word for what happened here, and it is not misfortune. It is structural.

Cybersecurity protection in the UK is distributed by employer, not by need. The protections that stopped Sam from ever seeing that email exist because Sam’s organisation can afford to buy them, maintain them and staff them. They are not public infrastructure. They are private expenditure, scaled to the size of the organisation that funds them.

The scale of that gap is measurable. The government’s Cyber security breaches survey, published on 30 April 2026, found that 76 per cent of large businesses have a formal incident response plan. Among micro businesses the figure is 21 per cent. Across all UK businesses, only 11 per cent use or invest in threat intelligence at all — which means the arrangement Sam’s employer has is not standard practice. It is a purchase, and most organisations cannot make it.

The sole trader, the freelancer and the small charity face the same threats as the organisations with the security operations centres, and they face them on their own account. The conversation about what individuals should do — stronger passwords, updates, wariness about links — is not wrong. It is aimed at the wrong layer. Sam was not protected by being careful. Sam was protected by a payroll.

For Morgan, the consequences are immediate: three days of lost work, and two client relationships that will take longer than three days to repair.

The phishing email does not check who you work for before it decides to arrive.

There is a second consequence that takes longer to see. Small businesses are not peripheral to the UK economy. Businesses with fewer than fifty employees account for 13.1 million jobs — 47 per cent of private-sector employment, on the Department for Business and Trade’s 2025 figures. When that part of the workforce is disproportionately exposed to cybercrime and disproportionately unprotected against it, the cumulative cost is not a collection of individual misfortunes. It is a structural tax on the least resourced part of the economy, levied by criminals who understand the asymmetry better than the policy conversation does.

None of this is about the technology failing. Sam’s system did exactly what it was built to do: the email was quarantined, the morning continued, the harm never arrived.

Enterprise security works partly because it learns the organisation it sits inside. It knows which invoices are normal, because it has watched a year of them. A freelance designer with the same shared folders and the same monthly invoice cycle is exactly that kind of predictable context. Nobody built it for Morgan. The problem was never that a single inbox is too complicated to protect — it is that a single inbox was never going to pay for it.

My opinion

I don’t think this is really about awareness, though the conversation usually pretends it is. Morgan didn’t fail to ask the right question — Morgan didn’t know there was a question to ask, and neither do most sole traders and freelancers I’ve worked alongside. The tools that would have stopped that email exist and work. What’s missing is the money, and the confidence to spend it on something that feels invisible until the day it isn’t. I think government should treat this the way it treats other public safety problems it doesn’t leave to the market alone — by funding working protection for the self-employed and the smallest businesses, not another campaign telling them to be more careful.

If you’ve been on the wrong side of it

• Does the protection you work behind reflect what you actually need, or what your employer happened to buy?

• If you have been through something like this, or watched someone close to you go through it, what did the gap between the threat and the available help look like from the inside?

• For anyone who has bought security for a small organisation: what did you get for the money, and what did you find you still could not buy at that size?

• And if you have moved between employment and working for yourself, what changed about your exposure that nobody warned you about?

The argument behind this article is developed in my book The Shadow System, specifically the chapters on The Corporate Digital Divide and The Unbearable Asymmetry in Part III, which examine how the gap between governed and ungoverned security has become one of the defining structural problems of the digital economy.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

On 9 July 2026, the FCC approved a satellite called Eärendil-1. It carries an 18-by-18 metre deployable mirror, built by a company called Reflect Orbital, designed to reflect concentrated sunlight down onto specific points on Earth after dark. It’s a demonstration for a stated long-term plan: a fleet of 50,000 identical satellites by 2035, selling sunlight on demand, anywhere, at night.

The approval went through over more than 1,800 public objections, and formal objections from the American Astronomical Society and DarkSky International. The FCC’s justification was a citation rather than a rebuttal: the 1934 Communications Act, which obliges the agency to encourage the development of new technology, and says nothing at all about the night sky.

What’s already true, before the fleet exists

The temptation is to treat this as a story about the future, one satellite ahead of a problem that hasn’t arrived yet. It understates what’s already measured. Roughly 14,000 satellites are already in orbit, more than 10,000 of them Starlink, and a further 1.7 million have been proposed across various constellations, most of them still awaiting launch, according to an ESO study published in July 2026. Fourteen thousand is already enough to matter; over a million would be a different sky entirely.

A 2021 study, published in Monthly Notices of the Royal Astronomical Society: Letters, found that satellites and space debris were already raising night-sky brightness by more than 10% above natural levels across large parts of the planet — a figure five years and thousands of launches out of date. A newer study, led by Anthony Mallama and Richard Cole of the IAU Centre for the Protection of the Dark and Quiet Sky and published in the same journal in 2026, found that nearly all satellites in the major constellations already exceed the IAU’s own brightness limits for professional astronomical interference — before Reflect Orbital’s fleet has launched a single additional unit.

What it already costs, and who has no vote

ESO modelling shows the Vera Rubin Observatory’s new survey camera, which began operating in 2026, could lose hours of usable data a night to bright satellite trails. Olivier Hainaut, the ESO astronomer whose simulations underpin the study, found that a single Reflect Orbital mirror satellite could electrically saturate the camera’s detector, leaving a ghost trail that ruins an entire exposure — without the beam ever pointing at the observatory directly. At full fleet strength, every image taken while its satellites are illuminated could be wiped out. The ESO calculates that its own facilities would see background sky brightness rise three to four times under the complete 50,000-satellite constellation. Research into Indigenous rights and outer space, compiled for the IAU Centre for the Protection of the Dark and Quiet Sky’s policy work, documents a further cost: a bright, moving night sky disrupts navigation and oral traditions that depend on a dark one.

There is a body whose job is to care about this. The IAU Centre for the Protection of the Dark and Quiet Sky, formed in 2022 and jointly hosted by the Square Kilometre Array Observatory and NOIRLab, researches the problem, recommends brightness thresholds, and helped bring the issue onto the agenda of the UN’s Committee on the Peaceful Uses of Outer Space, COPUOS, the closest thing satellite governance has to a global forum. Getting an item onto that agenda is not the same as having a rule adopted, and adoption is not the same as enforcement. The centre has no binding authority anywhere, and nothing compels a licensing body like the FCC to consult it, weigh its findings, or even acknowledge them before granting an approval.

That’s the actual gap. It isn’t a missing rule about satellites, or a loophole Reflect Orbital found and exploited. Licensing bodies like the FCC were built to assess a technical filing: spectrum use, orbital slot, signal interference. They were never built to weigh a diffuse, global, shared-commons harm with no single financial cost and no single injured party who can bring a claim. The Royal Astronomical Society said as much in May 2026, arguing that “nighttime should be recognised as an integral part of the natural environment, requiring protection on a par with daytime activities” — a status it holds nowhere.

The approval that starts a fleet of 50,000 mirrors has already happened. The body that granted it wasn’t required to weigh the night sky against the mandate it exists to serve, because nowhere does that requirement exist. What would it take for the sky itself to become a party with standing in a process built entirely around the applicant?

My opinion

This is what approval-first looks like when nobody has mapped the consequences. Lighting the sky for hours after dark, night after night, is not a neutral act — it will affect sleep, wildlife, crops, and cultures in ways nobody has finished studying, and the fleet is being built anyway. The question underneath the licensing gap isn’t really about satellites — it’s whether any single body should have the authority to make a decision that reaches every person on the planet without asking them first. I don’t think it should. Not yet, and not like this.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Google DeepMind’s Co-Scientist, formalised in a Nature paper in May 2026, compresses the early stage of scientific discovery, generating and refining a research hypothesis, from weeks to days. It works as a multi-agent system built on Gemini: separate AI agents continuously generate hypotheses, critique each other’s reasoning, and refine what survives, in a loop that used to depend on a human researcher’s slower cycle of reading, thinking and drafting. It isn’t a demo. Now rolling out experimentally to individual researchers through Gemini for Science, it has already produced hypotheses that survived contact with a wet lab: a novel drug-repurposing candidate for acute myeloid leukaemia, and a new epigenetic target for liver fibrosis, both independently confirmed rather than merely plausible-sounding. This is a real capability, working, now.

In the same literature this capability is meant to accelerate, something else is happening. Fabricated references in biomedical papers have risen twelvefold in two years. Maxim Topaz, an associate professor at Columbia University’s School of Nursing and Data Science Institute, led an audit of 2.5 million papers — published as a letter in The Lancet in May 2026 — that found roughly one in every 277 PubMed-indexed papers now contains some fabricated references, up from one in 2,828 three years earlier. Review articles, the papers other researchers lean on most heavily to orient themselves in a field, showed a fabrication rate 57% higher than other paper types. It’s already sitting in the permanent scientific record, not a future risk to be managed.

One documented case, precisely

Most of that twelvefold rise is inference from pattern, not confession. But at least one instance is on the record in detail. A 2026 retraction from the Journal of Medical Ethics was explicitly attributed to an author who had used generative AI to identify and understand referenced sources, without independently verifying them before submission. The citations looked right. Nobody checked whether they were.

That’s the mechanism in miniature: a tool that can write a fluent, well-formatted reference list faster than any human, attached to a verification step that still assumes a human wrote it carefully. Multiply that gap across an entire literature moving at AI speed, and the twelvefold rise stops being surprising. The higher fabrication rate in review articles matters more than the raw number suggests, because a review article is exactly the paper other researchers reach for first when they’re trying to orient themselves in an unfamiliar field — it’s the summary everyone trusts precisely because checking every citation in it individually is what the review was supposed to save them from doing.

The system nobody rebuilt

Peer review and citation-checking were designed for a slower, lower-volume era of publishing — one submission at a time, one careful human reader at a time. AI can now generate a polished hypothesis, a literature review, and a citation list faster than that process was ever built to check — and the researchers behind the audit say plainly that the checking side hasn’t caught up to the volume the generating side now produces. There is still no accepted, mandatory standard for flagging an AI-assisted finding as provisional before it enters the permanent scientific record.

My opinion

I don’t read this as a publishing-etiquette problem, the kind a stricter house style or a sterner editor’s note could fix. Verification is infrastructure, in exactly the same sense that the wet lab confirming Co-Scientist’s drug-repurposing candidate is infrastructure. Nobody would fund a hypothesis-generation system this fast and then leave the checking to whoever happens to notice. Yet that’s precisely what’s happening on the citation side, because the money and the attention are on the generation half of the pipeline, not the verification half. Until that changes, the speed is the risk, not the safeguard.

That gap has consequences beyond the literature itself. A funding committee weighing a grant application, a systematic review synthesising a field, or a clinical trial recruiter screening candidates has no reliable way to know whether the citation or hypothesis in front of them survived independent checking, like Co-Scientist’s confirmed drug-repurposing candidate, or only survived generation, like the reference list behind the Journal of Medical Ethics retraction.

Even Google is candid about this. The company acknowledges Co-Scientist’s own limitations in literature review and factual accuracy — meaning the most rigorously validated AI-assisted discovery tool currently available still depends on external checking that the surrounding system isn’t built to provide at the speed the tool now works at. The researchers behind the audit have recommended that citation-checking move earlier, built into the submission workflow before peer review begins rather than caught, if at all, after publication — a change that depends on publishers adopting it, which none has yet done at scale.

On the page, a hypothesis nobody has independently checked looks identical to one that’s been rigorously verified. Right now, nothing in the publishing system forces that distinction to be visible before the finding starts shaping what gets funded and which patients get recruited into which trials. How many findings currently doing that work have never actually been checked — and would anyone know?

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Ask people at almost any company whether they use AI for work, and the honest answer is usually yes — whatever the official policy says. MIT researchers who studied enterprise AI adoption this year found that employees at more than 90% of the companies they surveyed were already using personal ChatGPT or Claude accounts for work, regularly, some of them every day — while only 40% of those companies had bought so much as one official AI subscription. The organisation didn’t approve any of it. It just didn’t get in the way fast enough to stop it.

That gap between what a business has sanctioned and what its people have already gone and done for themselves, is the real story behind a number that’s become almost a cliché this year: roughly 95% of generative AI pilots fail to deliver any measurable return, against $30 to 40 billion in combined enterprise spending. The same research found something more specific underneath that headline figure. Pilots built on generic, off-the-shelf tools reach implementation 83% of the time, because they ask almost nothing of the organisation around them. Custom pilots — the ones meant to touch a real process — mostly never get past the pilot stage at all.

What the pilots are actually funding

Look at where the $30–40 billion has gone and a pattern appears. Most of it has gone into things that generate enthusiasm rather than change: a chatbot here, a summarisation tool there, a demo for the board that photographs well in a slide deck. Each one is cheap to start and, when it doesn’t work, cheap to quietly abandon.

The harder work doesn’t get funded at the same rate: auditing whether the underlying data can actually be trusted, naming the political reason a broken process has never been fixed, admitting a failure that predates the technology by years. None of that demos well, none of it fits inside a single budget cycle, and none of it feels like progress in the moment — even though it’s the only work that would tell you whether a tool could ever have helped in the first place. So it keeps not happening, and the pilots keep multiplying instead.

The mirror, not the tool

AI doesn’t reveal what’s wrong with a business. It runs straight into it. A process that was already unclear stays unclear, just faster. Data that was already messy gets amplified rather than cleaned up, because nothing about deploying a model answers the question nobody asked about where that data came from or whether it was ever trustworthy in the first place.

The technology isn’t lying about the organisation it’s dropped into. It’s showing that organisation exactly what was already there — at a speed that makes the thing much harder to keep ignoring.

The alternative isn’t complicated

None of the harder work requires new technology. Deciding who owns a piece of data, and whether it can actually be trusted, is a decision an organisation could have made before any pilot started. Deciding who’s accountable when a model gets something wrong is a governance question, not an engineering one. Giving staff an approved tool that’s as fast and as unrestricted as the one they’ve already found for themselves is a procurement decision, not a research problem.

What makes this work hard isn’t complexity. It’s that it requires someone to admit, out loud, that the process was already broken before the model arrived — and admitting that is exactly what a shiny new pilot lets everyone avoid doing.

Who is left holding the gap

The employees quietly running work through their own AI accounts aren’t doing it to make a point. They’re doing it because the sanctioned tool doesn’t do the job and the deadline hasn’t moved. That leaves them carrying two things nobody chose to give them: the actual risk of putting company information through a personal, unmanaged account, and the actual judgement calls — what’s fine to paste in, what isn’t — that no governance process has ever talked them through.

None of that is a technology failure. It’s a business that outsourced a hard decision to whoever was closest to the deadline, then called the resulting mess an AI adoption problem.

A better model won’t fix what a business has never been willing to look at directly. It will just show that back, more clearly, and faster than before — and leave someone else to carry what it finds.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Jordan had been using the platform for two years by the time the standard was announced. Not easily — with a screen reader that could parse about half of what the interface contained, and a set of workarounds so habitual they had ceased to feel like workarounds. Every week: export the dashboard data to a spreadsheet to read it, update the spreadsheet offline, re-enter the changes manually. Three extra steps for every task a sighted colleague completed in one. The platform was not designed to exclude Jordan. It was not designed for Jordan at all.

The standard arrived as a regulatory document with an eighteen-month compliance window. Jordan did not know about it immediately. There was no reason to. Standards are announced to the organisations that have to meet them, not to the people who will benefit from them.

In Sam’s team, the announcement landed differently.

The response the organisation had prepared

Sam was not against accessibility. Nobody in the team was against accessibility. The resistance was to the timing, the resource, and the assumption that eighteen months was enough to rewrite the parts of the product the standard required. The roadmap was committed. Three major features were already in progress. The compliance work was not budgeted. The concern was not that the standard was wrong but that it had arrived at the wrong moment.

That concern was genuine in its own terms. Within nineteen months, the team would need to audit the full product against the standard’s requirements, identify every failure point, prioritise the remediation work, retest with assistive technology, and document the compliance case. For a product of any complexity, that is not a small undertaking. The estimate that came back from the initial audit was uncomfortable.

What was also in that audit, if anyone looked carefully enough, was a list of things the team had wanted to do but had never been able to prioritise. The keyboard navigation issues that the accessibility standard required to be fixed had been in the backlog for eight months. The plain language requirement had been discussed at two design reviews and deferred both times. The colour contrast failures had been flagged by one of the designers the year before, logged, and then — as things are logged — left there.

The standard had not introduced new problems. It had made existing ones compulsory to address.

There is a pattern here that runs across the history of technology regulation. When the Disability Discrimination Act was introduced, the adjustments it required were widely resisted as disproportionate by the industries subject to them. The Web Content Accessibility Guidelines were described as technically unrealistic for large, complex products. The EU Accessibility Act was described in similar terms by the industries preparing for its June 2025 deadline. Each time, the argument is that the standard is too far ahead of where organisations are. Each time, the organisations that meet the standard find they were closer than they thought — and further from where they should have been.

From compliance to floor

The reframing that changed the team’s view was not philosophical. It came from the user testing.

The standard required testing with the people it was designed to protect. The team arranged the sessions. For most of them, they were the first time anyone had watched someone use the product with a screen reader.

What they saw was uncomfortable. They watched someone work through the same tasks the team had been designing for two years. The time it took, and the places where the screen reader fell silent because there was no text equivalent for a visual element. The workaround that had developed for the dashboard export — not because it was a natural way to work, but because it was the only way to work. The product had been declared a success on every internal metric. From outside the design assumptions that had produced it, it was a product with a secondary interface for a significant portion of its potential users, and that secondary interface was made of spreadsheets and manual re-entry.

The user testing did not produce guilt. It produced information. The team had not known. Now they knew. The standard had been the mechanism that produced the knowledge, and the knowledge changed what the work meant.

The reframe - from compliance burden to design floor - followed from that session. The standard was not telling the team how good the product could be. It was telling them how bad it was no longer permitted to be. One instruction sets a ceiling. The other raises a floor. The team had been treating the standard as the former. It was the latter. Once they understood that, the work changed from meeting a requirement to building from a new foundation.

What the floor revealed

The improvements came in waves. Screen reader compatibility - the most technically complex of the required changes - produced the first unexpected consequence. Making every interface element accessible to a screen reader required auditing the entire navigational structure of the product. The audit found redundancies, dead ends, and paths through the product that nobody had designed and that made no sense to any user. Fixing the screen reader compliance required fixing the navigation. The navigation that emerged was better for everyone.

The plain language requirement produced the second wave. Error messages rewritten for cognitive accessibility were clearer to all users. Support call volume for error-related queries fell within months of the changes going live. The calls had been telling the team something about the product that the product’s own metrics had not surfaced.

Colour contrast was the third. The standard required ratios that the team’s existing palette did not meet. The redesign of the colour system produced a product that performed better in direct sunlight, on low-quality displays, and - as subsequent user testing confirmed - for users over fifty whose colour sensitivity had shifted with age. The team had been designing for a calibrated monitor in a dim office. The contrast requirement forced a wider assumption about where and how people use things.

This is the curb-cut effect: standards designed for a specific group produce improvements that serve a wider population than the standard required. The lowered kerb that allowed wheelchair access also helped people with pushchairs, delivery workers with trolleys, and cyclists. The plain language, the contrast ratio, the keyboard navigation did the same thing. Jordan could use the product. So could everyone else, more easily than before. The floor that had been raised for Jordan turned out to be below where the product should have been all along.

The team had not set out to make a worse product than they were capable of. They had made the product their design process was equipped to produce, for the users their design process had in mind. The standard forced a wider design process. The wider design process produced a better product.

That sequence is not unique to accessibility. External requirements have forced expanded scope and produced better outcomes in electrical safety, food labelling, and data protection. The resistance is always first. The improvement follows.

What a standard is really testing

The question a well-designed standard asks of a technology product is not only whether it is accessible to people with disabilities. That is the specific requirement. The deeper question is whether the product assists the people who use it — all of them, not only the ones whose needs were visible in the original design brief.

A product that cannot be used by someone with a screen reader was, by definition, only helping the users whose circumstances it had accounted for. Meeting the standard did not add an accessibility layer to a complete product. It completed the product for the first time.

The second question is whether the product adds something real to the capability of the people who use it, or only enables interaction. Plain language that reduces cognitive load does not just make the product reachable - it makes it more useful. The user who previously needed three support calls to understand an error message needed none after the change. That is not accessibility as a separate category. That is a better product, full stop.

The third question - the one that runs underneath the history of the Disability Discrimination Act, the Web Content Accessibility Guidelines, and the EU Accessibility Act - is adaptability. A product designed for one profile of user, however carefully designed, is not adaptive. A product that has been required by a standard to respond to a wider range of human circumstance has, in meeting that requirement, become demonstrably more so. The standard did not produce adaptability. It required the work that produced it. That is what a well-designed standard does: it requires the work the organisation would not otherwise do, for the people the organisation had not otherwise considered.

My opinion

Technology follows the path of least resistance. So does design. We build for the user we imagine, in the circumstances we assume - and that user does not exist. Real people are older, less sighted, less confident, using screens in conditions we did not test for. A standard forces the reckoning with that gap. It should not have to.

Four questions for the next standard

1. Think of a regulatory or technical standard that applies to technology you build or use. Has meeting that standard improved the product for a wider population than the standard required? If you do not know - would it be worth finding out?

2. If your organisation is currently resisting a new accessibility or ethical design standard, what is the specific argument against it? Is that argument based on evidence of what the standard would cost and produce - or on an assumption made before the work was done?

3. If you are involved in setting standards - in government, in industry bodies, in procurement specifications - are the people the standard is designed to protect part of the process by which it is written? And if not, how confident are you that the standard actually requires what those people need?

4. The argument against technology standards is consistently the argument of the organisation that has not yet had to meet them. What would it take for your organisation to be the one that argues for higher standards rather than against them - and what would that mean for the people your technology is supposed to serve?

Jordan can use the platform now. Not with workarounds. Just - use it. The standard required that as a minimum. The team went further because, once they understood what the floor was, they found the ceiling was higher than they had thought. The standard did not limit what was possible. It clarified what was no longer acceptable.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

For most people with idiopathic pulmonary fibrosis, the question is not whether their lungs will fail but how quickly. There is no cure. The available treatments slow the damage; they do not stop it. Median survival after diagnosis is three to five years.

In June, at the BIO International Convention in San Diego, Insilico Medicine announced a collaboration worth up to $2.5 billion in milestone payments with SK Biopharmaceuticals for neuroimmune disorders — a year after publishing Phase IIa trial results for its lead IPF drug in Nature Medicine. The molecule at the centre of this is Rentosertib, a TNIK inhibitor designed to slow the fibrotic damage that kills IPF patients. It is the first drug candidate where both the target and the compound were identified entirely by AI. The platform synthesised and tested 79 candidate molecules; Rentosertib was the 55th. From target identification to clinical candidate: 18 months.

The question is what to do with that.

What 18 months actually means

Traditional drug discovery takes years before a compound reaches a trial. Identifying the right biological target is itself a years-long process — hypothesis, experiment, elimination. Then designing a compound that binds to it, screening variants, testing in the laboratory, a Phase I trial to establish safety. By the time a drug enters efficacy testing, four to six years have typically passed.

Insilico collapsed that phase. The AI generated and assessed virtual molecules computationally, identified the best candidates, and selected one for synthesis without the iterative cycle that makes traditional chemistry slow. The biology remained; the chemistry accelerated.

Rentosertib’s Phase IIa trial enrolled 71 patients across three dose arms, randomised, double-blind, placebo-controlled. Patients on the highest dose showed a mean improvement in lung function of 98.4 millilitres over the study period; the placebo group declined by 20.3 millilitres. For a disease measured in how fast function falls, that gap matters. The results are not definitive; no Phase IIa trial is. But they were published in Nature Medicine in June 2025, and they were positive.

Discovery has changed. The question is whether the rest of the process noticed.

The infrastructure was not built for this

After Phase IIa comes Phase III. Larger trials, typically thousands of patients, with full efficacy and safety data required for regulatory submission. Insilico opened Rentosertib’s own Phase III trial in July, enrolling 320 patients across 47 sites in China. The median time from Phase IIa completion to approval still runs several years, even with expedited review designations. Fast Track status accelerates the conversation with regulators, not the data requirement.

The regulatory requirement is appropriate: evidence from 71 patients is not enough to approve a drug for general use. Phase IIa is promise; Phase III is proof.

But the clinical trial system was not designed around a world in which a promising molecule can be found in 18 months. Recruiting patients, running trials across multiple sites, generating data, submitting for review: the queue was sized for a pipeline that moved at human speed. AI has shortened the front of that process without changing anything that follows it.

For someone diagnosed with IPF today, the arrival of AI drug discovery does not change their timeline. Rentosertib may eventually be approved, or it may not; Phase III trials fail more often than they succeed. The drug was discovered faster, but the path from discovery to patient has not changed. The bottleneck moved from the laboratory to the clinic.

What gets built next

The collaboration announced in June is, among other things, an industry bet that AI will produce candidates at a scale and pace the traditional pipeline never achieved. If that bet is correct — and the discovery-side evidence is accumulating — then the clinical trial system will face pressure it has not faced before: not a shortage of candidates but a surplus of them, each requiring the same years of Phase III data.

Some adaptation is already under way. Adaptive trial designs, accelerated patient matching, synthetic control arms built on existing patient data, earlier integration of real-world evidence from approved therapies — these approaches exist and are being explored. None of them is available to the patient who needs a drug now, and none changes the gap between AI-accelerated discovery and human-speed delivery.

The FDA and the EMA published joint guiding principles for AI in drug development in January, and the MHRA has spent the same period expanding its regulatory sandbox for AI tools. None of it touches the actual constraint. The guidance governs how AI models get validated for use inside a regulatory submission, not how many Phase III trials the system can run at once, or how fast a surplus of candidates gets through it.

My opinion

I think this is the pattern, not the exception. We keep running into the same asymmetry: what a technology like AI can now do at pace and at scale, set against governance and legal procedures that were never built to move at that speed. Some of that caution is earned — Phase III exists because it should be hard to approve a drug on the strength of 71 patients. But most of the controls now standing in the way were built for a more analogue world, and they have not moved into this one. Until they do, every acceleration in discovery just relocates the wait.

For IPF specifically, the constraint is concrete. Median survival after diagnosis is three to five years. The drug that might help was discovered in 18 months. The distance between those two facts is not a scientific problem. It is a system problem — and the system in question is not the one that found the molecule.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

The email arrived at 3.17 in the morning.

Sam hadn’t slept. They’d been sitting at their kitchen table for four hours watching the progress bar on their screen crawl — file after file locked, folder after folder gone dark. The attack had started somewhere in the accounts system, spread quietly through the network, and by midnight the company’s billing platform, its customer database, and eighteen months of project files were encrypted behind a key Sam didn’t have.

Then the email came. It was polite. It opened with an acknowledgement that this was a difficult situation. It explained, clearly and without technical jargon, what had happened and what the options were. It included a link to a payment portal with step-by-step guidance. It offered a live chat facility for anyone who needed help navigating the process. At the bottom, there was a note that a support representative was available around the clock and would typically respond within the hour.

Sam had never received customer service this good from anyone.

The support desk that worked

This is not an unusual story. The criminal organisations behind ransomware attacks have built support infrastructures that would be impressive in any legitimate technology business — documentation in multiple languages, and in some documented cases a rating mechanism through which affiliates — the criminal operators who carry out attacks on behalf of the platform — could evaluate the service they received.

They have done this because it works. A victim who cannot navigate the payment process is a victim who might not pay. Friction costs money. They removed it.

The DarkSide group, responsible for the 2021 attack on Colonial Pipeline in the United States, operated with the organisational structure of a funded technology company. They developed proprietary tools, leased their capabilities to vetted affiliates who carried out attacks on their behalf, and maintained a support function specifically to guide victims through the payment and decryption process. The attack locked down the billing and operational systems of the largest fuel pipeline on the US East Coast, triggered fuel shortages and panic-buying across multiple states, and prompted a federal emergency declaration. Colonial Pipeline’s CEO, Joseph Blount, confirmed to the Senate Homeland Security Committee that the company had paid the ransom — 75 bitcoin, approximately $4.4 million at the then-exchange rate. The ransom note had demanded approximately five million dollars.¹ The support desk was open the entire time.

The question this raises is not primarily about the criminals. Their logic is coherent. They are running a revenue model and they have optimised it. The question is about what their model reveals about everything else.

What the legitimate system was built to do

A person dealing with a digital crime — whether ransomware, fraud, or identity theft — typically encounters a response system that was designed for a different era and a different scale of problem. The bank’s fraud line connects them with an adviser working from a script that was written before this kind of crime existed in its current form. The police non-emergency line takes a report and assigns a reference number. The relevant government body sends an acknowledgement within five working days.

Each of these institutions is staffed by people trying to help within a structure that was not built to help at speed, at scale, or at the point of maximum need.

The criminal support desk is better not because criminals are more talented. It is better because it was designed entirely around what the paying customer needs at the moment they need it. That is a design choice. The legitimate systems were designed around something else — compliance, liability, institutional process, historical precedent — and the human being in distress is not at the centre of any of it.

“The ransomware gang had better customer service than the bank that was supposed to help you recover from them.”

When a fraud line or a government body responds to a digital crime victim, the response is designed to help that person complete a process — file a report, receive a reference number, trigger a workflow. That is not the same as helping them do what they actually need to do. The paper trail grows. The problem does not. And the response does not adapt: it does not know where this specific person is right now — technically, emotionally — and it was never built to find out.

The criminal support desk was built around exactly that question. Not because of any ethical commitment, but because its commercial success depends on it. That is the uncomfortable mirror it holds up. The legitimate system, by contrast, was designed to be defensible. That is an institutional virtue. It is not the same as a human one.

The Colonial Pipeline attack disrupted fuel markets across multiple states for six days, driving fuel prices above three dollars a gallon for the first time since 2014 and prompting panic-buying across the south-east of the United States.² The polite emails were in service of extortion.

What the comparison reveals

The argument is that something in the legitimate system has been designed wrongly, and that the criminal model makes the wrongness visible by contrast. When a fraud victim gets better guidance from the people who defrauded them than from the institution that holds their money, that is not an anomaly. It is a diagnostic. The system that was built to protect and assist that person has, somewhere in its design, prioritised something other than them.

That matters because the design is not inevitable. It is a series of choices, made by people, in institutions, at particular moments in time.

My opinion

Cyber crime is not the work of individuals in dark rooms. It is a structured business model, operated by organised criminals with the discipline and infrastructure of any revenue-generating enterprise. The support desk is not absurd — it makes complete commercial sense. Once you accept that framing, the question for every organisation changes. Not “could this happen to us?” but “what is our blast radius, and what do we have in place to recover?” This is not going to become simpler or rarer. Crime now operates at the speed of code. Our response still runs at the pace of policy and procedure. That is an asymmetric challenge, and naming it is the first step to treating it seriously.

This is the territory that The Shadow System covers in its opening section — the professionalisation of the criminal economy, and what that professionalisation reveals about the failures of the legitimate one.

Your experience is part of this

These patterns are not abstract. The design failures described here accumulate in individual moments — a fraud call that went nowhere, a report that produced only a reference number, a process that kept moving while the person at the centre of it stood still. If you’ve encountered any part of this, your account is relevant evidence.

• Does this story relate to you — have you ever found the process of reporting a digital crime or fraud more difficult than the crime itself deserved?

• If you have your own experience of how institutions responded when something went wrong digitally, share it. These patterns only become visible when enough people name them.

• If you could redesign the response a fraud or ransomware victim receives from legitimate institutions, what would be the first thing you would change?

• And the harder question: is this a technology failure, a culture failure, or an incentive failure — and does the answer change what we should do about it?

References

¹ Joseph Blount, CEO, Colonial Pipeline — Senate Homeland Security and Governmental Affairs Committee testimony, 8 June 2021. Blount confirmed the payment was 75 bitcoin (approximately $4.4 million at the time of payment); the original ransom demand was approximately $5 million. The FBI subsequently recovered approximately $2.3 million by seizing 63.7 Bitcoin (the value had fallen since the date of payment). Sources: CNBC; Nextgov/FCW; Department of Justice press release, 7 June 2021.

² Colonial Pipeline shut down 7 May 2021, restarted 12 May 2021. Fuel price and disruption data confirmed across multiple sources; note that a University of Kansas study (December 2021) found the longer-term gas price impact was more localised than early reporting suggested. Sources: CISA retrospective; University of Kansas study.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Google pays Reddit roughly $60 million a year for access to its forums. OpenAI pays Reddit around $70 million for the same thing, and separately signed a five-year, $250 million deal with News Corp. None of the people who wrote the posts, comments and articles now changing hands for that money were party to any of these deals. They wrote for free, on platforms that were free to join, and the platforms sold what they wrote.

That should be the fact that stops you, before any argument gets built on top of it. Live-access licensing deals like these have grown from two in 2023 to a projected thirty-four in 2026 — a market that didn’t exist three years ago, already moving hundreds of millions of dollars a year, for something the AI industry used to scrape for nothing.

Why the market suddenly needs “real” data

The reason is a 2024 finding, published in Nature, that AI models degrade when they’re trained on their own output — not goodwill toward journalism or forum culture. Ilia Shumailov, a computer scientist then at Oxford, and his co-authors showed that models fed a diet of AI-generated text suffer what they called model collapse: an early drift away from the true distribution of language, followed by the permanent loss of rare, low-frequency information — the unusual phrasing, the minority view, the edge case that only shows up once in a million real sentences. Train a model on text that was itself written by a model, repeatedly, and each generation grows narrower and worse.

The supply of text that hasn’t been through that loop is shrinking fast. By April 2025, an estimated 74.2% of new webpages already contained AI-generated text. Epoch AI has projected that the internet’s stock of fresh, verifiably human-written material could be exhausted sometime between 2026 and 2032. Put those two facts together and the licensing boom stops looking like opportunism and starts looking like the market pricing in a resource it only just discovered was finite.

What buying real data gets you

For a lab that can afford it, a licensing deal buys something specific: a continuously refreshed pipeline of text a real, currently living human being wrote, with a paper trail behind it. That’s a genuine practical advantage: slower degradation, a defensible answer to “what did you train this on,” and a supply of language that keeps including the rare, awkward, human sentence that a model trained on its own past output would quietly lose.

Regulation has started reaching toward this problem, but not from the direction that matters here. The EU AI Act’s labelling mandate, which takes effect in August 2026, will require AI-generated images, video and audio to carry machine-readable labels, with fines of up to €15 million for non-compliance. Text is explicitly exempted — the European Commission’s own reasoning is that reliable detection of AI-written text doesn’t exist yet. So the one piece of law built to tag AI output has nothing to say about the market this article is describing: hundreds of millions of dollars a year, changing hands over exactly the kind of text the mandate can’t touch.

What it leaves out, and for whom

The money flows to the platform, not the person. Reddit and News Corp get paid because they own the pipe the content flows through; the individual who posted the comment or wrote the article does not see a share of either deal, and has no legal or technical route to ask for one. Nothing built so far reaches past the platform to the creator.

Smaller and open-source developers are shut out from a different direction. A verified, licensed data pipeline costs tens or hundreds of millions of dollars to assemble. A developer who can’t pay that price is left training on the same open web that’s now more than 74% synthetic, with no way to check what’s clean and what isn’t. The gap this opens is concrete: a well-resourced lab’s model gets steadily better data, a smaller developer’s model gets steadily worse data, and neither the developer nor their users are told which side of that split they’re on.

Government hasn’t closed the gap either. The UK’s March 2026 Report on Copyright and Artificial Intelligence chose to keep the existing rules in place, favouring voluntary licensing and transparency over reform, even though 88% of the more than 10,000 people who responded through the consultation’s Citizen Space platform backed the option that would have required licences for AI training — the strongest protection on the table. The market has already answered the question of what this content is worth — hundreds of millions of dollars a year and rising. Nobody in government has yet answered the question of who, if anyone, is owed a share of it.

My opinion

I write a lot both here and in my published books, so I read this as someone whose words could end up in the same pipeline. The question of paying for what someone else created isn’t new — publishers have negotiated royalties for a century. What’s new is the scale: this isn’t licensing a book or a photograph, it’s the wholesale absorption of everything a person has ever written, folded into a model with no accounting for whose sentences did the work. Some form of monetary recognition has to follow the creator, not stop at the platform that happened to host them.

The AI industry has just told everyone, in real money, exactly what their unpaid words are worth. It has not told anyone how, or whether, they will ever see any of it.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Jordan is sitting with a member of staff who will use the new system every day. Not in a workshop. Not in a project meeting. In the room where the work actually happens, watching how it is done, listening to what is said about it.

The person Jordan is talking to says something that stops the conversation mid-note. Not a complaint — a small, specific observation. A workaround that everyone on the team uses, that nobody has mentioned in any survey or requirements session because it has become invisible through familiarity, and that had the visit not happened would have been absent from every document the technology decision was based on.

Jordan writes it down. It will change what is built.

What the standard process does not capture

The technology decision process here is not badly designed. The business case is thorough. The vendor shortlist is credible. The project workshops have run, and the requirements document reflects what was said in them. The governance process is in place.

What the formal process cannot capture is what is now being heard. The requirements document reflects what the previous system did — the starting point, not the need. The project workshops produced consensus around the visible problems, the ones that are easy to articulate in a meeting room to a project manager with a template. The satisfaction survey measured experience against the current system’s performance, not against what the work actually requires.

Each layer of mediation between the decision-maker and the person doing the work reduces the fidelity of what arrives. The user research output has been simplified for an executive audience. The support tickets have been categorised and aggregated. The workshop outputs have been translated into requirements language. By the time the information reaches the decision-maker, the texture has been processed out of it — the hesitation before an answer, the third thing said after the first two, the small specific observation that reveals the real constraint.

The visit was to hear what the processing had lost.

Going before the direction was fixed

The conditions that distinguish this kind of listening from the kind that confirms rather than informs are timing, directness, and repetition — and each runs against what the project calendar prefers.

The first condition is timing. Going before the strategy is fixed — before the vendor shortlist is final, before the requirements are locked, before the business case has closed around a particular direction — means the listening can still change things. Listening that happens after the direction is set validates rather than shapes. The project team hears what it hoped to hear, because the questions have been shaped by what is already being built, and the person being listened to can sense the direction and answers accordingly. The visit happened when it could still matter, which meant going when the schedule pressure made going feel like a distraction from the real work.

The second condition is directness. Going in person — not through a project manager or a user research agency — means sitting with the people, watching the work, hearing the words in the order and with the weight the speaker intended. There is a real role for structured user research in technology decisions. But there is something it cannot carry: the unprocessed signal that reaches a decision-maker only when the decision-maker is present to receive it.

The third condition is repetition. The first visit produced one set of things heard. The second produced different things, because the trust that had not existed in the first conversation had started to build. People answer differently when they know the person asking is genuinely interested rather than completing a process — and that knowledge does not form in a single visit.

In the third conversation, someone said something they had not said in the first two — not because they had withheld it, but because it had taken that long for the context to feel safe enough for honesty to reach what actually mattered. Single visits produce first answers. Repeated visits produce the ones underneath.

What was not in any document

The visit that mattered most was the third one.

The spreadsheet had come up in the first visit — the workaround every person on the team maintained alongside the current system, because the system could not hold notes in the right place at the right stage in the process. Three years in, nobody had mentioned it in any requirements session because it had become invisible through familiarity. The requirement was added on the way back.

The second visit changed a vendor criterion. Someone was working with three applications open simultaneously, constantly switching, returning to the system from a backgrounded state. The shortlisted vendors had been evaluated against their interfaces during structured demos — clean, navigable, well-presented.

The question of how each vendor’s product behaved when interrupted and resumed had never been asked. It was asked after that visit. Two vendors answered adequately. One did not. The shortlist changed.

What the third visit produced was different. Someone described the hardest part of their day — a step that happened immediately before they opened the system. Technically outside the system’s scope, but one the implementation could have absorbed if the handoff had been designed differently. That single observation reshaped the implementation in a way that addressed the friction at its source, rather than treating the system boundary as a fixed line. The first two visits had not produced it.

None of the changes were expensive. Each required someone to have been in the room. The requirements document that did not include the notes capability was not careless — it simply had no mechanism to surface what the direct conversation produced.

Together, the changes constituted a technology decision made on what was actually true, not on what the formal process had captured. Not complete information — that is never available. But closer to what was actually true than any document had been.

Three questions a leader cannot answer from data alone

Before a significant technology decision, there are questions that no amount of analysis will answer reliably without direct listening.

The first is whether the technology will genuinely help the people who most need it to work. Not the average user in the user story — the person whose day is hardest, whose work is most complex, whose need the requirements document has approximated but not named. That answer is in the room where the work happens. Direct listening — going there — is the only way to hear it. The leader who does not go will answer it from assumption, and the assumption will be derived from data simplified precisely where the difficulty is most real.

The second is whether the technology will genuinely add to what the people it serves can do — or substitute one set of constraints for another. The genuine gain in capability shows up in the specifics of what people find difficult and what they could do if that difficulty were removed. Those specifics are not in aggregated data. They emerged in the second visit, from a person describing a capability they currently lacked that the new system could provide if designed to carry it. That capability had not appeared in any document because no document had asked for it directly.

The third is what the system needs to adapt to — the actual range of people and circumstances it will encounter in use. The answer is not in the modal user’s profile. It is in the outlier’s experience, in the third conversation with the person who finally said the thing they had been circling. Going back is the only way to hear it. A leader who makes the decision without returning will design for the average and discover the outliers after the system is live, when adaptation costs considerably more.

My opinion

I have sat in enough rooms to know that the gap between the requirements document and what is actually needed is structural, not accidental. At one university, a single conversation around a table surfaced needs nobody had articulated — and capabilities some groups had no idea others could use. I have run procurement exercises where, over four days, real teams put real problems in front of vendors and evaluated how they responded. The vendors that won were not the ones with the best product. They were the ones that listened before they pitched.

Questions worth taking into your next decision

The four questions below are not rhetorical. They are worth answering — for yourself, before the next decision closes.

1. Think of the last significant technology decision you were involved in making. When did you last speak directly — not through a survey, not through a user research report, but in person or by phone — to someone who would be affected by that decision? What did you hear, and what did it change?

2. If you lead technology decisions, what is the mechanism in your process that brings direct, unmediated human experience to you before the strategy is fixed? If that mechanism does not exist — what is stopping you from creating it?

3. Think of a technology decision you made that did not land as intended. At what point in the process would direct listening have changed the direction? Was that opportunity available and missed, or was the process not designed to create it?

4. Is direct listening a luxury that busy technology leaders cannot afford — or is it the one thing that makes everything else more efficient by ensuring the direction is right before the resources are committed?

A McKinsey and University of Oxford study of large IT projects found that budget overruns averaged 45 per cent — but shortfalls in delivered value averaged 56 per cent. Those are not the same problem. The value loss lands quietly, across years of implementation, workaround, and eventual replacement.

Jordan’s listening did not guarantee the right direction. It made the wrong direction less likely — at a cost in time that is negligible against the cost of not going at all.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Sam walks the lower fields in winter before anything else. Not because it is efficient. Because it is the only way to know what the overnight rain has done — whether the water is moving through or pooling, whether the drainage that has been slowly improving for three years has held, whether the flood-prone corner that floods briefly and recovers quickly if left ungrazed for six weeks is beginning to saturate. This knowledge cannot be looked up. It was built over twelve years of walking this ground in all weathers, at all hours, through all seasons.

The soil sensor is in the barn. It has been there for fourteen months.

The phone would show no signal here even if they took it out. It has shown no signal on this part of the land for as long as they can remember. That is not a technical problem waiting to be solved. It is a characteristic of these forty acres, in this valley, in rural Wales. The land is what it is. They are doing what they have always done: feeling the soil, watching the drainage, reading what this particular ground is trying to say.

What the grant assumed

The precision agriculture package arrived two years ago with a government agri-environment grant attached and a marketing brochure that described “nature-inspired data systems designed to work with natural variation.” They spent three weeks setting it up: the soil moisture sensors pushed into the ground at the intervals the manual specified, the GPS-guided monitoring app configured on the tablet, the yield-mapping software calibrated against the field boundaries.

The sensors required persistent 4G signal to transmit data to the cloud platform. Signal on the lower fields drops to zero.

The app’s recommendations were calibrated for large-scale arable monocultures: uniform rows, single-variety plots, conventional seasonal rotation. This is a twelve-year permaculture polyculture — a food forest, a market garden, mixed livestock — designed around the specific drainage, soil, and microclimate of this ground. The app had no logic for any of it. The yield-mapping tool produced output that mapped onto nothing being grown there or recognisable on this land.

The problem was not the setup, or the interface. It was that the technology had been designed for a farm this land was not, and no amount of patient configuration was going to change that. Within eighteen months, the sensors were back in the barn and the grant had been spent.

The principle the industry never borrowed

Permaculture has a founding principle, stated plainly since Bill Mollison and David Holmgren developed the practice in 1978: observe and interact. You do not design until you know this place. You walk it across seasons. You watch what it does in rain and drought and frost. The design emerges from sustained attention to specific conditions — not from a template applied to all conditions.

The precision agriculture industry describes its products as nature-inspired. The language is in the brochures, the grant applications, the conference presentations. Adaptive systems. Intelligent sensors. Biomimetic design. What the industry has borrowed is the vocabulary. The founding principle — observe first, design second, never assume the average applies here — was left behind.

The reason is structural. Technology development is funded by scalability. A solution that works for eighty percent of farms is investable. A solution built from sustained observation of this specific land, and a different solution built from sustained observation of the farm next door, is not.

The observation step — the one that would reveal what this land actually needs — is designed out before a line of code is written. Scalable and specific are not the same thing.

In nature, there are no average organisms. Only specific ones, each the product of sustained adaptation to particular conditions. The precision agriculture platform was built on the logic that farms are variations on a theme. This land is not a variation. It is a place.

What twenty-five years of forecasts missed

In March 2026, a peer-reviewed study published in Agribusiness (Wiley) tracked twenty-six precision agriculture technologies across twenty-five years of adoption forecasting data. The researchers — Malone and colleagues — worked through the CropLife–Purdue Precision Agriculture Dealership Survey, the longest-running continuous assessment of precision agriculture technology in the United States. Their finding: input dealers overestimated adoption on nearly every technology, in nearly every year of the study. The overestimation gap widened in the early 2020s, as adoption on several technologies flattened or declined while dealers continued to project steady growth.

The industry’s explanation has been consistent: this is an adoption problem. Farmers are slow to change. They lack the skills, the confidence, the willingness to invest. Twenty-five years of the same explanation, in the face of twenty-five years of the same forecasting error, is not an explanation. It is a refusal to look at the other possibility.

The NFU’s 2025 connectivity survey puts a number on part of the gap. Only 22% of UK farmers have reliable mobile signal across their whole farm. Twenty-one percent of UK farms have broadband speeds below 10 Mbps — compared with a national average of less than one percent. The technology assumes infrastructure that the majority of small mixed farms in rural Britain do not have. These lower fields are not an outlier.

They are the norm.

The grant funding attached to the precision agriculture package does not ask whether the farm has 4G. It asks whether the farmer can demonstrate a commitment to sustainable intensification. They could. The application process had no field for “my lower fields have no signal,” any more than the monitoring app had a field for “food forest polyculture.” The system that allocated the grant and the system that failed to work were built on the same assumption: that the farm in front of them was a version of the average farm, and that any deviation from the average was the farmer’s problem to solve.

The kingfisher and the train

In the early 1990s, a Japanese engineer named Eiji Nakatsu was tasked with eliminating the sonic boom produced by what would become the Series 500 Shinkansen when it exited tunnels. The existing nose design was aerodynamically efficient in open air and catastrophically loud underground, where the pressure wave had nowhere to go. The engineer was a birdwatcher before they were a lead engineer, and the problem reminded them of something they had spent years watching: kingfishers moving from air into water without producing a pressure wave. The kingfisher’s beak is a precise solution to that exact problem — the transition between two media of different densities, at speed, without disturbing the surrounding medium.

The engineer did not spend a two-week research sprint watching kingfishers. They had been watching them for years, as a matter of personal attention rather than institutional programme. The insight was available because observation had been treated as real work, sustained over time, without a deliverable attached to it. The redesigned nose reduced the sonic boom, made the train 10% faster, and reduced electricity consumption by 15%.

The technology that arrived at this farm had borrowed nature’s language. It had not borrowed the engineer’s method. It had not spent seasons watching what these lower fields do after overnight rain. It had not asked what twelve years of permaculture observation had produced.

That accumulated knowledge — the only knowledge that could optimise this specific land — was invisible to the system’s design. It was not data. It was noise.

What observation asks of any system

The common thread between permaculture’s founding principle and the engineer’s decades of watching is not the sector or the technology. It is the decision to treat observation as real work, completed before design begins. The questions that follow are worth bringing to any system you deploy, inherit, or commission — because the answers will tell you whether the observation step was there, or whether it was designed out.

1. When a technology you use, or your organisation deploys, was designed — how much time was spent observing the specific conditions it would operate in, rather than the average conditions it was modelled on? Would you know if the answer was none?

2. What knowledge do you carry — about your work, your land, your patients, your tenants, your community — that exists only in your head and has no field in the system you’ve been given? What happens to a decision when that knowledge is absent?

3. If the technology you’re currently using was built for someone else — a different scale, different context, different connectivity — would the evidence be visible in how it performs, or only in what it cannot do?

4. Permaculture’s first principle has been available since 1978. What would technology design look like if it had the same founding requirement: that you do not design until you have observed?

Sam still walks the lower fields in winter. The soil sensor is in the barn. The land is doing what it has always done: varying, specific, unaverageable. The technology that was supposed to augment that knowledge had no knowledge of this place. It never did. It was built for somewhere else.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

For twenty-five years, the International Space Station has been humanity’s address in orbit — built by governments, run by governments, and costing more than any object ever constructed. Now a startup is building a replacement in a factory in California. The question is not whether the era of private space stations is coming. It is what that era is actually for.

The ISS and what comes next

In a factory in Long Beach, California, a team of engineers is building a space station. Not designing one. Not modelling one. Building one — machining aluminium panels, rolling and welding the pressure vessel that will contain breathable air, testing solar arrays and life-support systems. The station is called Haven-1. The company building it is called Vast. And if the schedule holds, a SpaceX Falcon 9 rocket will carry it to low Earth orbit sometime in early 2027.

Haven-1 is small. At 45 cubic metres of habitable volume — roughly the interior of a small tour bus — it is far from the football-field sprawl of the International Space Station, which took over a decade to assemble and cost upwards of $150 billion to build. It will host four astronauts at a time for missions of up to thirty days. Its operational life is three years. It does not try to be everything. It tries to prove that a private company can build and operate a human habitat in orbit — safely, on a realistic schedule, and without the resources of a national space agency.

That proof matters because the era it is trying to inaugurate has a deadline. NASA plans to deorbit the International Space Station around 2030. After twenty-five years as humanity’s permanent outpost in orbit — the most ambitious collaborative engineering project in history, shared by fifteen nations — the ISS will be guided out of the sky and into the Pacific Ocean. What replaces it, if anything does, will not be built by governments. It will be built by companies.

Whether that is exciting or alarming depends on what you think space is for.

The International Space Station was conceived as a symbol as much as a laboratory. Built in the years after the Cold War, assembling modules launched by American Space Shuttles and Russian Soyuz rockets, it expressed something about what cooperation between nations could achieve. The science it enabled was real and significant — advancing our understanding of human physiology in microgravity, producing pharmaceutical research that could not have been conducted on Earth, hosting thousands of experiments across biology, materials science, Earth observation and physics. But it was also, always, enormously expensive.

NASA spends roughly $3 billion per year to operate the ISS. That sum, compounded over the station’s lifetime, has represented a significant fraction of the agency’s total budget. The calculation NASA is now making is whether the science and presence in orbit justify that cost when private companies are willing to build and run orbital infrastructure and sell NASA access to it. The model shifts from landlord to tenant. Instead of owning and operating the station, NASA becomes a customer, buying crew time and research capacity from whoever builds the next generation of platforms. The $150 billion question is whether that works.

The ISS cost over $150 billion to build and costs $3 billion per year to run. NASA’s bet is that private companies can do this cheaper — and that competition will do what government monopoly could not.

Haven-1 is not the only contender. Axiom Space, which has already been awarded a NASA development contract, is launching a Payload Power Thermal Module to the ISS in early 2027; roughly nine months later that module will undock and rendezvous with Axiom’s habitat module to form an independent two-module station by early 2028. Starlab — a joint venture between Voyager Technologies and Airbus, with Northrop Grumman providing autonomous docking capability, and a design that will launch in a single flight aboard SpaceX’s Starship — is targeting 2028. Blue Origin’s Orbital Reef, billed as a “mixed-use business park in orbit,” is in development. Between 2026 and 2031, NASA plans to provide $1–1.5 billion in funded Space Act Agreements to support commercial station development — a programme it has recently revised, shifting from firm fixed-price contracts after officials concluded that a fully commercial business case does not yet exist in low Earth orbit.

The gap between the ISS’s retirement and the arrival of its successors is a genuine concern. NASA’s planning assumes overlap between commercial station operations starting in 2028 and ISS retirement in 2030. If commercial timelines slip — as they have already for Haven-1, which moved from a May 2026 target to Q1 2027 — that overlap narrows. Axiom Space chief executive Michael Suffredini has said publicly that the market may not be large enough for more than one commercial station. The company has faced funding difficulties. The question of whether the economic case for private space stations is as solid as the engineering ambition is one that has not yet been answered by the market.

What space is actually for

The commercial case for space stations rests on a set of claims about what can be done in orbit that cannot be done on Earth — and what that is worth. The most substantive of these claims concerns microgravity manufacturing.

Gravity shapes almost everything about how materials form and how biological processes operate. Protein crystals grown on Earth are distorted by gravity-driven convection; the same proteins grown in microgravity form larger, more perfect structures that reveal their molecular architecture more clearly. Drug development that depends on understanding a protein’s shape can benefit from having better data. Major pharmaceutical companies and JAXA have conducted crystallisation experiments on the ISS. One JAXA study of proteins associated with Duchenne muscular dystrophy led directly to a drug candidate — TAS-205 — now in Phase 3 clinical trials. California startup Varda launched a manufacturing satellite in June 2023, producing crystals of ritonavir — an HIV medication — in orbit and returning the capsule to Earth in February 2024.

Optical fibres grown in microgravity achieve purity levels impossible on Earth. Alloys and semiconductors manufactured without sedimentation can have more homogeneous microstructures. What changes with commercial stations is not the science. It is the access — more capacity, lower cost, broader range of customers, and the possibility of operating manufacturing processes at scale rather than as research demonstrations.

Space tourism is the other commercial pillar. Axiom Space has already sold private astronaut missions to the ISS, at prices reported to start around $55 million per person. Haven-1 will accommodate private visitors. Blue Origin’s Orbital Reef explicitly frames its design around commercial hospitality. The market here is narrower — it requires the kind of wealth that makes $55 million an acceptable discretionary expenditure — but it is real, and it serves the less glamorous purpose of underwriting operational costs that the research market alone may not cover.

The science done on the ISS has already produced pharmaceutical advances, materials breakthroughs and discoveries in human physiology that could not have been achieved on Earth. The question is whether the commercial model can sustain access to that environment after the ISS is gone.

The honest uncertainty is about the size and reliability of the non-government market. NASA’s contribution — buying crew time, hosting government science, funding development — is the anchor. If that anchor holds, the economics of commercial stations are plausible. If NASA’s budget is cut, or its priorities shift, or the agency decides it can achieve its goals without a permanent crewed outpost in low Earth orbit, the business cases of every company in this race become significantly harder.

Who inherits the sky

The ISS was many things, but one of them was political. The same hardware that carried American astronauts carried Russian cosmonauts, European scientists, Japanese engineers, Canadian robotics technologists. The station’s assembly required fifteen nations to agree on standards, interfaces, safety protocols and operational procedures. It survived the post-Cold War period, survived Shuttle disasters, survived the tensions that followed Russia’s 2014 invasion of Ukraine — though the latter ultimately prompted the partners to plan for Russian withdrawal and ISS retirement. Whatever its limitations as a scientific platform, it demonstrated that sustained human cooperation in space was possible.

The commercial stations being built now are American. They are funded by American venture capital and American government contracts. They will be operated under American regulatory frameworks and serve primarily American and allied institutional customers. China, which has built its own Tiangong station — permanently crewed since 2022 — is not a customer. The transition from government-operated to commercially-operated orbital infrastructure in the Western world is also, in a geopolitical sense, a transition from international to national. That is not necessarily wrong. But it is a change that deserves to be named.

The science also changes. The ISS was an instrument of national prestige and international diplomacy, which meant science happened on it even when the immediate commercial return was unclear. Commercial stations will prioritise the experiments customers pay for. That will include pharmaceutical companies crystallising proteins, universities running long-duration biology experiments, technology companies testing hardware in the radiation and vacuum of orbit. It will also mean that certain categories of basic science, whose value is not easily priced and whose benefits arrive decades later, will struggle to find a berth.

What the commercial model has to prove

The case for commercial stations is clear enough on paper. A research environment that currently costs $3 billion per year to run could be made more accessible, more competitively priced, and faster to iterate — if the market works as advertised. The pharmaceutical researcher, the materials scientist, the physician studying bone density in long-duration spaceflight, the engineer testing life-support systems for deep-space missions: all of them benefit if orbital access becomes cheaper and easier rather than harder and more expensive. The platform itself does not matter. What matters is whether the transition sustains the access.

The honest question is whether private operation adds something or merely substitutes one expensive monopoly for another. The ISS cost $3 billion per year under government management. The benchmark for commercial stations is not zero. It is whether, over time, the cost per hour of orbital research comes down, whether the range of customers expands beyond governments and billionaires, and whether the scientific output justifies the investment.

The harder question — the one the market cannot easily answer — is what happens to research that does not have an immediate commercial return. A drug candidate for a neglected tropical disease. A long-term study of bone loss that informs treatment for earthbound osteoporosis patients. A materials science experiment whose applications are twenty years away. These are the kinds of research that the ISS accommodated because governments decided they mattered. Whether commercial operators will find business models that serve them too is not yet clear.

Sources

Haven-1 and Vast

• Haven-1 — Wikipedia

• Vast Delays Haven-1 Launch to 2027 — Payload Space

• Vast Advances Haven-1 Into Integration Phase — Vast Space

Axiom Space

• Axiom Station — Wikipedia

• Vast and Axiom awarded new private missions to ISS — NASASpaceFlight

• Axiom Space Secures $350M in Financing — Axiom Space

• Axiom Space puts a brave face on claims of dire finances — The Register

• Axiom Space raises $350M backed by Qatar fund — CNBC

Starlab

• Starlab Space Station — Wikipedia

• Private Starlab space station moves into full-scale development — Space.com

• Voyager Space Announces Teaming Agreement with Northrop Grumman for Starlab — PR Newswire

NASA Commercial LEO Destinations programme

• Commercial Space Stations — NASA

• NASA releases details on revised next phase of commercial space station development — SpaceNews

• Payload Field Guide: Commercial LEO Destination — Payload Space

Varda Space / ritonavir

• HIV drugs manufactured in space return to Earth — Healio

• Return of the Ritonavir — Varda Space Industries

• Rocket Lab successfully returns spacecraft capsule to Earth — Rocket Lab

JAXA protein crystallisation / Duchenne muscular dystrophy

• Creating new and better drugs with protein crystal growth experiments on the ISS — NASA

• Creating New and Better Drugs with Protein Crystal Growth Experiments — SpaceNews

ISS context and commercial transition

• Commercial Efforts Flail to Replace Aging International Space Station — Engineering

  ---

  You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Alex is in the middle of a routine task when it happens. Not an epiphany — a small, specific thing. The step that used to require four form fields now requires one. Alex completes the task and then pauses, trying to remember when the change appeared, because it was not announced and it was not the subject of a release note and Alex only notices it now because the friction that used to be there is gone.

Somewhere in the background of that pause is a support call from eighteen months ago. Alex had reported exactly this — the unnecessary duplication in the form, the way the same information was asked for in three different places. It had taken fifteen minutes longer than it should have. Alex had said so. The call had ended with an apology and a reference number.

And now the form was different.

What usually happens to feedback

The standard experience of providing feedback to a technology product is an exercise in patience with diminishing returns. The survey that arrives after every interaction, rated one to five, with a final optional text box that almost no one fills in because nothing has ever clearly happened as a result of it. The support ticket that generates an automated acknowledgement, a reference number, and silence. The complaint that is logged, managed, and resolved — which means it is closed, not that the thing that caused it has been changed.

None of these are careless. They are systems doing what they were designed to do. The survey tracks satisfaction, the ticket resolves the case, the complaints process manages liability. None of them were designed to change the product. That distinction matters more than it sounds, because a system designed for one purpose will not accidentally perform another, however good its data.

The difference in Alex’s case was not that the organisation cared more. It was that the system had been designed for a different purpose: to move information from the person experiencing the product to the people building it, and to do something with what arrived.

The three conditions that make a loop

A genuine feedback loop breaks the moment any one of three conditions is missing.

• Feedback that works has to be easy to give — not a survey sent after the fact, not a formal process initiated at a separate URL, but a mechanism present at the moment the friction is felt. Alex did not fill in a survey. Alex called to report something broken, while it was broken, with the specific detail of what had gone wrong. The signal was clean because it was immediate and unsolicited — the kind of feedback that surfaces the actual experience rather than a retrospective reconstruction of it.

• Getting that signal to the people who can act on it is harder than it sounds. In most organisations, the path is long and obstructed: feedback arrives at customer service, which resolves individual cases and closes tickets, while patterns that would be visible to the product team — if they received the same data — never reach them. Alex’s organisation had a short path: a shared channel between support and engineering, a standing item in the sprint review for user-reported friction, a team that treated incoming complaints as product development input rather than cases to be managed. The information arrived with its context intact and at the people who could do something about it.

• Visibility is what closes the loop. Not announced, not marketed — just present. The next time Alex used the service, the thing that had been difficult was different. Without that, the person who provided the feedback has no reason to believe the loop exists and no reason to provide the next signal honestly. A feedback system that receives, processes, and acts — but never demonstrates that it has done so — produces exactly the same user behaviour over time as one that does none of those things. Trust requires evidence, not intention.

All three conditions were present in Alex’s case. Remove any one and the loop would have broken.

What two years of listening produces

Over two years, the service Alex uses has changed in small, specific ways that trace directly back to what people who used it said.

The file upload failure message used to say “Error: upload rejected.” Alex had reported this after twenty minutes trying to diagnose which of five files was causing the problem. The message now says which file failed, why it failed, and what the accepted formats and size limits are. The change took a morning to implement. The information needed to make it had been sitting in the support queue for eight months.

The address entry flow used to require four separate fields in an order that did not match the way most people read their own address. The most common support call in one quarter was people unsure which field their flat number belonged in. The form now uses postcode lookup. Completion time for that step fell by two-thirds.

A third change is smaller and easy to miss. The notification sent when a request was submitted used to say “Your request has been received.” Alex had written, in a feedback form, asking whether anyone actually looked at these requests — because the notification contained no information about what would happen next, or when, or who to contact in the meantime. The notification now includes an expected response time and a named contact for queries. It took one line of copy and the answer to a question that no one on the product team had thought to ask, because no one on the product team had ever waited on the other side of the process they had built.

The cumulative effect is a service that feels attended to. Not perfect — attended to. That quality is not cosmetic. It is the specific evidence that someone received what was said and acted on it, and it is the foundation of whatever trust Alex now places in the service.

What the design is actually for

1. Think of a digital service you use regularly. Can you point to a specific change that service has made in the past year that you believe was a response to real user experience? If you cannot — what does that tell you about whether the feedback loop exists?

2. If you build or manage a technology product, how does feedback from real users reach the people who can act on it? Is the path direct and short, or does it pass through layers that filter, aggregate, and delay?

3. If you have ever provided feedback to a technology product — a support call, a complaint, a survey response — were you ever aware of a change that resulted from it? How did that experience affect your relationship with the product and the organisation?

4. Continuous improvement is standard practice in manufacturing and in software development. Why is continuous listening — treating user feedback as an ongoing product development input rather than a post-launch compliance exercise — still the exception in most technology organisations?

Most feedback systems are designed to manage the organisation’s relationship with feedback rather than to let feedback drive the product. The organisations that reverse that design — that make the path from user experience to product change as short and unobstructed as possible — build technology that improves in the direction that matters. Alex noticed. That noticing is rarer than it should be.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Alex puts the phone face-down on the table while eating dinner. Not because there’s an argument brewing, not because anyone has asked. Just a small, deliberate act. The meal lasts forty minutes. At some point — it turned out to be fourteen minutes in — the phone gets picked up. There was no notification. Nothing rang. The impulse arrived with no identifiable origin, like a hand that reached before the mind could object.

It doesn’t feel like a problem. Just a habit. Something to work on. A matter of willpower.

That framing is exactly what the people who designed that impulse were counting on.

The story we were given

We have been taught to understand compulsive phone use as a failure of self-control. The language we reach for is personal: distracted, addicted, weak, unable to focus. Self-help books tell us to build better habits. App stores sell us tools to limit our screen time. The diagnosis is individual, and so the cure is individual.

This framing is not neutral. It is convenient — and it is convenient for a specific set of people.

The companies whose platforms you cannot stop checking did not produce compelling products by accident. They built something more precise: a compulsion. And they built it using the same mechanisms that keep a gambler at a slot machine at three in the morning, long past the point of enjoyment.

What the engineers understood

Variable reward schedules are the most reliably powerful mechanism in behavioural psychology. The principle is simple: when the timing and size of a reward is unpredictable, the brain cannot habituate to it. It cannot stop paying attention. It keeps seeking.

A vending machine that always delivers the same snack is satisfying but not compulsive. A slot machine that sometimes pays out, in unpredictable amounts, at unpredictable intervals, is something else entirely. The unpredictability is not a design flaw. It is the entire point.

Social platforms were redesigned around this principle deliberately. The pull-to-refresh gesture works like a slot machine lever — and what it delivers operates on the same principle: sometimes a funny post, sometimes a message from someone you love, sometimes nothing worth reading. The brain cannot tell in advance. So it keeps pulling.

The Hook Model — one of the most cited models in social platform design — describes the loop explicitly: trigger, action, variable reward, investment. It was not developed to help you get things done. It was developed to make sure you come back.

What the misdiagnosis costs

When we believe the problem is personal, we look for personal solutions. We download another app. We set a screen time limit. We move our phone to another room at night. Some of these things help, at the margins. None of them changes the underlying architecture.

The system benefits from our misdiagnosis. They want the conversation to remain about you.

The human cost is not abstract. Attention fragmented over years becomes harder to sustain. The capacity for deep work, for genuine presence in a conversation — these are not personality traits immune to outside pressure. They are capacities that can be eroded. And they are being eroded, deliberately, at scale.

Children who grow up inside these systems will have spent their formative years in an environment specifically engineered to prevent them from consolidating attention. The long-term consequences of that are only beginning to be understood.

Against the user

A system that manufactures compulsion rather than enabling choice is not assisting the person using it. The platform’s interest and the user’s interest diverged the moment engagement became the goal rather than the outcome of delivering something worth using.

These platforms amplify a biological vulnerability — the brain’s inability to habituate to variable rewards — and direct it toward commercial ends. That is not augmentation in any useful sense. It is a diminishment dressed as a feature.

The algorithm adapts, yes — but it adapts to exploit you more efficiently. It learns what keeps you engaged, not what serves you. These are different objectives. In most people’s lives, they are opposing ones.

A system that does not assist, does not genuinely add something, and does not adapt to serve is not a neutral tool that some people misuse. It is a system designed, tested, and refined to work against its users’ interests while appearing to serve them. That is where the responsibility lies.

The questions underneath the habit

• When you reach for your phone without knowing why, what exactly are you looking for?

• Has there been a time when you tried to reduce your phone use and found it harder than you expected? What story did you tell yourself about why?

• If you discovered that a product you used every day had been deliberately engineered to make it difficult for you to stop using it — what would you want to be different, and who would be responsible for making that change?

• Is the design of the technology in your life something you have ever actively chosen, or is it something that happened to you?

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

A case documented in the medical literature described a patient with pancreatic cancer who missed out on a potentially life-prolonging treatment because their scan result was delayed by eighteen months. The scan had been taken. The images existed. No one had looked at them in time.

That case was not unusual enough to be a scandal. It was unusual enough to make the journal. The background against which it sits — a radiology system under sustained and worsening pressure, in which delays are the norm rather than the exception — has become so familiar that it no longer registers as news. It is simply the condition patients navigate when something goes wrong inside their heads and they need to know what it is.

In September 2025, more than 386,000 people in England were waiting longer than six weeks for a diagnostic test. The total number waiting for any diagnostic test exceeded 1.7 million. The Royal College of Radiologists has documented a 30% shortfall of clinical radiologists in England — roughly 2,000 doctors short of what is needed to meet current demand — with the gap projected to reach 39% by 2029. Forty-six percent of NHS acute trusts are not meeting even the interim waiting time target.

A system that can read a brain MRI scan in seconds has arrived. Not a prototype. Not a laboratory demonstration. A system trained on more than 220,000 real clinical scans, tested across more than 30,000 studies in a running health system over a full year, published in February 2026 in Nature Biomedical Engineering. The question is not whether the technology works. It is why, given that it works, the gap between scan and diagnosis remains so wide for so many patients.

What the AI can actually do

The system is called Prima. It was developed by a team at the University of Michigan led by neurosurgeon Todd Hollon, and it works differently from most AI diagnostic tools that have preceded it.

Earlier AI models in radiology were built for specific tasks: detecting a particular kind of lesion, flagging a certain cancer type, screening for a defined abnormality. They were useful for the thing they were designed for and nothing else. Prima was trained on everything — every MRI study taken at the University of Michigan Health system since radiology records were digitised, covering more than 220,000 studies and 5.6 million imaging sequences. It was trained not just on the images themselves but on the patients’ clinical histories and the reasons each scan had been ordered by a physician. It works, as its developers describe it, the way a radiologist works: integrating what the image shows with what the patient’s medical history tells you.

Across more than 50 different radiological diagnoses — stroke, brain haemorrhage, tumours, multiple sclerosis, aneurysms and more — Prima outperformed other state-of-the-art AI models, achieving a mean diagnostic accuracy of 92% across that range. It can identify which cases require urgent attention and immediately alert the appropriate specialist: a stroke neurologist if the scan suggests a stroke, a neurosurgeon if it suggests a bleed. Feedback is available as soon as imaging is complete. The result that would otherwise take days or weeks to arrive takes seconds.

Prima was trained on every MRI taken at a major health system since records were digitised. It works the way a radiologist works — integrating image data with clinical history. The result takes seconds.

UCL and King’s College London published parallel work in December 2025. Their model, trained on 60,000 brain MRI scans and the radiology reports written about them, accurately distinguished normal from abnormal scans approximately 19 times out of 20. A randomised multicentre trial across UK hospitals started in 2026 to assess how the system performs in live clinical workflows. These are not the same research programme, but they point in the same direction: AI that can read a brain scan fast and flag what matters is operationally real, not theoretical.

For acute neurological conditions, stroke above all, the stakes of speed are measurable in the specific. In stroke care, the standard principle is time is brain: for every hour of untreated ischaemic stroke, roughly 120 million neurons are lost. The gap between scan and treatment decision is a gap in which irreversible damage accumulates. A system that compresses that gap from hours to seconds is not a marginal improvement in workflow efficiency. It is a different category of clinical event.

Why the gap persists

If the technology exists and the problem is severe, the natural question is why the gap between them remains so large. The answer is not simple and it is not flattering to any single part of the system.

Regulatory approval is one obstacle. AI diagnostic tools that make clinical decisions about individual patients are classified as medical devices in the UK and the EU. The approval process is rigorous — necessarily so, given the consequences of getting it wrong — but it is not designed for the pace at which AI research is producing results. The models that receive approval tend to be narrow-task systems, because narrow tasks are easier to validate; a foundation model like Prima, which makes diagnoses across 50 conditions, presents a different and more complex challenge. The UCL team’s next step — a randomised multicentre trial in UK hospitals — is the appropriate pathway, and it is one that takes years.

Integration is a second obstacle. An AI model that produces a result in seconds is only useful if that result reaches the clinician who needs it, in a format they can act on, connected to the patient record system they are already using. NHS radiology departments typically operate on legacy IT infrastructure. The connections between imaging systems, electronic patient records, and clinical communication tools are incomplete, inconsistent, and often not interoperable. The radiologist shortage means that even when AI flags a case as urgent, there may not be a specialist immediately available to review the flag and make the treatment decision.

Liability sits beneath both. When an AI system incorrectly categorises a scan — flags something urgent that is not, or misses something that is — the question of who is responsible is unresolved in most legal and professional frameworks. The EU AI Act, whose main provisions came into application in August 2026, classifies medical AI diagnostic tools as high-risk and requires documentation of training data, bias checks, and human oversight policies — though the provisions specifically covering CE-marked medical devices are not due to apply until 2027. Radiologists operating under professional accountability for their diagnoses have every reason to be cautious about delegating to a system whose error profile is not yet fully understood.

The technology compresses the gap from hours to seconds. The regulatory pathway, the IT infrastructure, and the liability framework compress it from years to months. That asymmetry is the problem.

The Royal College of Radiologists has been clear that AI represents one of the most significant changes in medical care since the inception of the NHS — and equally clear that far from replacing radiologists, AI creates new roles and responsibilities for them. The challenge is not the technology’s relationship with the profession. It is the system’s capacity to absorb a disruptive tool without the disruption falling, as it usually does, on the patients.

Who bears the cost of the gap

Between October 2024 and September 2025, an average of more than 500,000 people were waiting each month for a CT or MRI scan in England. That figure increased by nearly 19,000 per month compared to the previous year. More than 74,000 of those waiting longer than six weeks were waiting for a CT or MRI scan necessary to diagnose or rule out cancer and other time-critical conditions.

The people who bear the cost of the gap are not evenly distributed. Delays in imaging disproportionately affect patients in regions with the fewest radiologists, which are not the regions with the least need. The East of England had the highest proportion of patients waiting too long for a diagnostic test in September 2025. Rural and lower-income areas consistently face longer waits than urban centres with large teaching hospitals. The patients most likely to experience an eighteen-month wait for a scan result are not the patients with the most access to private imaging, second opinions, or the capacity to advocate loudly for themselves in a system under pressure.

The documented case of the pancreatic cancer patient is one example of a pattern that oncologists and radiologists describe consistently. One consultant oncologist, quoted in a Royal College of Radiologists report, put it plainly: delays in scans and treatment were resulting in missed or late cancer diagnoses, with some patients’ conditions deteriorating to a stage where treatment was no longer possible. These are not patients who would be well-served by a technology that exists in a Michigan laboratory and a multicentre trial and not yet in their district general hospital.

Sources

• Prima AI system: Hollon et al., “Learning neuroimaging models from health system-scale data”, Nature Biomedical Engineering, February 2026.
  https://www.nature.com/articles/s41551-025-01608-0

• Michigan Medicine press release: “An AI model that can read and diagnose a brain MRI in seconds.” https://www.michiganmedicine.org/health-lab/ai-model-can-read-and-diagnose-brain-mri-seconds

• UCL and King’s College London brain MRI model: UCL News: “AI model helps speed up diagnosis of brain disorders”, December 2025. https://www.ucl.ac.uk/news/2025/dec/ai-model-helps-speed-diagnosis-brain-disorders

• King’s College London: “AI brain scan model identifies stroke, brain tumours and aneurysms.” https://www.kcl.ac.uk/news/ai-brain-scan-model-identifies-stroke-brain-tumours-and-aneurysms

• NHS diagnostic waiting times (September 2025): Royal College of Radiologists: “Nearly half of NHS trusts missing test waiting time target as backlogs grow.”
  https://www.rcr.ac.uk/news-policy/latest-updates/nearly-half-of-nhs-trusts-missing-test-waiting-time-target-as-backlogs-grow/

• NHS Digital: “Diagnostic waiting times and activity for September 2025.”
  https://digital.nhs.uk/data-and-information/publications/statistical/nhse-diagnostics-waiting-times-and-activity-data/for-september-2025

• Radiologist workforce shortfall: Royal College of Radiologists: “2024 workforce census reports lay bare the challenges facing radiology and clinical oncology.”
  https://www.rcr.ac.uk/news-policy/latest-updates/2024-workforce-census-reports-lay-bare-the-challenges-facing-radiology-and-clinical-oncology/

• Stroke neuron loss: Saver JL, “Time Is Brain — Quantified”, Stroke, 2006.
  https://www.ahajournals.org/doi/10.1161/01.str.0000196957.55928.ab

• EU AI Act — medical device provisions: Kennedys Law: “The EU AI Act implementation timeline: understanding the next deadline for compliance”, 2026.
  https://www.kennedyslaw.com/en/thought-leadership/article/2026/the-eu-ai-act-implementation-timeline-understanding-the-next-deadline-for-compliance/

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Jordan noticed it on a Tuesday morning, about four months in. They had opened the app the way they opened other apps — without thinking about it, on the way to something else. The result was there when they looked: a number, with a short explanation of what it meant in relation to where it had been last week. Jordan read it in the time it takes to make coffee. There was nothing to decode.

That moment — a test result arriving with context, at a time they could actually do something with it — was not dramatic. It did not change the condition. Jordan has a condition that requires regular monitoring, and it will require monitoring indefinitely. But something had shifted in the daily experience of managing it. The technology had been built to serve that experience, not to process it. For the first time in years of navigating health systems, the difference was legible.

The system Jordan already knew

Before the app, the management of Jordan’s condition was a patchwork held together by their own effort. Test results arrived on paper, sometimes days after the event, with reference ranges printed beside the numbers but no interpretation of what those numbers meant in context. They went into a folder maintained on their own — disconnected from the records at the GP surgery, disconnected from the records at the specialist clinic. The folder existed because Jordan had created it. No one had suggested it; no system supported it.

Appointments were bookable, but the booking required persistence — the kind that is straightforward when you are well and difficult when you are managing a condition that affects concentration and energy. Jordan had learned which times to call, and which numbers to try when the first was engaged. These were not skills anyone had expected to need.

The condition itself was manageable. What was not manageable, not sustainably over years, was the administrative overhead the health system had attached to it. The paper trail that was entirely the individuals responsibility. The records that told a different story depending on which part of the system you were looking at. The appointments that were partly administrative events — catching up on information that should have been shared already, establishing a baseline that should have been continuous.

The ordinary cost of managing a long-term condition in a system not designed for the person managing it is a cost that most people carrying those conditions pay, quietly, without counting it as a cost. The honest context is that health technology has not, on the whole, served people in this situation well. Two thirds of NHS patients and carers reported at least one administrative problem in the past year, according to the King’s Fund in 2025 — disproportionately those managing long-term conditions. That history is part of why the difference, when it came, was so immediately recognisable.

The question they asked first

The team that built the app began with a question rather than a specification. Not: what should this app do? But: what does it feel like to manage this condition, and where does the friction concentrate?

Those two questions produce different software. The first produces a list of features. The second produces an understanding of what is actually hard about the problem — which is rarely what the first question reveals, and rarely what someone without the condition imagines from the outside.

The answer, when they gathered it, was specific. The friction was not in the absence of information but in its format and its timing. Information arrived too late, without context, through channels not connected to anything else. The friction was in the administration of care, not in the care itself. The app was designed to reduce that specific friction, in that specific form, for people in that situation.

What came next followed from that question. The design was tested throughout development with people who found digital tools genuinely difficult to use — not only with people who were already comfortable with health apps and quick to forgive poor design. The team was accountable, internally, for whether the people using it could actually manage their condition better — not only for whether the app was delivered within budget and adopted within target timescales. And the design was iterated after deployment, based on how people used the app in practice rather than how they had been expected to use it.

None of that requires a large budget or an unusually skilled team. It requires a set of priorities — a decision, made at the start and maintained throughout, that the person using the technology matters more than the efficiency of the organisation providing it. That decision is not always made. When it is not, the technology reflects the absence.

What actually changed

Jordan’s condition is no better or worse than it would otherwise have been. What is better is the experience of managing it, day to day, in the space between clinical appointments.

Test results that used to require interpretation now arrive pre-interpreted — with enough context to inform a decision, to notice a pattern, to decide whether something warrants a call before the next scheduled appointment. No more calls to the surgery to ask what a number means. The conversation with the clinical team is more useful because the patient arrives with information, and the team arrives without gaps to fill. What used to be a partly administrative appointment is now a clinical one.

The relationship with the clinical team has changed because the technology removed the noise that used to surround it. The paper trail is gone. The catch-up is gone. The time spent establishing what the records should already show — gone. That is not a small improvement in a clinical relationship. It is a different kind of care.

There is also more time. Time that used to go to navigating systems that were not working — chasing results, maintaining the folder, ringing back — is available for other things. The calculation is not large on any single occasion. Over years, it compounds.

None of this is dramatic. The condition is still there, still requires attention, still shapes daily life in ways that are not going to change. But the daily weight of managing it is lighter than it was, and that lightness is the direct result of a design team that started with actual circumstances rather than a generic idea of what a patient requires.

What the design actually did

The app does one thing well: manage this condition, in these circumstances, with these constraints. Not a generic patient managing a generic condition — a specific person, with the particular shape of their daily life. The help it provides is specific to where the friction actually is, rather than matched to what the average health app is assumed to need. That specificity is the direct consequence of the question they asked first.

The app adds something that was not there before. The person using it can now participate in their own care in a way that was previously impractical — reviewing their own data with context, preparing for clinical conversations with real information, communicating between appointments without the friction that made that communication more trouble than it was worth. The participation remains theirs. The app did not take over the management of the condition. It provided better tools to manage it.

That distinction, between technology that substitutes for a person’s agency and technology that extends it, was a design decision, not an accident. It was made at the start.

The app also responds to the individual. Over time it reflects what this person uses, finds useful, needs at different moments. The personalisation is not surface-level — it changes how the app functions rather than how it looks, making it more useful over time rather than more impressive in a product demo. That responsiveness requires a feedback loop between user and team — one that exists only because someone was accountable for outcomes. Which is, again, the decision made at the start.

Four questions for the next project

1. Think of a piece of technology — an app, a service, a system — that has genuinely helped you, in a way that felt specific to you rather than generic to all users. What made it different? Can you identify the design decision that produced that experience?

2. If you build or commission technology, how does your design process ensure that the person who will use it is the starting point rather than the endpoint? What specific mechanisms — genuine user research, inclusive testing, accountability for the human outcome — are in place to hold that priority throughout development, not only at the requirements stage?

3. What would it take, in your organisation or sector, to make conditions like these the standard rather than the exception? Not in five years — in the next project. What is the single first change that would make the most difference?

4. Is the gap between technology that burdens and technology that helps primarily a technical gap, a resource gap, or a priorities gap? And if it is a priorities gap, what does that tell us about who has the power to close it?

The conditions that produced an app like this are not secrets. They are known, replicable, and not especially expensive. The gap between what exists and what is possible is not technical. The reason it persists is a question for the people who make the priorities, not the people who write the code.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Alex submitted the third query through the portal on a Thursday afternoon, copying in the reference numbers from the first two. The housing benefit calculation still didn’t match what Alex had been told over the phone three months earlier. The difference was not large — roughly forty pounds a month — but over four months it had accumulated into an overpayment notice that Alex was now expected to repay.

Each of the three queries had been acknowledged within twenty-four hours. Each had been assigned a case reference. Each had received a response. The most recent response explained, in clear language, how housing benefit calculations worked. It did not explain why Alex’s calculation differed from the figure given on the phone. The case had been marked as resolved.

The week after Alex received the overpayment notice, the council’s customer services director presented the quarterly performance report. Digital channel adoption was up. Average response time was down to 1.4 working days. Cost per interaction had fallen 28%. Case resolution rates were running at 94%. The dashboard was green across every metric. Nobody in that meeting had Alex’s name in front of them.

The metrics the council was reporting were accurate. An average response time of 1.4 days is faster — measurably, demonstrably faster — than what came before. A 28% reduction in cost per interaction is a real achievement. A 94% case resolution rate sounds like a service that is working well.

What the dashboard doesn’t count

What none of those numbers captured: whether the question had been answered. The 94% resolution rate counted cases that had received a response and been closed. It did not count whether the response had resolved the issue. A case could be opened, acknowledged, responded to with a standard letter, and closed as resolved — three times — without the person ever getting what they needed. That is not a failure in the data. It is a failure in what the data was designed to count.

The calls that ended without resolution. The cases where the automated response was technically accurate and substantively irrelevant to the specific question asked. The people who received four replies and gave up on the fifth attempt. None of that appears in the performance report. Not because it was hidden, but because the system was never designed to see it.

The people running the service have no signal that anything is wrong. The people funding it have no reason to ask. The people using it — the ones for whom it is not working — have no channel to make the failure visible, except to keep submitting queries that will each receive a response and remain unresolved.

The politics of choosing what to count

This is where Goodhart’s Law applies — the principle that the economist Marilyn Strathern summarised as: when a measure becomes a target, it ceases to be a good measure. Response time was a reasonable proxy for service quality until it became a target. Once it was a target, the system optimised for it. Cases were responded to quickly. Whether the responses were useful was a different question — and not one the target required anyone to ask.

KPI structures for technology services are not purely technical documents. They are political ones. The metrics chosen reflect what the organisation is prepared to be accountable for, and organisations, rationally, choose metrics they can improve. Response time can be reduced through better routing and automation. Case closure rates can be improved by defining “resolution” in ways that are achievable. Cost per interaction falls as volume grows. All of these things can be improved without making the service more useful to the person at the end of it.

The human outcome — did the person get what they needed? — is harder to measure because it requires asking them. That takes time, raises uncomfortable questions, and produces data that is more difficult to present as success. So it is left out. Not maliciously. Structurally. The governance body that approves the metrics does not require it. The reporting cycle that reviews them does not miss it. And the dashboard that summarises them has no row for it.

There is also a governance dimension to this. The people who approve the KPI structure are presented with a set of metrics that look complete because they cover all the dimensions the previous governance model required: speed, volume, cost, compliance. Nobody in that room has ever been asked to approve a metric for whether the service actually resolved the person’s problem. Nobody has been told that such a metric is missing. The governance process validates what it is shown. What it is not shown does not fail — it simply does not exist in the conversation.

This is how a service comes to be celebrated and broken at the same time. The celebration is for what is measured. The brokenness is in what is not. The two things co-exist perfectly until someone submits a third query and receives a fourth standard letter.

The cost that doesn’t appear in the report

The pattern repeats across services. A case that the digital channel cannot resolve eventually gets escalated to a phone call — the kind of contact the digital system was built to reduce. The query is answered. The discrepancy is corrected. The outcome the portal was supposed to deliver arrives via the channel the portal was supposed to replace, weeks later than it should have.

The quarterly performance report will show digital channel adoption at its highest level. What it will not show: the volume of digital cases subsequently resolved by phone; the number of people who returned to a human channel because the digital one had not worked; or the real cost of a failed digital interaction, which includes not just the cost of the digital transaction but the cost of the phone call that followed it.

The service looks more efficient than it is. The measurement system has been designed in a way that makes that possible — by measuring the transactions that occurred rather than the outcomes they were supposed to produce. When a governance body reviews a performance report and sees green across every metric, the natural response is to keep doing what is working. The investment flows toward the metrics that are already performing well. The things that are failing silently receive nothing, because they are not visible as failures.

That is not a data quality problem. It is a design choice. The absence of an outcome metric is itself a decision, with consequences that accumulate quietly in every case the service was not built to see. And because the consequences are not measured, they do not generate pressure to change the measurement. The organisation that cannot see its own failures cannot feel the cost of them, and cannot direct resources toward addressing them. The green dashboard is not just a reporting artefact. It is a protection against accountability — not deliberate, but effective.

My opinion

I have watched organisations celebrate dashboards that tell them almost nothing about whether their services are working. The instinct to measure is right. The flaw is in what gets chosen. SLAs, KPIs, OKRs all share the same structural problem: they measure what the organisation can control. An automated acknowledgement meets the response SLA. The service still failed. I’ve witnessed a large outsourcing organisation hit their KPIs and SLAs every single month for response times…by putting an automated email response thanking the user for their query.

The organisations closing this gap are measuring outcomes, not transactions — asking users directly whether their problem was resolved. Some are going further, tracking social impact metrics: what actually changed for the person, beyond whether the case was closed. A number is a starting point, not a conclusion. Before accepting one, ask what human reality it represents. If you cannot answer that, the number is not yet telling you what you think it is.

The questions the system was never required to answer

The question a measurement structure should start with is whether the technology is actually helping the specific person using it — not whether it is processing their request efficiently. Those are different things. A system that processes twelve thousand requests per quarter and resolves sixty percent of them is not necessarily more helpful than one that processes eight thousand and resolves ninety. But a performance report that counts volume and response time will describe the first system as the better one.

The second question is whether the technology adds something real to the person’s capability, or only reduces costs for the organisation. A reduction in cost per interaction is not the same as an improvement in the quality of the interaction. Faster and cheaper are not synonyms for better. A measurement system that tracks only cost and speed cannot answer whether the technology is doing what it was built to do. It can only confirm that the technology is doing what it was built to do cheaply and quickly — which is a different question entirely, and a less important one.

The third question — and the one most completely absent from most KPI structures — is whether the system responds to the specific circumstances of the person in front of it. The 94% resolution rate is a population-level figure. It tells you nothing about the person in front of it. It tells you nothing about the six percent for whom the service did not work, or what those cases had in common, or whether they were disproportionately the people who needed the service most. A system that performs well on average while failing consistently at the edges is not performing well. It is performing adequately for people whose situations fit the expected pattern, and invisibly for everyone else. Those are not the same thing, and no dashboard that measures only the average will ever surface the difference.

The question behind the numbers

1. Think of a technology or service you use regularly — at work, or as a citizen or customer. What do you think is being measured about your experience of it? Does that match what you would want to be measured?

2. If you are responsible for reporting on technology performance in your organisation, when did you last ask whether the metrics you report capture the experience of the person at the end of the system — not just the performance of the system itself?

3. If you could add one metric to your organisation’s technology reporting — something currently unmeasured that would tell a more complete story — what would it be, and why is it not already there?

4. A system that hits every KPI while failing the people it serves: is that a success or a failure? And if the answer is complicated, what does that tell us about how we have defined success?

A service can hit every target on its dashboard and still fail the person who needed it. The performance report will describe it as operating at record efficiency. Both things are true. They are not the same truth, and the distance between them is where the accountability isn’t.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

It leaves no trace. No alerts fire. No systems misbehave. The attack, if it is happening, looks identical to normal network traffic. The adversary captures encrypted data — patient records, financial transactions, strategic communications, diplomatic cables — and stores it. Then they wait. They wait for a quantum computer capable of breaking the encryption to exist. When it does, they decrypt everything they collected. The breach happened years earlier. The victim had no idea.

This is called harvest now, decrypt later. It is not a theoretical future scenario. The US Department of Homeland Security, the UK’s National Cyber Security Centre, the European Union Agency for Cybersecurity, and the Australian Cyber Security Centre have all issued official guidance stating that adversaries are already doing this. State actors and advanced persistent threat groups are collecting encrypted data at scale, banking on quantum decryption becoming available within the next decade or two.

The good news, and there genuinely is good news, is that the solution exists and has been standardised. In August 2024, after an eight-year international evaluation process, the National Institute of Standards and Technology (NIST) published three post-quantum cryptography standards that organisations can implement today to protect their systems against quantum attacks. The standards are ready. The algorithms are tested and approved. The path is clear.

The bad news is that most organisations have not yet completed, or in many cases begun, their cryptographic inventories.

Why current encryption breaks

Almost all public-key cryptography — the kind that secures your bank login, your email, your medical records, and the connections between servers that underpin the internet — relies on the difficulty of certain mathematical problems. RSA encryption, which is used to protect the majority of internet traffic, depends on the fact that factoring the product of two very large prime numbers is computationally infeasible for classical computers. The numbers are so large, and the factoring process so slow, that it would take longer than the age of the universe for a classical machine to break a properly implemented RSA key. The security is not derived from complexity of design. It is derived from the limits of classical computation.

A quantum computer running an algorithm called Shor’s algorithm — published by Peter Shor in 1994 (SIAM Journal 1997), thirty years before any machine existed that could run it at useful scale — can factor large numbers exponentially faster than any classical approach. What takes a classical computer millions of years takes a quantum computer, at sufficient scale, hours. RSA breaks. Elliptic curve cryptography, which is used in everything from TLS web security to blockchain transactions, breaks. Diffie-Hellman key exchange, used to establish secure connections, breaks. The mathematical foundations that the entire architecture of digital security is built on become vulnerable.

The security of RSA encryption is not derived from complexity of design. It is derived from the limits of classical computation. A quantum computer changes those limits. When it does, the encryption breaks.

The timeline for when a quantum computer capable of doing this will exist is genuinely uncertain. The most optimistic estimates suggest the early 2030s. More cautious assessments say the mid-to-late 2030s. Some researchers place it later still. What is not uncertain is that quantum computing is advancing rapidly, that the below-threshold error correction result demonstrated by Google in December 2024 validates the theoretical path, and that once a cryptographically relevant quantum computer exists, breaking encryption will be a matter of running software, not of further hardware development.

The harvest now, decrypt later threat means this uncertainty is beside the point for data that needs to remain confidential for more than a decade. A diplomatic communication encrypted today that must remain secret for twenty years is already at risk — because if an adversary captures and stores it now, they do not need to decrypt it today. They need to decrypt it before the twenty years are up. Medical records, legal correspondence, intellectual property, strategic financial plans, national security communications — all of these have lifespans that could extend into the quantum era. The relevant question is not when quantum computers arrive. It is how long your data needs to remain protected.

The standards are ready. The migration is not.

NIST’s three finalised post-quantum cryptography standards — ML-KEM (for key encapsulation), ML-DSA (for digital signatures), and SLH-DSA (a hash-based signature scheme) — are built on mathematical problems that quantum computers cannot efficiently solve. They have undergone eight years of rigorous international evaluation, involving hundreds of cryptographers across academia, industry and government. They are not experimental. They are ready to implement.

The transition timeline is clear. NIST’s guidance calls for quantum-vulnerable algorithms to be deprecated and ultimately removed from its standards by 2035, with high-risk systems expected to migrate considerably earlier. The NSA’s Commercial National Security Algorithm Suite 2.0 framework requires quantum-safe algorithms for all new US national security systems by January 2027, full application migration by 2030, and complete infrastructure migration by 2035 — with compliance cascading through defence contractors, federal agencies and regulated industries. The UK’s NCSC has published equivalent guidance with target dates for migration activities. The European and Australian cybersecurity agencies have done the same.

The problem is execution. Identifying where quantum-vulnerable algorithms are used across an enterprise turns out to be surprisingly difficult. Cryptography is embedded in operating systems, in hardware, in libraries, in protocols, in custom-built applications, in supplier systems, in legacy infrastructure that predates anyone currently working at the organisation. The protocols most at risk — TLS, which secures most internet traffic; IKE, used in VPNs; S/MIME, used in email; code signing — depend on public-key cryptography and will all need to be migrated. Some of these migrations are straightforward software updates. Others require hardware replacement, which takes years and significant capital expenditure.

The complexity is compounded by interdependency. The first step in any migration is a cryptographic inventory: a complete map of every system, protocol and library that uses cryptography. Most organisations do not have one. Building it requires examining every layer of the technology stack, including components provided by third parties who may themselves not yet have assessed their quantum exposure. The Federal Reserve’s research division published analysis in 2025 examining the harvest now, decrypt later threat to distributed ledger networks, finding that blockchain transactions committed today cannot be retroactively re-encrypted — making the permanence that gives blockchains their value also their greatest vulnerability to quantum attack.

There is also a practical distinction about symmetric encryption. AES-256 and the SHA-3 family of hash functions are not broken by quantum computers — or rather, they are weakened but not broken, in a way that is addressed simply by using larger key sizes that are already standard practice. The algorithms that break catastrophically under quantum attack are the asymmetric, public-key schemes. Organisations that have already moved to AES-256 throughout are not in as exposed a position as those still running older symmetric standards. But the public-key infrastructure that secures their communications, authenticates their users and signs their code is still vulnerable.

The standards to protect against quantum attacks were finalised in August 2024. The first step is a cryptographic inventory — a complete map of every system using quantum-vulnerable algorithms. Most organisations do not have one.

The new post-quantum algorithms come with larger key sizes and signature sizes than their classical predecessors, and some introduce performance overhead. Migrating to ML-KEM and ML-DSA is not a drop-in replacement for RSA everywhere — it requires careful implementation, particularly in constrained environments like embedded firmware, Internet of Things devices, and legacy systems with limited processing capacity. The industry’s guidance is to begin with hybrid deployments where possible: running the new post-quantum algorithms alongside existing classical schemes, so that security is maintained against classical attacks during the transition and gradually shifted to quantum-resistant-only. This is more complex than a clean cutover, but it is the responsible approach for systems that cannot afford any gap in protection.

What is actually at risk

The data categories most exposed to the harvest now, decrypt later threat share one characteristic: they need to remain confidential for a long time. Government and military communications — diplomatic cables, intelligence assessments, operational plans — are the most frequently cited category, because state actors with the capability to harvest at scale are the most likely harvesters, and national security data has the longest confidentiality requirements. The exposure is direct and serious.

Healthcare data carries similar risks. Medical records are retained for decades and contain some of the most personal information people generate. A patient’s HIV status, mental health history, genetic conditions, reproductive choices — all encrypted in hospital systems today, all potentially sitting in an adversary’s storage waiting for the tools to read them. The same applies to legal communications protected by attorney-client privilege, to research data underpinning pharmaceutical competitive advantage, to financial records containing the kind of material non-public information that could move markets if exposed years later.

The most alarming aspect of harvest now, decrypt later attacks is that they leave no visible signs. No files are corrupted. No systems are disrupted. No ransom note arrives. The breach is silent, and the damage is deferred. An organisation can be entirely unaware that its most sensitive communications from the past decade have been collected and are sitting in an archive waiting for decryption. By the time the damage becomes visible — when a quantum computer makes the archive readable — it is too late to protect the data that was already taken.

This is not a risk profile that fits the usual mental model of a cyber attack. Most security investment, most incident response planning, most board-level conversation about cyber risk is oriented around attacks that are visible and immediate. Harvest now, decrypt later is invisible and patient. The discipline it requires is different: not reactive, but anticipatory. The investment it requires must be made before any damage occurs, against a threat whose timing is uncertain, for a payoff that consists of nothing happening.

Who gets left exposed

What is not yet clear is who gets there first, and who is left last. The largest technology companies, banks and government agencies are already in the process of migration. What is not yet certain is whether the organisations that hold the most sensitive data about ordinary people — healthcare providers, legal systems, social service agencies — have the resources, the expertise, and the urgency to migrate before the threat becomes critical. A well-resourced adversary harvesting encrypted data at scale is not particularly interested in the records of major academic medical centres, which will likely be early adopters. It is interested in the records of district hospitals, GP practices and national health registries, which may be last.

Mosca’s Theorem, named after cryptographer Michele Mosca of the University of Waterloo’s Institute for Quantum Computing, states the problem precisely: if X is the number of years your data must remain secure, Y is the time required to complete your migration, and Z is the time before a cryptographically relevant quantum computer exists, then you must start migrating before X + Y exceeds Z. Every year of delay reduces the margin. For organisations with long-lived sensitive data, the margin may already be thin.

An organisation that has not yet begun its cryptographic inventory does not know which of its systems are most exposed. That inventory — a complete map of every system, protocol and library that uses cryptography — is the first step NIST, the NCSC, and every cybersecurity agency has now recommended. Identify the systems. Assess which data has the longest confidentiality requirement. Prioritise in order of exposure and sensitivity. Do not wait for a quantum computer to exist before beginning.

Sources

Standards and official guidance

• NIST Post-Quantum Cryptography Standards (August 2024): ML-KEM (FIPS 203), ML-DSA (FIPS 204), SLH-DSA (FIPS 205) https://csrc.nist.gov/projects/post-quantum-cryptography

• NIST — HQC selected as fifth algorithm (March 2025): https://www.nist.gov/news-events/news/2025/03/nist-selects-hqc-fifth-algorithm-post-quantum-encryption

• NIST IR 8547 (Initial Public Draft, November 2024): Transition to Post-Quantum Cryptography Standards — deprecation timeline to 2035 https://csrc.nist.gov/pubs/ir/8547/ipd

• NSA — Commercial National Security Algorithm Suite 2.0 (CNSA 2.0): Quantum-safe deadlines for national security systems; January 2027 procurement gate https://media.defense.gov/2022/Sep/07/2003071836/-1/-1/0/CSICNSA_2.0_FAQ.PDF

• UK National Cyber Security Centre — Timelines for migration to post-quantum cryptography (March 2025):Three-phase roadmap: assessment by 2028; prioritised upgrades 2028–2031; full migration by 2035 https://www.ncsc.gov.uk/guidance/pqc-migration-timelines

• US Department of Homeland Security — Post-Quantum Cryptography: Official DHS guidance including harvest now, decrypt later threat confirmation https://www.dhs.gov/quantum

• ENISA — Post-Quantum Cryptography: Current State and Quantum Mitigation: EU Agency for Cybersecurity guidance; EU coordinated roadmap targets high-risk systems by 2030, full transition by 2035 https://www.enisa.europa.eu/publications/post-quantum-cryptography-current-state-and-quantum-mitigation

• Australian Cyber Security Centre — Post-Quantum Cryptography: Official ACSC guidance on harvest now, decrypt later threat https://www.cyber.gov.au/resources-business-and-government/maintaining-devices-and-systems/system-hardening-and-administration/post-quantum-cryptography

Research and primary sources

• Shor, P. (1994 / 1997) — Polynomial-Time Algorithms for Prime Factorization and Discrete Logarithms on a Quantum Computer: Note: first presented at FOCS 1994 (the date used in the article); the formal journal paper was published in SIAM Journal on Computing, 1997. https://arxiv.org/abs/quant-ph/9508027 SIAM Journal version: https://epubs.siam.org/doi/10.1137/S0097539795293172

• Google Quantum AI — Willow processor announcement (December 2024): Below-threshold error correction result — first quantum chip to demonstrate error rates that improve as qubit count scales https://blog.google/innovation-and-ai/technology/research/google-willow-quantum-chip/

• Federal Reserve Finance and Economics Discussion Series 2025-093: “Harvest Now, Decrypt Later” — analysis of harvest threat to distributed ledger networks; blockchain immutability as a specific vulnerability https://www.federalreserve.gov/econres/feds/files/2025093pap.pdf

• Michele Mosca — Mosca’s Theorem: Co-founder and Deputy Director, Institute for Quantum Computing, University of Waterloo https://uwaterloo.ca/institute-for-quantum-computing/

Market data

• MarketsandMarkets — Post-Quantum Cryptography Market (2025): PQC market projected to reach $2.84bn by 2030, CAGR 46.2% https://www.marketsandmarkets.com/PressReleases/post-quantum-cryptography.asp

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

The agenda ran to eleven items. Security was seventh. By the time the board reached it, forty minutes had been spent on a supplier contract and a facilities question about the car park. The CISO had twenty minutes. They used eighteen of them.

The security review recommended twelve measures. The board approved three, deferred four, and excluded five as beyond the year’s plan. The meeting ended. Someone went to get coffee. Nine months later, one of the deferred items was the entry point for a ransomware attack that shut the business down for eleven days.

Nobody in that room had been negligent. Nobody had ignored the risk. The decision was made by competent people using the information they had, inside a process designed to balance competing priorities.

The CISO’s brief was forty-two slides. The executive summary was four pages. The risk register mapped each of the twelve recommendations to a likelihood score and a financial impact range. It was thorough. It was also built to satisfy a compliance requirement, not to help a non-technical board member understand which of those twelve things actually mattered most, and why, and what the difference would look like if they got it wrong.

Item seven on the deferred list was a patch management protocol for legacy systems. The brief described it as “medium risk, estimated remediation cost £18,000, recommended timeline Q2.” It sat between a network segmentation proposal and an employee phishing simulation programme. The board chair asked whether it could wait until the budget review. The CISO said it could be managed. The item was deferred.

That sentence — “it can be managed” — actually means: I believe the risk is tolerable in the short term, given everything else on the table. It does not mean: this is safe. The distinction between those two things disappears in the room. It stays disappeared until something goes wrong.

Organisations have been told for twenty years that security is a board-level responsibility. The intention was good. Bring it out of the IT department. Get it in front of decision-makers. Give the CISO a seat at the table. According to DSIT surveys, only 27 to 31 per cent of UK businesses have a board member with specific responsibility for cybersecurity — and this after a decade of sustained regulatory pressure to change it.

What happened instead is that security became a governance exercise. A risk register. A compliance checklist. A set of items to be approved, deferred, or removed, using the same mental framework that the board uses to evaluate a marketing spend or a property lease. The question is always: what does this cost, what does it protect, and can it wait?

The catastrophic failure of that framework is that cybersecurity risk does not behave like the risks that framework was built for. A lease that is deferred costs rent. A security measure that is deferred creates a window. The attacker does not wait for the budget cycle.

The deferred patch was not a theoretical vulnerability. It was a known gap, documented, scored, and filed. It sat in a folder. The attacker found a different route to it than the folder.

The decision that enabled the attack was made in a Monday morning meeting that ended with someone going to get coffee.

Eleven days offline. Not a dramatic number until you translate it. Supply chain halted. Customer orders unmet. Staff unable to work. The IR firm billing £2,400 a day. The ransom demand. The legal fees. The forensic audit. The regulatory notification. The reputational cost that does not appear in any spreadsheet.

The CISO did not lose their job. The board did not face personal liability. The organisation recovered, at significant cost, and updated the risk register. The patch management protocol was implemented in full. It had cost £18,000 on the original recommendation. The recovery cost many times that.

The people in the meeting were not careless. They were doing exactly what the system asked them to do. The system asked them to balance risks against budgets in a room built for a different kind of risk. That is not a character failure. That is a design failure.

What the brief was actually built for

A forty-two slide deck translated into a four-page summary with a risk register does not help a non-technical director make a well-informed decision about which deferred item will cost them thirty-eight times the remediation price. It helps the CISO demonstrate that they have fulfilled their duty of care. Those are not the same thing. The board member who approved the deferral did not understand the difference between “medium risk” and “a known gap in a legacy system with active exploit code circulating in the wild.” The brief did not help them understand it.

The governance structure was built to add something to security oversight — documentation, audit trail, the appearance of rigour. What it does not reliably add is security, because security requires that decisions reflect the actual threat environment rather than the budgetary one. A compliance framework and a threat actor are operating on entirely different timelines. The register records what was known. The attacker acts on what was left open.

The risk assessment treated every organisation of similar size and sector as roughly equivalent. It did not adapt to this company’s specific exposure: which of their legacy systems was most likely to be targeted, what the attacker’s likely entry path would be, what the actual blast radius of a successful attack on that specific infrastructure would look like. A one-size-fits-all risk score applied to a specific organisation on a specific day in a specific sector is not risk management. It is risk documentation.

The problem is not that boards are incompetent or that CISOs are failing. The problem is that the governance architecture built to manage cybersecurity risk was designed for a world where risk is relatively static and consequences are proportionate. Cybersecurity risk is neither. The architecture needs to change, not the people trying to work inside it.

My opinion

The more we rely on digital services, the more we place ourselves in the hands of organisations whose governance structures were not built to protect us. CISOs are operating in conditions no single role should carry. My position is that security risk belongs at the top of every board agenda, treated as a living state rather than an annual exercise — because the threat changes every day. Every new service deployed without security built in from day one creates exposure that did not exist the day before. Secure by design is not a technical principle — it is a condition of doing business.

The questions the risk register doesn’t raise

• Have you ever been in a meeting where a security recommendation was deferred for cost or time reasons? Not because anyone thought it was unimportant — but because the table was full and the budget was finite and it seemed manageable for now?

• If you have sat in a room like that one: what would it have taken for the conversation to go differently? What would the CISO have needed to say, or show, or do, to change the outcome?

• Should the people who make security deferral decisions face personal accountability when those decisions lead to a breach? And if the answer is yes — what would that accountability look like in practice, and who would apply it?

• The gap between the risk register and the real threat environment is not a secret. The people who write the registers know it. The question is whether the governance structures we have built are capable of closing it — or whether they were designed for something else entirely.

The dynamics in that boardroom — the compressed agenda, the deferred risk, the gap between documented exposure and lived consequence — run through The Shadow System. The book takes the argument further than a single meeting.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Somewhere beneath the Vijf Eikentunnel in Rijen, a railway underpass in the Netherlands, concrete is doing something concrete has never done before. A crack has formed — the kind that appears in any heavily loaded structure over time, the kind that, left alone, lets water in and corrosion begin and the slow degradation that ends, years or decades later, in a repair bill or a closure or a failure. Except this concrete is not waiting to be repaired. Dormant bacteria embedded in the mix have been activated by the water entering the crack. They are feeding on calcium-based nutrients and producing limestone. The crack is sealing itself.

The idea that a material could detect its own damage and fix it without any external intervention sounds like a property borrowed from living organisms. That is not a coincidence. The research behind self-healing materials draws explicitly on the biology of healing — on how skin closes a wound, how bone knits after a fracture, how trees seal damage to their bark. The scientific ambition is to replicate those capabilities in the inert materials that hold the built world together: concrete, steel, polymers, coatings.

The built world needs this badly. Corrosion and material degradation cost the global economy an estimated $2.5 trillion every year — roughly 3.4% of global GDP, according to a 2016 NACE International study, the most comprehensive global assessment of corrosion costs undertaken. That figure covers the direct costs of repair and replacement and a portion of the indirect costs of disruption, but not the full human cost of infrastructure failures, delayed maintenance, or the carbon embedded in the concrete and steel that gets thrown away and rebuilt. The scale of the problem is so large and so distributed that it rarely registers as a single issue. Cracks in bridges, rust in pipelines, spalling on car park decks, moisture ingress in tunnel linings — these are experienced as individual nuisances, not as a systemic failure of how we build things.

Self-healing materials are a systemic answer to a systemic problem. The question is how close they are to being usable at scale, and what it would take to get them there.

How materials learn to heal

There are two distinct approaches to making a material that can repair itself, and the distinction matters for understanding what each approach can and cannot do.

The first approach is extrinsic: healing agents are stored within the material and released when damage occurs. The most widely studied version of this embeds microcapsules — tiny containers filled with a liquid resin, a polymer, or a mineral-producing compound — directly into the material during manufacture. When a crack propagates through the material, it ruptures the capsules along its path. The healing agent flows into the crack, reacts with a catalyst or with the surrounding material, and solidifies. The crack is sealed. The material continues to function. The bacteria-based concrete works on a similar principle, with dormant bacterial spores taking the place of chemical capsules: water activates the bacteria, the bacteria produce calcite crystals, the crystals seal the gap.

The second approach is intrinsic: the molecular structure of the material itself contains bonds that can break and reform. When damage occurs at the microscopic level, these dynamic bonds — hydrogen bonds, reversible covalent bonds, supramolecular interactions — release and reconnect, the material flows slightly toward the damaged region and the damage heals. This approach works well in polymers and elastomers, where molecular mobility is sufficient to allow the necessary flow. It does not, in general, work in rigid materials like concrete or steel, where the molecular architecture does not allow that kind of movement.

Dormant bacteria embedded in concrete activate when water enters a crack. They produce limestone crystals that seal the gap. The material fixes itself without any human intervention. This is not a laboratory demonstration. It has been built into real structures.

The most advanced real-world deployments of self-healing materials are in concrete, and the most credible research has come from Delft University of Technology in the Netherlands, where professor Henk Jonkers and his team developed the bacteria-based approach over more than a decade. The world’s first building constructed with bacteria-based self-healing concrete was a lifeguard station by a lake in the Netherlands, a demonstration project built by the Delft group. Two further full-scale demonstrator projects followed.

Basilisk, the commercial spinout from the Delft research, expanded deployment of its bacteria-based healing agent into large-scale infrastructure projects in 2025. These are not controlled laboratory experiments. They are structures in the field.

The laboratory results are substantive. For cementitious mixes with inorganic self-healing admixtures designed to seal cracks up to 0.3 millimetres wide, permeability-based healing indices — a measure of how effectively the crack has been sealed — regularly exceed 90% at 28 days. Superabsorbent polymer systems tested against 0.3-millimetre cracks lowered total water flow through the crack by up to 98%.

Beyond concrete, self-healing polymer coatings on car surfaces can repair minor scratches using near-infrared light from sunlight. Self-healing tyres — which use a soft polymer layer that flows into punctures and reseals them — are available commercially from several major manufacturers. Self-healing polymer electrolytes for lithium-ion batteries are an active area of research, addressing the micro-cracking of electrodes that limits battery longevity.

The gap between the laboratory and the bridge

The science works. The engineering challenges that stand between a working laboratory material and wide deployment in real infrastructure are substantial but not unsolvable, and the most important one is cost.

Self-healing concrete costs more to produce than conventional concrete. The bacteria, the encapsulating agents, the additional admixtures — all of these add to the price per cubic metre. The premium varies by system and scale, but it is real, and it creates an immediate procurement problem. Infrastructure projects are costed on initial build price, not on lifecycle cost. A bridge specified with self-healing concrete costs more upfront. The maintenance savings accrue over decades, to a different budget, potentially under a different political administration. The person signing the initial contract and the person who would benefit from lower repair bills twenty years later are usually not the same person, and often not the same organisation.

This is not a technical problem. It is a procurement and accounting problem. Whole-life costing — evaluating infrastructure based on the total cost over its operational lifetime rather than just the initial build price — is standard guidance in most developed countries’ infrastructure policy frameworks. It is applied inconsistently. The Treasury Green Book in the UK explicitly requires whole-life cost assessment for public projects. Modelling of US concrete lifecycle emissions found that a 50% increase in structural longevity would have reduced associated CO₂ emissions by an estimated 14% over the twentieth century. Self-healing concrete that extends maintenance intervals well beyond that would compound the reduction further. A bridge built for a 50-year maintenance cycle rather than a 25-year one does not just save money. It reduces the carbon embedded in repeated repair cycles, the disruption of lane closures, and the probability of a failure caused by a crack that was never found.

The second constraint is crack width. Bacteria-based and capsule-based self-healing systems work best on cracks up to around 0.3 to 0.5 millimetres wide. Beyond that threshold, the healing agent cannot bridge the gap effectively. This matters because the cracks that most threaten structural integrity are not always small ones. A mature concrete structure that has been under load for thirty years may develop cracks that exceed what current self-healing systems can address. The technology is most powerful as a preventive intervention — stopping small cracks from becoming large ones — not as a remedial one for structures that are already significantly damaged.

The third constraint is one-time use. Most extrinsic self-healing systems — the capsule-based and bacteria-based approaches — can heal a given location once. The capsules rupture and are gone. The bacteria are consumed. If a crack opens again at the same location, the healing capacity has been depleted. Intrinsic polymer systems, where dynamic molecular bonds reform after each damage event, can in principle heal the same location multiple times. Research into multi-cycle concrete healing systems is ongoing, but the one-time-use limitation of the most commercially advanced systems is a real constraint on where and how they are deployed.

What infrastructure could look like

The consequences of widespread adoption of self-healing materials are not just financial, though the financial case is real. Cement production is responsible for approximately 8% of global CO₂ emissions — one of the largest single industrial sources of greenhouse gases. A consequential driver of cement consumption is the repair and replacement of structures that have degraded faster than their design life suggested. Self-healing concrete that doubles the effective maintenance interval of a structure halves, roughly, the cement demand associated with keeping that structure operational.

Infrastructure degradation disproportionately affects communities where the capacity to fund maintenance is limited — rural road networks, ageing urban bridges in lower-income municipalities, water and sewage infrastructure in countries where regular inspection and repair is not operationally feasible. A pipeline that monitors and seals its own small leaks does not require the inspection regime that a conventional pipeline does. A bridge deck that arrests crack growth autonomously does not generate the maintenance costs that have accumulated as structural degradation.

In England and Wales, a 2025 industry assessment found that 17% of the local road network is in poor condition — structural deterioration that has built up because maintenance funding has not kept pace with deterioration. That gap is not primarily a funding problem. It is also a materials problem.

The field is also moving rapidly beyond concrete into applications with different but equally consequential implications. Self-healing coatings on offshore wind turbines would reduce maintenance costs for one of the most expensive and difficult-to-access classes of infrastructure. Self-healing insulation on undersea power cables would extend their operational life considerably. Self-healing polymer housings for electronic devices would reduce the rate at which functional devices are discarded due to physical damage — one contributor to the 62 million tonnes of electronic waste generated globally in 2022, a figure rising by 2.6 million tonnes every year, according to the Global E-waste Monitor 2024. Self-healing battery electrodes, if they reach sufficient maturity, could extend the calendar life of electric vehicle batteries, reducing the extraction demand for lithium and cobalt that currently constrains sustainable EV adoption.

Cement production causes roughly 8% of global CO₂ emissions. A consequential part of that is repair and replacement of degraded structures. Self-healing materials that double maintenance intervals could cut that demand substantially.

What it would take to build differently

Self-healing materials help structures do their job for longer, with less human intervention. The bridge that arrests its own crack growth reduces the workload on engineers and maintenance teams whose schedules currently scale with every year of infrastructure age. The pipeline that seals its own pinhole leaks reduces the deployment of crews to locate and repair them. The scale of what needs doing is large, and the technology for at least some of the most important use cases has cleared its proof-of-concept stage.

Whether self-healing materials extend what infrastructure can do — allowing structures to be built in environments or to lifespans that would otherwise be impractical — or whether they are primarily cost-reduction tools for existing practice, the honest answer is both. A marine structure with self-healing concrete can be specified for a longer service life in a more aggressive environment than one without it. A pipeline coating that repairs micro-abrasions can be run in conditions that would quickly degrade conventional coatings. These are real expansions of what is buildable. But much of the near-term value will come from the more straightforward task of making ordinary infrastructure last longer and cost less to maintain.

What the industry has not yet seriously addressed is whether self-healing systems should be deployed uniformly at all. At present, they are — the same bacteria concentration, the same capsule density, regardless of where the highest damage risk sits in a given structure. A bridge deck sees very different stress profiles across its span. A pipeline coating experiences different corrosion rates at different locations based on soil chemistry and moisture. A system that modulated its healing capacity to match the risk profile of the structure — higher concentration in the highest-risk zones, lower elsewhere — would do more with the same material. This is technically feasible in principle and not yet routinely implemented in practice.

What the procurement system doesn’t capture

1. Infrastructure procurement in most countries evaluates initial cost, not lifecycle cost — even where policy frameworks require whole-life assessment. Self-healing materials are more expensive upfront and cheaper over time. What would it take to change the procurement incentive, and who would need to make that change?

2. Corrosion and material degradation cost $2.5 trillion per year globally. That figure is distributed across thousands of individual failures, repair contracts, and maintenance budgets. Is there a version of infrastructure ownership — public, private, or hybrid — that better captures the value of materials that last longer?

3. The communities with the most degraded infrastructure are often the ones with the least capacity to pay for repairs. Self-healing materials are currently priced as a premium product. What would a deployment model look like that directed this technology toward the places with the highest need, rather than the highest willingness to pay?

4. Self-healing materials borrow the concept of biological healing and apply it to inert matter. The research at Delft started with a question: if bone can heal, why can’t concrete? What other properties of living systems — adaptation, response to damage, distributed sensing — are materials scientists currently trying to replicate, and which of those would you most want to see applied to infrastructure?

My opinion

Electric vehicles are heavier than the combustion cars they replace — sometimes by several hundred kilograms — and roads specified for the last generation of vehicle are already absorbing loads beyond their original design. The UK road network, in many places, looks like potholes held together by asphalt. That is not a funding problem in isolation. It is also a materials problem, and one that becomes more urgent with every year that the vehicle fleet changes. Self-healing materials are not a long-horizon technology. They exist, they work, and they are cheaper than the alternative of letting the infrastructure fall further behind.

The lifeguard station by the lake in the Netherlands is still standing. The self-healing concrete in its foundations has been doing its work, quietly, since it was built — sealing the small cracks before they became large ones, without any human action, without any inspection finding them first.

That is a plain description of a consequential idea. Infrastructure that degrades slowly and repairs itself is not just cheaper to maintain. It is safer, more resilient, and less carbon-intensive. The hard part is not the science. It is getting the science out of the demonstration project and into the specification.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Casey stood in the kitchen on a Sunday morning wanting to do something small and

ordinary. The lights were already on — the right level, the right temperature, the system having made the calculation sometime before waking. The heating was already at the temperature the house knew a Sunday morning required. The coffee machine had already started its cycle.

Not one of these things had been consciously chosen that morning.

The intent had been to make a deliberate choice about the day — to begin it with a moment of decision rather than discovery. To turn the lights up brighter than usual, perhaps, because it was overcast outside and the mood called for it. To make the coffee later than normal because there was no urgency. Small acts of self-determination in an ordinary morning.

The system had already decided. It had observed enough previous Sunday mornings to know what this one required. The preferences were not wrong — they were, in fact, accurate. They reflected genuine past behaviour, real choices made on real mornings over months. The system had learned well.

That was the problem.

The disappearance of friction

There is a version of this story that reads as convenience. The house anticipates. The morning runs smoothly. That is what the technology was designed to do, and it is doing it.

What that version of the story does not account for is what friction was for.

The small acts of choosing — the light level, the temperature, the pace of a morning — are not merely mechanical. They are the continuous practice of preference: the low-stakes exercise of deciding who you are and what you want on any given day. They are how a person remains in contact with their own inclinations. When those acts are automated away, the automation does not just remove effort. It removes the opportunity to notice whether the choice would have been the same.

Casey’s house had learned a version of Casey built from past behaviour. Not from yesterday’s behaviour, from the cumulative pattern of months. A version that was accurate in aggregate and potentially wrong in any specific instance. The system had no mechanism to register that people change, that moods shift, that a Sunday in January and a Sunday in March might call for different things even if the previous fifty Sundays had called for the same ones. The actual person was no longer quite required.

The moment of choosing had disappeared so gradually that it was only noticeable in its absence. Not when the system got it wrong, it rarely got it wrong in any obvious way. But when the desire arose to do something differently, and the system had already done it, and the act of undoing it felt like more effort than simply accepting what was already there.

That is the trap. Not coercion, convenience so complete it becomes its own constraint.

From the kitchen to everywhere

The smart home is the most intimate version of this pattern, because the home is the space where people are most fully themselves. But it is only the beginning of the scale.

Step outward from the kitchen and the same architecture is running at every level. The streaming platform that has watched enough viewing history to know what will be selected before the selection is made. The news feed that has modelled enough reading behaviour to decide what is worth knowing today. Each of these systems is doing the same thing as the smart home, in a different register: replacing the act of choosing with the delivery of a prediction so accurate that choosing no longer feels necessary.

Step further out. The e-commerce platform that has learned spending patterns well enough to surface only what will be bought — not what might be wanted, not what might be surprising, but what the model says is coming next. The navigation app that routes not by the driver’s preference but by the algorithm’s assessment of the optimal path, such that after a year of use a driver may no longer know the roads they travel every day. The workplace tool that pre-populates responses, pre-schedules meetings, pre-drafts the email — reducing the human contribution to the act of approval.

At each step the friction removed is real. The convenience is not manufactured, it is delivered. And at each step something else is happening simultaneously: the system’s model of the person is deepening, the person’s practice of self-determination is narrowing, and the gap between who the system thinks they are and who they actually are is quietly widening.

The model is built from what was. The person is living in what is. Those two things are not the same, and the longer the system runs without the person having reason to notice the divergence, the larger the gap becomes.

Step to the outermost scale. The political information environment, the social media feed, the search result page, all of them running the same logic at civilisational scope. Not just anticipating individual preferences but shaping them, because a preference exercised without friction, without encounter with the unexpected, without the resistance of a world that does not already agree, is a preference that stops developing. The system does not just reflect who you are. Over time, it shapes who you remain.

The distinction nobody offered

The question of how much anticipation is enough is not rhetorical. It is a real dilemma.

Friction is not inherently good. The effort of choosing is not valuable in itself, it depends entirely on what is being chosen and why. The friction of adjusting the thermostat is not the same as the friction of deciding what to read or what to believe or what kind of person to become. Not all convenience is erosion. The task is to distinguish between the friction that was serving a purpose and the friction that was simply in the way, and to recognise that those two things look identical until the moment of removal.

Most people have not made that distinction deliberately, because it was never offered to them as a choice. The anticipatory systems arrived and removed friction and the question of which friction was worth keeping was never asked. The default was anticipation. The person who wanted to preserve the act of choosing had to actively resist the convenience of not choosing, and resistance is itself a form of friction that most people do not have time or inclination to apply.

The deeper consequence is not about convenience it is about the model. Every anticipatory system builds a representation of the person it serves. That representation is the basis for every subsequent prediction, every pre-emptive action, every choice made on the person’s behalf. The richer and more accurate the model, the more completely it can substitute for the person’s actual judgment. And the more completely it substitutes, the less opportunity there is for the person’s actual judgment to develop, correct, or surprise itself.

A person who has never had to choose their own route no longer knows the roads — a 2024 study confirmed this is not a metaphor: habitual GPS use produces a measurable, causal decline in spatial memory, even in people who previously navigated well. A person whose reading is pre-selected no longer encounters what they did not know they were missing. A person whose home has already decided never fully inhabits the decision to be at home. The capacity that is not exercised does not remain dormant. It diminishes.

That is the cost that does not appear in the product specification. That is the human loss that convenience optimises away.

Who is this for

The question the anticipatory system never asks, because it was not designed to, is who benefits from the anticipation.

The person using the system receives convenience that is real. But the system is not running in their service alone. Every preference recorded, every choice predicted, every behaviour anticipated is data. Data that trains the model, refines the prediction, deepens the platform’s understanding of what the person will do next. The more accurate the anticipation, the more valuable the person becomes as a data source and the more dependent they become on a system whose primary obligation is not to them.

The smart home manufacturer knows the daily rhythms of the household. The streaming platform knows attention patterns more precisely than the viewer knows themselves. None of this knowledge is held in the person’s interest by default. It is held in the platform’s interest, governed by the platform’s terms, subject to the platform’s commercial decisions about how that knowledge is used, shared, or sold.

The anticipatory trap is not just about agency. It is about the terms on which that agency has been surrendered and to whom. Casey’s house is learning Casey — but who owns what has been learned, and what happens to it when the subscription ends, the company is acquired, the terms are changed? The system was presented as a tool serving the household. The data flowing out of it tells a different story about who is serving whom.

The system that anticipates your needs perfectly has built a more accurate model of you than you have of yourself. That is impressive. It is also alarming, not because the anticipation is wrong, but because the entity holding that model is not you, does not answer to you, and did not ask your permission to build it in quite the way it has.

The friction that was removed was not just effort. Some of it was the resistance that kept you in contact with your own preferences. Some of it was the randomness that kept you encountering the unexpected. And some of it was the inconvenience that made you notice, every now and then, that a choice was being made at all.

What it was replaced with is not neutral. It serves someone. The question is whether you know who.

What you might find you’ve forgotten

Think about the systems in your daily life that anticipate your preferences — the home, the phone, the feed, the platform. When did you last make a deliberate choice in a domain where anticipation now operates? Not overriding a suggestion — making an original decision, from scratch, without a pre-populated option in front of you?

The model these systems hold of you was built from your past behaviour. How accurate is it to who you are now? Have you changed in ways the system has not registered? Have your preferences shifted, your interests expanded, your circumstances moved, and has the anticipation moved with you, or is it still running on data from a version of you that no longer quite exists?

Who holds the model? Not in terms of which company’s name is on the product, in terms of what happens to what has been learned about you. Is it portable? Is it deletable? Is it yours, in any meaningful sense? Or has the convenience of the service been partly funded by the permanent transfer of something you did not consciously offer?

And the question underneath all of these: if the anticipation were removed tomorrow — if the systems stopped predicting and simply waited for instruction — what would you find you had forgotten how to choose?

My opinion

The technology itself is not the problem. What makes ambient systems genuinely difficult is their invisibility. A tool you can see, you can use deliberately. A system that operates below the level of conscious attention has already made the choice about when and how it operates. The line between the two is not technical — it is a design decision, made by someone other than you, before you arrived.

Most ambient technology was not designed to be controlling. It was designed to be easy. But easy and controlling share a direction of travel. The easier a system makes a decision, the less often you will make that decision yourself — and at some point along that gradient, the decision stops being yours. It stopped leaving room for it.

Not everyone needs this. That sounds obvious but it is rarely said plainly. There are households where ambient technology solves a real problem — accessibility, health monitoring, the household stretched across too many responsibilities to manage routine decisions without help. In those circumstances it is doing what it was built for. As a default for everyone, replacing the ordinary friction of a morning with a prediction of what the morning should be, it is a different proposition — one that takes something from the quality of ordinary life that is hard to name until it is gone.

The question is not whether to use these systems. It is whether you know what you are giving in exchange. Convenience offered at no visible cost is rarely free. What is given in return is usually the data that makes the system more accurate, and the habit of not deciding that makes it more necessary. Both transfers happen gradually, without announcement, and by the time they are noticeable, reversing them requires more effort than most people are willing to apply on a Tuesday morning.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Morgan opens the app on a whim. It is one of those data-transparency tools that have become briefly fashionable — the kind that pulls together what advertisers, brokers, and platforms believe about you and shows you a summary. Morgan is curious, not anxious. It takes about thirty seconds to load.

What comes back is strange. The profile is accurate in the way a caricature is accurate — it captures something recognisable while getting most of the important things wrong. Morgan is categorised as a “high-anxiety urban professional,” likely to respond to time-pressure messaging and financially cautious. There are interest tags: travel, wellness, mid-range electronics. A risk band. An inferred household income. Morgan reads it twice, unsure whether to laugh or feel uneasy. Neither feels quite right.

When behaviour becomes identity

The discomfort is worth paying attention to. Not because the data is false — most of it is close enough to be recognisable — but because it is being treated as complete. The profile was assembled from clicks and purchases and search patterns, and somewhere along the way it stopped being a description of behaviour and became a verdict about a person.

That shift happens quietly. A system designed to predict preferences begins to define them. A tool built to segment audiences starts to assign identities. Nobody decided this was the purpose. It emerged from the logic of optimisation, from the pressure to know more in order to sell more, to lend more wisely, to assess risk at scale. The profile is a side effect of a process that was never really about the person at all.

The profile is accurate about behaviour and wrong about the person. The danger is not the bad data — it is that the system believes data is enough.

A trace treated as a portrait

Here is the underlying problem: data captures what someone did, in a particular moment, under particular circumstances. It does not capture why. It cannot hold the week someone was caring for a sick parent and spending on delivery food because there was no time to cook, or the period of job uncertainty that explains a cautious financial pattern, or a single impulsive holiday purchase that rewrote the travel interest score. The data is a trace. The system reads it as a portrait.

This matters because these profiles are not just sitting in a database somewhere. They are active. They shape what someone is offered, what rate they are quoted, sometimes what opportunities appear and which do not. A mortgage application, a job screening, an insurance premium — in each of these, the statistical shadow is treated as the person. And the person has almost no practical way to correct it, because most of the time they have no idea it exists.

The portability of these profiles makes it worse. An assumption made in one context follows into another. A risk score generated for one purpose bleeds into decisions it was never designed to inform. The data travels; the context does not. What began as a description of what someone did last Tuesday is now being used to decide what they deserve next year.

What the profile decides

The human cost here is not always dramatic. Sometimes it is a loan offered at a higher rate than someone with an identical financial history but a different data footprint would receive. Sometimes it is a job application that does not progress past the screening layer. Sometimes it is simply the low-grade unease of being looked at by systems that cannot see you, being assessed by processes you cannot access, being defined by a version of yourself that you did not author and cannot revise.

For those who are already disadvantaged — people in financial precarity, people with gaps in their histories, people whose lives do not fit the patterns the models were trained on — the distortion is not low-grade. It is consequential. The system is not malicious. It is simply indifferent to the difference between a data point and a human being. That indifference is its own kind of harm.

What the design is actually doing

In the original use case — recommendations, relevant offers, faster credit decisions — there is a genuine service on offer. But it holds only while the profile is accurate, and only while the person has some say in what it contains. Where the profile is wrong, or where it follows someone into contexts where it has no business being, the service becomes a constraint.

The efficiency gains for the institutions using these systems are real. What has been added for the individual is less clear: a permanent, portable verdict about who someone probably is. Whether that addition has value depends entirely on who is doing the measuring — and for whom the system was designed.

Where it fails most visibly is in adaptation. The profile is static in the ways that matter most. It captures behaviour but not situation. It updates when you click, but it does not reset when your life changes. It does not know the difference between who you were and who you are, and it has no mechanism for registering who you are trying to become. A system that cannot hold that kind of ambiguity is not adaptive. It is just persistent.

This is one of the questions at the centre of The Cognitive Crucible — not just what the system records, but what it misses by design.

My opinion

We are living through a period of digital identity fragmentation. Pieces of us — clicks, purchases, search patterns — are being gathered by third parties and reassembled into a composite. The problem is not just the assembly. It is the premise: that this composite represents a person.

It does not. We are not one thing. We are many, depending on context — different at work and at home, under pressure and at ease, at a particular age, in a particular relationship. A system that flattens all of that into a single profile is not describing us. It is describing a version of us that has never existed.

Taking back agency over this will be hard. It may require something counter-intuitive: choosing, deliberately, to define and manage our own digital identities — profiles that represent genuine aspects of who we are in a specific context, rather than ceding that definition to systems that were never designed with us in mind.

The questions underneath the discomfort

There is a reasonable question underneath the discomfort the opening described. Not “how do I get off this list” — that is almost certainly impossible — but something more fundamental: who is this profile actually for?

The profile was not built for the person. It was built for the systems that process them. That distinction changes the nature of the conversation. The question is not whether the data is accurate. The question is whether data should be sufficient.

1. Have you ever encountered a version of yourself through a system — a recommendation, a decision, a risk band — that felt both recognisable and completely wrong? What did you do with that feeling?

2. If you could see your full data profile across every platform and institution that holds one, what do you think you would find — and would you want to know?

3. If you could change one thing about how these systems work — how they are built, what they are allowed to infer, how they can be challenged — what would it be?

4. The system knows your clicks. Should that be enough to know you? And if not — what else should be required before a system is allowed to make decisions about your life?

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

The town of Wylfa sits on the northern tip of Anglesey, a Welsh island connected to the mainland by two bridges and a reputation for nuclear power. The original Wylfa power station opened in 1971 and ran for more than four decades, supplying electricity to households and businesses across the region. When it closed in 2015, the local economy lost something difficult to replace — not just jobs, but a particular kind of work: skilled, well-paid, long-term.

Plans for a new nuclear plant at Wylfa have circled since before the original station was switched off — Horizon Nuclear Power, then Hitachi, then a succession of other investors, none of whom made it past the preliminary stages. In June 2025, something different: the UK government selected Rolls-Royce SMR as its preferred partner to build the country’s first small modular reactors. By April 2026, a commercial contract had been signed and Wylfa confirmed as the site for up to three SMR units. If the timeline holds, the grid connection comes in the mid-2030s.

That “if” is doing a lot of work. The nuclear industry has a long and painful history of timelines that do not hold, costs that do not stay contained, and promises that outlast the political will required to keep them. The people of Wylfa — and the communities near every proposed SMR site — have reason to ask hard questions before deciding whether this time is genuinely different.

The honest answer is: it might be. And that is a more interesting answer than either the enthusiasts or the sceptics want to give.

Why nuclear has a second chance

To understand what small modular reactors are trying to solve, it helps to understand what broke the large ones.

Nuclear power has three advantages over almost every other energy source that are not marginal. It produces virtually no carbon emissions during operation. It generates electricity around the clock, regardless of whether the wind blows or the sun shines. And its energy density is extraordinary — a single kilogram of uranium fuel contains roughly two million times the energy of a kilogram of coal. These are the reasons nuclear provides about a tenth of the world’s electricity despite decades of political difficulty and public mistrust.

The problem is construction. Large nuclear plants take fifteen years to build in Western countries. The most recent to complete in the United States, the Vogtle plant in Georgia, came in seven years late and roughly $22 billion over its original $14 billion budget. The Hinkley Point C plant currently under construction in Somerset — the UK’s first new nuclear station in a generation — has seen its estimated cost rise from £18 billion when contracts were signed in 2016 to £35 billion in equivalent terms, approaching £49 billion at current prices. These are not outliers. They are the pattern.

The causes are structural: one-off bespoke designs that require learning everything from scratch on each new site; an atrophied supply chain that lost its skills during the decades-long pause in construction; and regulatory frameworks designed around 1970s technology that treat every new application as if nothing has been learned since. The result is that nuclear power, which ought to be the most reliable weapon in the decarbonisation arsenal, has spent thirty years being too expensive and too slow to be deployable at scale.

Nuclear’s advantages are extraordinary. Its construction record in the West is, on the evidence, one of the most expensive and difficult in modern engineering.

Small modular reactors are an attempt to escape that trap by changing the nature of the construction problem. Rather than building a bespoke cathedral on each site, the idea is to build a reactor in a factory — standardised, modular, replicable — and ship it to where it’s needed. A conventional reactor might generate 1,000 to 1,600 megawatts of electricity. An SMR typically produces between 50 and 470 megawatts per unit. You can add more units as demand grows, and the first unit can begin generating revenue while subsequent ones are still being assembled.

The factory model matters for cost. Nuclear’s expense problem stems partly from the fact that large reactors have never been built in sufficient numbers to drive down costs through learning and repetition. The steel and concrete and specialist components are expensive partly because the people and processes that handle them do so rarely. A factory model — producing the same design, repeatedly, under controlled conditions — is how aviation and automotive manufacturing drove down costs over decades. It is, at least in theory, how nuclear might do the same.

Where the theory meets the track record

There is a problem with the factory model, and it is not a small one: it has not yet produced a nuclear reactor on time or on budget, anywhere in the world.

As of early 2026, three SMRs have been built globally — one in China, two in Russia — plus one still under construction in Argentina. None were built on time. Russia’s floating plant, the Akademik Lomonosov, began construction in 2007 and reached operation only in 2019, at roughly three times its original budget. Argentina’s CAREM reactor broke ground in 2014 with a planned completion of 2017 and was still incomplete as of 2026, construction halted twice by funding crises, with costs significantly above initial estimates. China’s pebble-bed reactor at Shidaowan took approximately eleven years from construction start to commercial operation. These are first-of-a-kind projects, and some overrun is expected when building genuinely new technology. But the pattern of delay holds across different countries, different designs, and different political contexts.

The most instructive cautionary tale is NuScale Power. The American company was the first to receive US Nuclear Regulatory Commission design certification for an SMR, achieving that milestone in 2022 — a genuine achievement after years of rigorous review. NuScale appeared to be on the cusp of proving that the new model worked. Then, in 2023, its flagship project — a planned plant in Idaho — was cancelled. The project began with an estimated cost of around $3.6 billion for a 720-megawatt plant. Before cancellation, with the design rescaled and costs revised, that figure had risen to $9.3 billion. The customers, a consortium of public utilities, could not make the economics work. NuScale has since received regulatory approval for an uprated 77-megawatt design and is pursuing new projects, but the Idaho cancellation cast a long shadow over the industry’s cost claims.

The sceptical case against SMRs is not that the physics is wrong — it is that the economics may be structurally difficult. Smaller reactors lose the cost advantages that come with scale: building one 1,000-megawatt plant is cheaper per unit of output than building four 250-megawatt ones, because you only need one set of site preparation, one control room, one regulatory approval process. A 2025 study by Kim and Macfarlane — the latter a former chair of the US Nuclear Regulatory Commission — found that most cost projections for SMRs assumed high-volume serial manufacturing that has never materialised in nuclear, and that more realistic estimates placed the cost of electricity well above $100 per megawatt-hour, in some cases significantly more. For context, onshore wind in the UK currently costs around £40–50 per megawatt-hour.

There is also the waste question. Some analyses suggest that SMRs could produce greater volumes of radioactive waste per unit of energy generated than large conventional reactors — a consequence of their smaller scale and, in some advanced designs, the use of different fuel cycles. This is contested by the industry, but it is not a settled matter, and it matters to any community being asked to accept a reactor nearby.

None of this means SMRs cannot work. It means the confident promises being made — by governments, by technology companies, by investors — are running ahead of the evidence. The International Energy Agency’s scenarios have commercial SMR deployment beginning around 2030 — meaningful on paper, but, in the context of a global electricity system measured in thousands of gigawatts, still a rounding error in the near term.

What hangs on getting this right

The stakes are high enough that the question of whether SMRs work is not abstract.

Renewable energy — wind and solar — has become extraordinarily cheap and is now the cheapest source of new electricity generation in most of the world. But it is intermittent. The sun does not always shine; the wind does not always blow. Storing renewable energy at grid scale, for days or weeks at a time, remains an unsolved problem at the volumes the transition requires. This is sometimes called the baseload problem: the need for power sources that generate electricity continuously, regardless of weather, to underpin a grid that cannot afford to go dark.

Nuclear is one of the very few zero-carbon technologies that addresses the baseload problem directly. It generates power around the clock, in any weather, for decades. Hydropower does the same, but its geography is fixed. Large-scale geothermal is similarly constrained. If the world is serious about reaching net zero while keeping the lights on reliably, the honest arithmetic suggests that some form of nuclear will be part of the answer — and that SMRs, if they can deliver on their promise, could be a significant part.

The tech sector has reached a version of this conclusion independently. Amazon has committed over $500 million to SMR development. Google has partnered with Kairos Power to bring 500 megawatts of SMR capacity online by 2035, with the first unit targeting 2030, to power its data centres. Microsoft has signed agreements in the same space. The irony is pointed: AI’s insatiable demand for always-on power is one of the forces driving investment in nuclear energy’s revival.

The geopolitical dimension adds a further layer. Russia and China have both invested heavily in nuclear technology as a strategic export — Russia through Rosatom, which has construction contracts in over twenty countries; China through its state-owned enterprises. The UK-US Atlantic Partnership for Advanced Nuclear Energy, signed in September 2025, explicitly includes joint safety assessments and a shared commitment to eliminate dependence on Russian nuclear fuel by 2028. The race to build credible Western SMR supply chains is not purely a climate story. It is also a question of whose technology runs the world’s energy infrastructure in the decades ahead.

Against this backdrop, the communities living near proposed SMR sites find themselves in a familiar position — asked to accept proximity to technology that carries real and perceived risks, on the promise of jobs and cheap energy, with a track record that gives them every reason to ask for guarantees that are difficult to give. Research published in Energy Research & Social Science in 2025 put it plainly: the burdens and benefits of nuclear transitions are rarely distributed fairly, and existing frameworks to address this are inadequate. Wylfa’s community has been promised a new nuclear future before. The question is whether this version of the promise is better-designed, or just better-marketed.

What the technology still owes

At its best, an SMR helps communities, grids and industries access clean, reliable power that they cannot get from renewables alone. The UK government’s £2.5 billion commitment, the US government’s backing of multiple designs, and the private investment pouring in from the tech sector all reflect a real assessment that this technology could help solve a real problem. That is not spin. It is the outcome of serious people looking hard at the arithmetic of net zero.

Whether it adds something genuinely new is a sharper question. The factory model is the central claim — that modular construction changes the economics of nuclear in the same way that modular construction changed the economics of consumer electronics and automotive manufacturing. The honest answer is that this claim is plausible but unproven at commercial scale. The first Western SMR to be built on time and on budget will be the most important data point in energy policy this decade.

The question most consistently overlooked is whether this technology responds to the specific circumstances of the communities and contexts it serves — to their energy needs, their employment profiles, their risk tolerances, their histories with the industry. The SMR industry’s governance frameworks, on the evidence of current proposals, are not yet doing this. Regulatory processes are improving — the US ADVANCE Act of 2024 directed the NRC to streamline its approach, and by the end of 2025 it had met 30 of its 36 planned deliverables. But the frameworks for distributing the benefits and managing the disruptions of deployment — particularly for the communities most directly affected — remain underdeveloped.

The people of Wylfa have skills, knowledge and a relationship with nuclear power that spans generations. They are not a passive audience for a technology decision made elsewhere. Whether the SMR programme being planned for their island treats them as participants or subjects will say something important about whether the industry has learned anything from its history.

Sources

• Wylfa power station: UK Government, GOV.UK: “Wylfa closes after almost 45 years” — https://www.gov.uk/government/news/wylfa-closes-after-almost-45-years

• ANS Nuclear Newswire, 10 June 2025: “UK’s own Rolls-Royce wins SMR competition” — https://www.ans.org/news/2025-06-10/article-7102/uks-own-rollsroyce-wins-smr-competition/

• NucNet: “UK Picks Rolls-Royce For Domestic Small Modular Reactor Rollout” — https://www.nucnet.org/news/uk-picks-rolls-royce-for-domestic-small-modular-reactor-rollout-6-2-2025

• New Civil Engineer, 13 April 2026: “Rolls-Royce SMR secures Wylfa contract and £599M government loan” — https://www.newcivilengineer.com/latest/rolls-royce-smr-secures-wylfa-contract-and-599m-government-loan-13-04-2026/

• Neutron Bytes, 19 April 2026: “UK & Rolls Royce Sign Deal for Three SMRs at Wylfa” — https://neutronbytes.com/2026/04/19/uk-rolls-royce-sign-deal-for-three-smrs-at-wylfa/

• IEA, Nuclear Power page (nuclear generated 9% of global electricity in 2024): https://www.iea.org/energy-system/electricity/nuclear-power

• IEA, Global Energy Review 2025: https://www.iea.org/reports/global-energy-review-2025/electricity

• Utility Dive, 2026: “After 2 years, ratepayer pain and political fallout from Georgia’s nuclear plant Vogtle” — https://www.utilitydive.com/news/after-2-years-ratepayer-pain-political-fallout-georgia-nuclear-vogtle/817792/

• The Current GA, May 2026: “Two Years After Completion, Plant Vogtle Still Looms Over the Nuclear Debate” — https://thecurrentga.org/2026/05/12/two-years-after-completion-plant-vogtle-still-looms-over-the-nuclear-debate/

• EnergyTransition.org, April 2026: “The billion-dollar boondoggle: how Vogtle became the US’s monument to nuclear folly” — https://energytransition.org/2026/04/the-billion-dollar-boondoggle-how-vogtle-became-the-uss-monument-to-nuclear-folly/

• New Civil Engineer, February 2026: “Hinkley Point C’s cost climbs to £35bn with confirmation Unit 1 will power up in 2030” — https://www.newcivilengineer.com/latest/hinkley-point-cs-cost-climbs-to-35bn-with-confirmation-unit-1-will-power-up-in-2030-20-02-2026/

• Construction Wave, February 2026: “Hinkley Point C costs approach £49bn as project faces M&E delays” — https://constructionwave.co.uk/2026/02/23/hinkley-point-c-costs-approach-49bn-as-project-faces-me-delays/

• US Department of Energy: “NRC Certifies First U.S. Small Modular Reactor Design” — https://www.energy.gov/ne/articles/nrc-certifies-first-us-small-modular-reactor-design

• Utility Dive, 2023: “NuScale, UAMPS terminate small modular reactor project in Idaho” — https://www.utilitydive.com/news/nuscale-uamps-terminate-small-modular-nuclear-reactor-smr-project-idaho/699281/

• Boise State Public Radio, November 2023: “NuScale nuclear reactor project in Idaho canceled” — https://www.boisestatepublicradio.org/news/2023-11-10/idaho-small-nuclear-reactor-project-canceled

• E&E News: “NuScale cancels first-of-a-kind nuclear project as costs surge” — https://www.eenews.net/articles/nuscale-cancels-first-of-a-kind-nuclear-project-as-costs-surge/

• US Department of Energy: “NRC Approves NuScale Power’s Uprated Small Modular Reactor Design” — https://www.energy.gov/ne/articles/nrc-approves-nuscale-powers-uprated-small-modular-reactor-design

• World Nuclear News: “China’s demonstration HTR-PM enters commercial operation” — https://www.world-nuclear-news.org/Articles/Chinese-HTR-PM-Demo-begins-commercial-operation

• NextBigFuture, December 2023: “China’s Pebble Bed Reactor Finally Starts Commercial Operation” — https://www.nextbigfuture.com/2023/12/chinas-pebble-bed-reactor-finally-starts-commercial-operation.html

• Wikipedia: Akademik Lomonosov —https://en.wikipedia.org/wiki/Akademik_Lomonosov

• Bellona.org, 2015: “New documents show cost of Russian floating nuclear power plant skyrockets” — https://bellona.org/news/nuclear-issues/2015-05-new-documents-show-cost-russian-nuclear-power-plant-skyrockets

• Power Magazine: “Russia Sees Floating Nuclear Power Plant Costs Balloon” — https://www.powermag.com/russia-sees-floating-power-plant-costs-balloon/

• Wikipedia: CAREM — https://en.wikipedia.org/wiki/CAREM

• Nuclear Engineering International: “Argentina’s CAREM-25 SMR faces setbacks” — https://www.neimagazine.com/news/argentinas-carem-25-smr-faces-setbacks/

• Buenos Aires Herald: “Construction of first Argentine-made nuclear power reactor halted amid layoffs” — https://buenosairesherald.com/business/construction-of-first-argentine-made-nuclear-reactor-halted-amid-layoffs

• ScienceDirect, Progress in Nuclear Energy, 2025: “Challenges of small modular reactors: A comprehensive exploration of economic and waste uncertainties associated with U.S. small modular reactor designs” — https://www.sciencedirect.com/science/article/abs/pii/S0149197025003877

• CleanTechnica summary, September 2025: “Small Modular Reactors and the Big Questions of Cost & Waste” — https://cleantechnica.com/2025/09/10/small-modular-reactors-and-the-big-questions-of-cost-waste/

• UK Government / BEIS, 2024: “Onshore Wind and Solar PV: Cost of Electricity Report Update 2024” — https://assets.publishing.service.gov.uk/media/68ba91f411b4ded2da19fe92/onshore-wind-and-solar-pv-cost-electricity-report-update-2024.pdf

• IEA: “The Path to a New Era for Nuclear Energy” — https://www.iea.org/reports/the-path-to-a-new-era-for-nuclear-energy

• IEA data chart: “Small modular reactor global installed capacity by scenario and case, 2025–2050” — https://www.iea.org/data-and-statistics/charts/small-modular-reactor-global-installed-capacity-by-scenario-and-case-2025-2050

• ESG Today: “Rolls-Royce Signs Deal with UK to for First Fleet of Small Modular Nuclear Reactors” (references Amazon/X-energy investment) — https://www.esgtoday.com/rolls-royce-signs-deal-with-uk-to-deliver-fleet-of-small-modular-nuclear-reactors/

• Google/Kairos Power: “Google and Kairos Power partner to develop nuclear energy” — search Google’s official press release at blog.google for the specific announcement.

• GOV.UK: “Golden age of nuclear delivers UK-US deal on energy security” — https://www.gov.uk/government/news/golden-age-of-nuclear-delivers-uk-us-deal-on-energy-security

• Al Jazeera, 18 September 2025: “US and UK sign major nuclear power deal: What does it include?” — https://www.aljazeera.com/news/2025/9/18/us-and-uk-sign-major-nuclear-power-deal-what-does-it-include

• Nuclear Innovation Alliance: “Regulatory Implementation Summary: NRC Progress Under the ADVANCE Act”(December 2025) — https://nuclearinnovationalliance.org/regulatory-implementation-summary-nrc-progress-under-advance-act

• NRC ADVANCE Act dashboard: https://www.nrc.gov/about-nrc/governing-laws/advance-act/dashboard

• NRC History: “Too Cheap to Meter: A History of the Phrase” — https://www.nrc.gov/reading-rm/basic-ref/students/history-101/too-cheap-to-meter

• Wikipedia: Too cheap to meter — https://en.wikipedia.org/wiki/Too_cheap_to_meter

The people report comes toward the end of the board pack. Somewhere after the financial performance slides and before the risk register. It contains an engagement score — this quarter seventy-one, up three points from the previous period — and a brief commentary from the people director explaining the movement. The board notes the improvement. Someone asks whether the score is above or below sector average. The answer is above. The item closes.

Six months later, a significant portion of the senior leadership team has left. The reasons, when they eventually surface, are not complicated: a sustained period of organisational change had been managed without involving the people most affected in the decisions being made, and those people had concluded, rationally, that the direction was not one they wanted to follow. None of this was invisible. It was legible, in specific terms, in the free-text responses that sat beneath the engagement score the board had accepted as reassurance.

The board had asked the right governance question — how engaged are our people — and received an accurate answer. The number was real. What it failed to convey was whether the organisation understood the conditions producing it, and whether those conditions were stable or deteriorating beneath a score that was still, technically, above average.

This is not an unusual story. It is a governance pattern.

What a people metric is designed to do

Board-level people reporting has evolved to serve a specific function: to give non-executive directors sufficient assurance that the human risk in the organisation is being managed. The engagement score performs that function efficiently. It is comparable across periods, benchmarkable against sector peers, and defensible in the event of a regulatory or reputational challenge. It answers the governance question the board is equipped to ask.

The problem is that the governance question and the strategic question are not the same question. Is engagement above or below threshold is a governance question. Do we understand what our workforce, our customers, or our communities are actually experiencing well enough to anticipate what is coming is a strategic one. The first produces a number. The second requires intelligence. And most board people reporting has been designed to produce the former while giving the impression it addresses the latter.

The distinction matters because boards make consequential decisions about people — through strategy, through resource allocation, through the acquisitions and restructures they approve — and the quality of those decisions depends on the quality of the understanding they rest on. A board that believes a score of seventy-one means everything is broadly in order is making different decisions than a board that understands why the score is seventy-one, which parts of the organisation are driving it, and what would need to change for it to move in either direction.

The score alone cannot support the second set of decisions. It was not designed to.

What boards are not seeing

The gap between what boards receive and what would constitute genuine people intelligence is wide in most organisations, and wider still in specific situations: post-acquisition integration, significant organisational change, sustained periods of uncertainty, or industries where the workforce is distributed, contracted, or otherwise at a distance from the centre.

In each of these situations, the aggregate score becomes less representative precisely when it most needs to be reliable. The integration that is going well for the acquiring team and badly for the acquired workforce produces an average. The restructure that has settled the leadership layer while creating deep anxiety in the operational layer produces an average. The number consolidates the variation into a single point and presents it to the board as a picture of the whole.

What the board is not seeing, in most cases, is the distribution beneath the average — where the score is concentrated, where it is fragile, which parts of the organisation are carrying the weight of others, and what the people in the lower quartile are specifically saying about their experience. Nor is it seeing the qualitative register: the specific concerns, the named conditions, the articulated reasons that would allow a board to distinguish between an engagement dip that is transient and one that is structural.

Boards that have tried to close this gap by asking for more granular data — scores by function, by tenure, by level — get closer to the distribution problem but do not solve the intelligence problem. Granular numbers are still numbers. They show you where the variation is. They do not show you why.

The question no one asks in the boardroom

There is a version of this problem that is specific to the board’s role, and it sits just beneath the surface of every board pack.

Boards are responsible for oversight, not management. The distinction is important. A board that receives an engagement score and accepts it as adequate people intelligence is not failing to manage — that is not the board’s job. It is failing to ask the questions that oversight requires: what does this number actually tell us, what does it not tell us, and what would we need to know to be confident we are not surprised by something we should have seen coming.

In most organisations, no one in the boardroom asks those questions routinely. The people director presents the data they have been asked to present, in the format that has been agreed, against the benchmarks that have been established. The non-executives receive it in a pack that also contains financial performance, risk, and compliance items that require more active scrutiny. People data gets the time allocated to it, which is rarely enough to interrogate it deeply.

The result is that people risk is governed at a level of abstraction that would be considered inadequate in other domains. A board that accepted a single aggregate technology risk score without understanding what it represented, where the exposure lay, and what conditions were driving it would be considered negligent. The equivalent in people governance is standard practice.

The shift required is not primarily about tools or data. It is about what boards decide to demand. Engagement intelligence that is specific, distributed, and grounded in what people actually say — rather than how they respond to a scale — is achievable. Gobby, a UK platform built on qualitative-first surveys, shows what this looks like in practice: respondents answer in their own words, then validate each other’s responses rather than having them averaged by the system. What reaches the board isn’t a number — it’s what the workforce actually said. It requires a different kind of question at board level, and a people function equipped and willing to answer it.

The score will always be easier to present than the understanding. The question is whether the board is prepared to settle for it.

My opinion

I left an organisation once because the leadership had changed five times in five years and no one seemed to know what direction they were heading in. When I left, no one asked why. The exit process was procedural. I was not a data point anyone was going to interrogate. In another business I left because of a executive level change which resulted in an organisational restructure, after the very first meeting with the new executive I knew what was coming and it wasn’t going to be pretty. The thing which I loved about that business was the one thing that was going to be destroyed - Culture. So I resigned, said my piece and everything I said came true.

The people who collect this kind of information are usually not hiding anything deliberately. They are responding to incentives. They have learned, over time, what the board finds useful and what it finds uncomfortable. They present the former. They convert the latter into a score.

That is not intelligence. It is data dressed as reassurance. And if a board cannot hear the difficult alongside the comfortable — if the culture in the boardroom makes hard truths unwelcome — then the problem is not the measurement. It is a leadership failure at the top of the organisation.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Sam has been reading the same paragraph for the third time. The words are there. The comprehension isn’t. Something, unclear what, has been pulling at the edge of attention for the last twenty minutes without ever resolving.

The message arrives that Sam isn’t going to read yet. The alert that Sam already half-knows exists. Some unfinished conversation sitting just below the surface of thought, unanswered, unresolved, quietly taking up space in the mind that was supposedly committed to this page.

Sam isn’t distracted. Sam is present. Physically here, eyes on the text, doing the right thing. The presence is real. The attention is somewhere else entirely.

The attention you were certain you had

In 1999, two psychologists at Harvard asked participants to watch a short video and count how many times a basketball was passed between players wearing white. The task was specific and demanding.

Halfway through the video, a person in a full gorilla suit walked into the frame. Faced the camera. Beat their chest. Walked out. Fifty per cent of participants never saw it. Not because it was subtle. Because their attention was committed elsewhere, and the gorilla wasn’t part of what they’d been asked to track.

We believe we’re present because we’re physically in a situation. The eyes are open. The body is there. But presence and attention are not the same thing, and the gap between them is where a great deal of modern life is quietly disappearing.

Inattentional blindness — the brain’s tendency to miss even obvious stimuli when focus is elsewhere — isn’t a rare laboratory phenomenon. It’s the operating condition of most people’s daily lives.

The brain doesn’t see everything in front of it. It samples, filters, and constructs a picture based on what it has been primed to look for. When that priming mechanism is overwhelmed — when there are too many things competing for focus — the filter becomes less a tool of selection and more a tool of exclusion.

What has changed since the gorilla experiment isn’t the cognitive architecture. It’s the scale and sophistication of the systems competing to own the priming. Every notification, every feed, every platform is fighting to be the thing your brain gets told to track. When they win, everything else becomes the gorilla. Invisible. Present. Missed.

We don’t lose the ability to pay attention. We lose the choice about where it goes.

The things that disappear without announcement

The gorilla in modern life isn’t a person in a suit. It’s the conversation not had. The warning not noticed. The colleague who needed support and didn’t get it because you were technically in the room while mentally somewhere else entirely.

It’s the child who asked a question and got a distracted non-answer, then stopped asking. The friend who was struggling and mentioned it once, sideways, and you were there but not there when it came up. The decision made quickly because the depth of focus required to make it well wasn’t available that day.

The cost isn’t always visible. It rarely arrives labelled. It accumulates in the quality of what you produce, the depth of the relationships you maintain, and the gradual narrowing of what you actually notice about the people and moments that are directly in front of you.

We have been taught to understand this as a personal failing. A concentration problem. A discipline issue. The technology industry is clear, in its habits if not its language, that this framing suits it perfectly. Because if the problem is you, you don’t look at the design.

What the design is actually doing

A system that competes for the priming of your attention, that inserts itself into the filtering mechanism the brain relies on to distinguish signal from noise, is not supporting you. It is working against the very cognitive function it claims to support.

The argument that social platforms augment connection holds only if the connections they produce are of comparable quality to those they replace or displace. The gorilla experiment suggests a direct relationship: the more demanding the competing task, the more complete the blindness. Connection that comes at the cost of presence isn’t connection. It’s substitution.

The deepest failure is this: any system that makes you selectively blind to your immediate environment cannot claim to respond to your actual circumstances. The person sitting with their child. The person driving. The person in the conversation that matters. A system that doesn’t know — or doesn’t care — which of those moments it’s interrupting has no basis for calling itself designed for you.

This is one of the questions at the centre of my book The Cognitive Crucible — not just what technology does, but what it costs us without announcing the price.

My opinion

Attention is the most finite resource we have. I don’t think the platforms taking it are confused about what they’re doing — the design is not accidental, and the framing of distraction as a personal discipline problem is not a misunderstanding either. What concerns me most is the picture of society these platforms have constructed: not a reflection of how people actually live, but of what drives engagement. The two are not the same thing, and recovering the ability to tell them apart — to notice what is actually in front of you rather than what you’ve been primed to track — is, in the most direct terms I can offer, how you get your cognitive agency back.

What this asks of you

The point is not to understand inattentional blindness. The point is to know where it is operating in your own life.

1. What have you noticed yourself missing lately — at work, at home, in conversation — that you’d have caught if you weren’t distracted?

2. Is there someone in your life who has been trying to tell you something you haven’t fully heard? What would it take to change that?

3. What would you need to change about your environment to make genuine presence your default rather than your exception?

4. If someone was watching how you spend your attention across a typical week, what would they conclude about what matters most to you?

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

A pathologist examining a breast cancer biopsy can read a great deal from the slide. The cells — stained pink and purple — arrange themselves in patterns that decades of training make legible. Some areas are dense with malignant cells: the tumour’s centre. Others show the advancing edge, where cancer is pushing into healthy tissue. Some regions are infiltrated by immune cells; others appear almost entirely free of them.

What the traditional biopsy slide cannot show is what those cells are actually doing — which genes each one is expressing, what proteins it is making, what signals it is sending to its neighbours, why the immune cells in one region are actively attacking the tumour while the immune cells in another region appear to have gone quiet. The stain shows structure. It cannot show function.

For most of the history of molecular biology, researchers addressed this limitation by doing what they could: taking the tissue, grinding it up, and measuring the average gene expression of the resulting mixture. This approach called bulk sequencing produces useful information. It also, inevitably, hides the detail that matters. A tumour cell that is resistant to a drug and a tumour cell that is sensitive to the same drug look identical in a bulk sample. Their individual signals cancel each other out. What the measurement captures is the average, and in a heterogeneous tissue, the average misleads.

Single-cell and spatial biology exist to fix that problem. Single-cell technologies analyse gene expression at the resolution of individual cells. Spatial technologies add location: they preserve the tissue architecture and record not just what each cell is doing but where in the tissue it sits. Together, they produce something that did not exist before — a map of a living tissue that shows, at microscopic resolution, which genes are active in which cells in which locations, and what those cells are saying to each other.

Nature Methods named spatially resolved transcriptomics its Method of the Year in 2020. Since then, the technology has moved from the leading edge of basic research to commercial platforms available in well-equipped laboratories worldwide. In February 2025, Illumina announced a spatial technology product planned for commercial release in 2026 with a capture area nine times larger than existing technologies and four times greater resolution. The field has tipped from pioneering to mainstream. What that means for understanding disease — and ultimately for treating it — is the subject of this article.

The tumour is not the problem. The neighbourhood is.

The central insight that spatial biology has delivered to cancer research can be stated simply: a tumour is not a mass of identical malignant cells. It is an organised community of different cell types — cancer cells in various states, immune cells of multiple kinds, structural cells called fibroblasts, blood vessel cells, and others — all occupying distinct spatial niches within the tissue, interacting constantly with their immediate neighbours, and behaving very differently depending on where they are.

This was suspected for a long time. What spatial biology has done is make it visible and measurable.

One of the most striking findings from spatial transcriptomics research is that the same tumour can simultaneously contain cells that are highly sensitive to a treatment and cells that are highly resistant — and that this is not purely a matter of genetics. Location matters. A cancer cell sitting in the centre of a tumour, surrounded by other cancer cells and relatively few immune cells, behaves differently from a genetically identical cancer cell at the tumour’s leading edge, where it is in constant contact with immune signals and structural proteins from the surrounding healthy tissue. The edge cell may be more aggressive and invasive; the centre cell may be more metabolically active but less mobile. Their gene expression profiles differ. Their drug sensitivities differ. The treatment that kills one may leave the other intact.

A tumour contains cells that are sensitive to a drug and cells resistant to the same drug — not just because of genetics, but because of location. Cells at the leading edge behave differently from cells at the centre. Spatial biology makes this visible.

A 2023 paper in Nature Communications on oral squamous cell carcinoma demonstrated that the gene expression profile of cells at the tumour’s leading edge — what researchers call the leading edge signature — was consistently associated with worse clinical outcomes across multiple cancer types, while the tumour core signature was associated with improved prognosis. The spatial position of the cancer cells relative to the boundary between tumour and healthy tissue carries predictive information that no amount of bulk sequencing could have extracted, because bulk sequencing cannot preserve the information about where each cell was sitting.

Immune cells show equally striking spatial organisation. Immunotherapy — the class of treatments that effectively unlocks the immune system to fight cancer — works by removing the signals that cancer cells use to suppress immune activity. It works well for some patients and not at all for others. Spatial biology research is revealing why. In patients who respond to immunotherapy, spatial analysis shows specific arrangements of immune cells clustered in particular locations around the tumour — what researchers call immune hubs or tertiary lymphoid structures. In patients who do not respond, those structures are absent or differently arranged. The presence or absence of the structures is not detectable by any analysis that does not preserve the spatial information. A blood sample, a bulk biopsy, even conventional single-cell sequencing without location data cannot tell you whether the right immune cells are in the right places.

Spatial transcriptomics of small cell lung cancer tumours has identified two distinct subtypes of cancer cells within the same tumours: one associated with treatment resistance and high proliferative activity, the other associated with treatment sensitivity and greater immune cell contact. The resistant subtype appears to reshape its local environment to drive away immune cells — to domesticate the macrophages in its vicinity and convert them into accomplices in suppressing the immune response. This finding was invisible to conventional analysis. It required spatial mapping of the tumour to become apparent.

The platforms are ready. The clinical pipeline is not.

The commercial field for spatial biology has developed quickly and is now genuinely competitive. 10x Genomics offers the Xenium In Situ platform, which can detect up to 5,000 genes across tissue samples at subcellular resolution. Bruker’s CosMx platform can profile over 19,000 genes — the entire protein-encoding transcriptome. Visium HD, also from 10x Genomics, resolves spatial data to 2-micrometre bins. Illumina’s forthcoming platform will use a capture area nine times larger than current technologies, integrated with its Connected Multiomics software suite for analysis.

The availability of these platforms has shifted the technology from something that required specialist physics laboratories to something that any well-resourced research institution can operate. The research output has followed. A review of spatial omics in cancer research published in Cancer Cell in January 2026, by researchers at the MD Anderson Cancer Center, described how spatial technologies now map tumour architecture with a fidelity that earlier methods could not achieve, resolving functional niches and spatial communities, converting spatial patterns into mechanistic insights.

The transition from research tool to clinical instrument, however, has several steps remaining. The data volumes generated by spatial transcriptomics are enormous. A single spatial experiment can produce hundreds of gigabytes of data. Analysing it requires computational infrastructure and bioinformatics expertise that most clinical settings do not yet have. The integration of spatial transcriptomics data with conventional pathology, with genomics, with clinical records, and with the kind of standardised workflows that regulatory agencies require before a diagnostic test can be approved — all of this is in progress but not yet complete.

The cost has been falling but remains significant. Running a spatial biology experiment on a clinical sample costs more than conventional sequencing and takes longer. For research settings, this is manageable. For routine clinical use — for the pathologist examining a cancer biopsy to guide treatment decisions for a patient who needs an answer within days — the workflow, cost and analytical requirements are still a barrier. The question of where spatial biology first enters clinical practice is the most practically important one in the field right now.

The most plausible near-term pathway is through what researchers call companion diagnostics: tests that are performed alongside a specific treatment to determine who is most likely to benefit. The patterns that spatial biology can detect — the presence or absence of immune structures around a tumour, the proportion of cells with certain gene expression signatures at the tumour edge, the spatial arrangement of drug-resistant subpopulations — are exactly the kind of information that could guide decisions about whether to use immunotherapy, which chemotherapy to start with, or whether a tumour is likely to respond to a newly approved targeted agent.

Why drug resistance looks different from here

Drug resistance is the central unsolved problem of cancer treatment. Most patients who are initially helped by a drug eventually stop responding to it. Understanding why, and predicting in advance who will stop responding and when, is the problem that determines whether a patient can be switched to an effective treatment before their cancer becomes untreatable.

Spatial biology is offering new ways to think about this problem. The conventional model of drug resistance focuses on genetics: a cell acquires a mutation that makes it insensitive to the drug, and that cell proliferates while the drug kills everything else. That model is real and important. But spatial biology research is demonstrating that it is incomplete. Resistance is also ecological.

A resistant cell does not only survive because of its own genetic properties. It also survives because of where it is in the tissue — who its neighbours are, what signals they are sending, whether the local immune environment is permissive or hostile. Research in melanoma has shown that distinct resistant fates — different ways a cell can become drug-resistant — are not randomly distributed through a tumour. They are spatially organised, associated with specific microenvironments. The cells in regions of necrosis behave differently from the cells adjacent to functional blood vessels. The resistant fate a cell adopts appears to be partly determined by its location.

This has immediate implications for treatment design. If resistance is partly spatial, then treating a tumour as a single uniform target with a drug calibrated to the average cell will predictably fail. The cells most resistant to treatment are often those in locations that the drug reaches least effectively and where the microenvironment most actively suppresses the immune system. A treatment strategy informed by spatial biology might target not just the cancer cells but their supporting environment: the macrophages they have domesticated, the fibroblasts that form barriers, the immune exclusion zones that protect the most resistant subpopulations.

If resistance is partly spatial, treating a tumour as a uniform target will predictably fail. The cells most resistant to treatment are often in the locations least accessible to drugs and most protected by their microenvironment.

Spatial analysis of breast tumours combined with pharmacogenomic profiling has identified that cancer cells in immunosuppressive microenvironments — the regions of the tumour where the immune system has been most effectively shut down — show specific sensitivity to cell cycle arrest agents and a class of drugs targeting the PI3K/AKT/mTOR pathway. This is information that could not be derived from bulk sequencing, because the immunosuppressive microenvironment and the cancer cells within it cannot be distinguished in a ground-up tissue sample. Seeing the map reveals what the average conceals.

Beyond cancer, spatial biology is beginning to reveal the tissue architecture of conditions that have been poorly understood precisely because they are diseases of organisation rather than diseases of individual cells. Pulmonary fibrosis — scarring of the lung tissue — involves molecular dysregulation at specific locations where epithelial cells are actively remodelling. Identifying exactly where in the tissue this remodelling is happening, and which cells are driving it, is exactly what high-resolution spatial transcriptomics can provide. Early data from the Illumina platform has shown that the scale and sensitivity of the technology enables the whole transcriptome to be studied across large tissue sections, identifying localised molecular dysregulation that earlier technologies could not resolve.

Who the insight reaches

Spatial biology is doing something specific for researchers and eventually clinicians: it makes visible biology that was previously invisible. Not incompletely understood — invisible. The question of whether two cancer cells in different parts of a tumour are behaving differently could not be asked until there was a technology that could preserve spatial information and measure gene expression simultaneously. Now it can be asked and answered. That matters for everyone whose work depends on understanding why treatments fail.

The distance between the research laboratory and the patient is where the real question sits. At present, the insights from spatial biology are largely flowing through academic research and pharmaceutical development. The direct beneficiary is the next generation of clinical trials — better-designed studies, better patient stratification, drugs tested against tumours whose architecture has been mapped rather than averaged. That is real progress. But the pathway from spatial biology discovery to a changed treatment decision for a specific patient is still long, and the technology to shorten that pathway — affordable, fast, clinically integrated spatial diagnostics — is not yet routinely available.

Whether spatial biology becomes a tool for everyone or a research instrument for well-resourced institutions is the question that will determine its real value. The core insight — that a patient’s tumour has a spatial architecture that determines how it will respond to treatment — is as true for a patient at a regional cancer centre as for one at a major academic medical centre. The Illumina platform under development is explicitly designed to reduce cost and integrate with existing sequencing infrastructure, which suggests the field is at least trying to address the accessibility barrier. The data infrastructure required to analyse the results — the computational pipelines, the bioinformatics expertise, the clinical interpretation frameworks — presents a more persistent challenge.

What the map cannot yet tell us

1. Spatial biology reveals that the same tumour can simultaneously contain cells that respond to a treatment and cells that resist it — and that location, not just genetics, partly determines which is which. How does that change how you think about the phrase ‘this treatment works for this cancer’?

2. If the spatial architecture of a tumour predicts whether immunotherapy will work, and that spatial information can only be obtained by analysing the biopsy tissue before treatment starts — who should have access to that analysis? What would it mean for cancer equity if spatial diagnostics are only available at specialist centres?

3. Spatial biology generates enormous datasets from individual tissue samples. Making sense of that data requires AI tools, bioinformaticians, and infrastructure that most clinical settings do not currently have. What is the bottleneck that most needs addressing before this technology reaches routine clinical use?

4. The insight that drug resistance is partly ecological — shaped by the local environment around cancer cells rather than only by the cells’ own genetics — suggests that some treatments may need to target the microenvironment, not just the tumour. What does that mean for how cancer drugs are designed and tested?

My opinion

The gap between what spatial diagnostics can reveal and where that capability is available is not a technical problem. It is the same problem that follows every powerful diagnostic technology: it arrives at the best-resourced centres first and, without deliberate intervention, tends to stay there. Cancer does not follow a map of specialist centres. It arrives wherever it arrives — in rural hospitals, in regional oncology units, with patients who cannot travel to the institutions that define the frontier. The field needs to be working towards diagnostics that are portable, affordable, and deployable anywhere, not because that would be admirable, but because the technology’s actual value depends on it.

The pathologist looking at the traditional biopsy slide can see structure. Spatial biology lets you see function. It lets you read which cells are talking to which other cells, and what they are saying, and what happens in the neighbourhood where the immune system has given up.

The dimension is location. And it turns out to matter enormously.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

There is a particular kind of silence that follows a consultation. Not the silence of consideration, where something heard is being weighed. The silence of completion, where a process has been concluded and the findings have been filed. The residents submitted their responses. The staff completed the survey. The service users attended the focus group. And then, at some point later — sometimes months, sometimes longer — a decision was announced that bore little visible relationship to what anyone said.

Most people who have been consulted by a large organisation have experienced this. Many have stopped expecting otherwise. That learned scepticism is not apathy. It is a rational response to a pattern that repeats across sectors with enough consistency to suggest it is not accidental.

The pattern has acquired a name — consultation fatigue — but the name understates what is actually happening. Fatigue implies a willingness that has been worn down. What most organisations have produced is something more deliberate: a process designed to look like listening without requiring the organisation to change course as a result of what it hears.

The mechanics of being heard without being heard

The structure of most consultations gives the game away.

Questions are written by the organisation, which means the answer categories reflect the organisation’s existing understanding of the problem. Respondents are asked to rate, rank, or select from options that were designed before anyone outside the room was consulted. Free-text fields exist, but are rarely analysed systematically — they are too difficult to process at scale and too unpredictable in what they might surface. The consultation closes, the quantitative results are tabulated, and the report that emerges reflects the organisation’s original position with a margin of adjustment around the edges.

This is not cynicism dressed as analysis. It is the logical consequence of designing a consultation process around the outputs you need to produce rather than the understanding you need to reach. In local government, that output is a mandate — a defensible record that affected parties were involved before a decision was made. In the NHS, it is a compliance requirement — evidence of patient and public involvement in service design. In the private sector, it is an engagement score — something to report to the board or include in a regulatory submission. Each has a legitimate purpose. None of them requires the organisation to be genuinely changed by what it hears.

The people being consulted know this. The communities most affected by decisions are typically the ones with the least trust in the processes designed to involve them — because they have the most experience of those processes. When a community has watched a planning consultation produce a predetermined outcome, or seen a workforce survey followed by a management communications campaign rather than structural change, the rational response is to disengage. Not because the issues don’t matter to them. Because they have learned that engaging produces no different outcome than not engaging.

What the same pattern looks like across sectors

In local government, the planning consultation is the clearest case. Residents are invited to comment on a proposal that is already substantially formed. The statutory requirement is to consult; there is no equivalent requirement to be materially influenced by what is said. Communities that oppose a development learn to flood the process with responses because volume is the only variable that demonstrably affects outcomes — not the quality of the argument, not the specificity of the concern, not the evidence that the proposal conflicts with the community’s stated needs.

In healthcare, service redesign consultations run on similar logic. A trust deciding to reconfigure services will conduct patient engagement as part of the process. The engagement is genuine in the sense that it happens; it is limited in the sense that the financial and operational case for the reconfiguration is typically established before the engagement begins. What the consultation can change is the detail. What it rarely changes is the direction.

In the private sector, the employee consultation — particularly in a restructure or acquisition — operates under legal requirements in the UK that mandate a period of consultation before certain decisions are implemented. The requirement is to consult; it is not to reach agreement, and it is not to change course if the arguments against the proposal are sound. HR teams are well practised in running consultations that satisfy the legal threshold while protecting the operational decision already made.

In the charity sector, the gap is perhaps most uncomfortable to acknowledge. Organisations built on the principle of community voice frequently struggle to implement it in practice. Governance structures centre the board and the executive. Impact measurement centres the funder’s reporting requirements. The beneficiary’s account of their own experience sits at the end of a chain of representation — reported by a caseworker, summarised by a manager, condensed into an impact metric by a communications team — that has progressively removed the original voice by the time it reaches anyone with decision-making authority.

The thing that peer validation changes

The problem is not only that organisations fail to act on what they hear. It is that the processes they use mean they often do not hear clearly in the first place.

A consultation that asks closed questions gets closed answers. A free-text field that sits at the end of a fifteen-minute survey gets responses from the minority who feel strongly enough to write something additional. An engagement that channels responses through a manager or a facilitator introduces the interpretation of the intermediary before the data reaches analysis. By the time findings are presented to the people who will make the decision, the original voice has been translated several times.

The methodological shift that matters here is not just collecting open responses — it is validating them within the community before they reach the analyst. When people can see each other’s responses and indicate what resonates, two things happen that closed surveys cannot produce. First, the synthesis reflects the community’s own sense of what matters, rather than the researcher’s coding scheme. Second, people who contribute to a process and see their contribution acknowledged by peers are less likely to conclude that the exercise was performative. The experience of being heard is partly structural — it depends on whether the process was designed to hear or designed to record.

Gobby is one tool built on this logic — open responses rather than closed questions, peer-powered voting so participants can indicate what resonates, AI-assisted synthesis that reflects the community’s own sense of priority rather than the analyst’s coding scheme. It is still used by a minority of organisations. The majority continue to run processes that produce defensible records rather than genuine intelligence. The technology to do otherwise is not the constraint. The will to act on what might be heard is.

That is the harder problem, and no tool solves it. What better processes do is remove the organisation’s ability to claim it didn’t know. If the findings are specific, community-validated, and clearly stated, the decision to set them aside becomes a conscious choice rather than an administrative oversight. That is a different kind of accountability — and it is the reason some organisations prefer the consultation that changes nothing.

My opinion

I have been in enough of these rooms to know the difference between an organisation that doesn’t know how to listen and one that has decided not to. Most consultation theatre is the second kind. Tools like Gobby do something useful: they remove the claim of ignorance. If the findings are specific and community-validated, the decision to set them aside is no longer an oversight. It is a choice. And the organisations that consistently make it know exactly what they are doing.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Morgan comes on shift at 11pm and the menu is gone. Not crashed — gone. The update that rolled out at 6pm moved it somewhere else, and the section of iOPS that handles electronic file transfer to the Crown Prosecution Service is no longer where it was yesterday. There is a person in the custody suite waiting to be processed. Morgan has a legal time limit. Morgan picks up the paper form.

This is Greater Manchester Police in July 2019. iOPS — the Integrated Operational Policing System — went live sixteen months behind schedule. The custody module, which was supposed to manage custody records and transfer files electronically to the CPS, was not working as specified. The announcement had said it was ready. The go-live had happened. Nobody said out loud that it wasn’t done.

It wasn’t done.

Deployed, not delivered

The workaround that custody sergeants had been using for three months no longer worked. The update had changed the interface without adequate retraining. There was no mechanism to refuse the tool — the option was to make it work, improvise a new workaround, or escalate a complaint that had no clear landing point. Officers improvised.

GMP’s own independent review named the consequences: technical shortcomings, inadequate training, and operational disruption that left vulnerable people without referrals to support services. Referrals for high-risk domestic abuse victims fell by 50%. Referrals to Greater Manchester Victims’ Services fell by 87%. Officers spent time running duplicate paper processes alongside the digital system that was supposed to replace them. The system did not help the workforce do the work — it created a second job beside it.

iOPS had gone live not because it was ready but because the organisation needed it to have gone live. Sixteen months of delay had a contractual cost, a political cost, and a considerable cost in credibility. Ready was redefined as deployed, and the workforce inherited the gap between the two.

Why no one says it isn’t ready

A software product that is finished cannot justify an ongoing roadmap. It cannot retain an engineering team. In a subscription market, completion is a liability — a done product is one that doesn’t need what comes in the next release cycle. Agile methodology and continuous deployment gave the industry a vocabulary for this: not unfinished, but iterating. Not incomplete, but evolving. The language made incompleteness sound like a sign of rigour.

Organisations accepted this because the alternative was more expensive. To say the system was not ready meant saying the procurement decision was wrong. GMP’s contract had already been signed, extended, and renegotiated. Admitting the custody module was unfit meant unpicking an obligation that was difficult to unpick. Deploying and adapting was easier.

There was no real mechanism to refuse. The tool was the tool. The workarounds became part of the job. The cost was absorbed shift by shift by custody sergeants who had no power to put it back where it came from.

The cost that didn’t appear in the contract

Birmingham City Council signed a contract for an Oracle ERP system projected at £19 million. By 2026 the total cost is forecast to reach £216.5 million. Two and a half years after go-live, the system was still not, in the auditors’ own words, safe and compliant. In September 2023 Birmingham declared bankruptcy — the largest local authority insolvency in UK history. The system was never ready either.

NHS IT outages disrupted more than 274,000 patient interactions across six major trusts in a single year. In each of those incidents, clinical staff reverted to paper. Not because paper was preferred, but because the digital system had failed and the paper process was the only thing keeping the work going. More than a quarter of UK central government digital systems are now classified as outdated, at an estimated cost to the public sector of £45 billion annually in foregone productivity.

Office workers endure an average of 3.6 technology interruptions a month — software updates, system failures, workarounds required. Nobody counted this cost when the contracts were agreed, because the cost landed on the worker. The vendor’s contract was satisfied. The product had been delivered. That it didn’t work as specified was the next problem, billable separately.

What it was designed for

The most basic test for any system deployed to workers is whether it actually helps them do the specific work it replaced. iOPS failed that test on the night the update moved the workflow officers had learned. The tool stopped being a tool and became an obstacle with extra steps. The work still needed doing. Staff found a way to do it. The system got the credit for a deployment that happened.

The next test is whether what the vendor calls an improvement actually improves anything for the person using it. iOPS received updates. Each notification used the word “improvement.” Custody sergeants had not requested the features those updates added. The gap between what the vendor was building and what the workforce needed is not a failure of communication. It is the expected result of a development roadmap driven by the vendor’s commercial cycle, not by the legal obligations of a sergeant at 11pm.

The final test is whether the system responds to the specific person in front of it, or expects the person to fit. iOPS updated on the vendor’s schedule. Its release windows did not account for shift patterns, operational hours, or the consequences of breaking a workaround mid-cycle. The specific circumstances of night-shift custody work — legal time limits, vulnerable people waiting — were not part of the specification for when an update went live. The system was not designed around the people using it. They were expected to fit.

Sources

• HMICFRS: Greater Manchester Police Integrated Operational Policing System Inspection, 2020. https://hmicfrs.justiceinspectorates.gov.uk/publications/greater-manchester-police-integrated-operational-policing-system-inspection/

• Computer Weekly: “Police IT system failure creates significant backlogs”, 2020. https://www.computerweekly.com/news/252479783/Police-IT-system-failure-creates-significant-backlogs

• Data Centre Dynamics: “Total cost of Birmingham City’s Oracle system failure to reach £216.5m by 2026”, 2024. https://www.datacenterdynamics.com/en/news/total-cost-of-birmingham-citys-oracle-system-failure-to-reach-2165m-by-2026-report/

• The Register: “Birmingham City’s Oracle ERP system still not ‘legally safe’”, April 2024. https://www.theregister.com/2024/04/17/birmingham_city_councils_oracle_erp/

• The Register: “Birmingham City Council goes under after Oracle disaster”, September 2023. https://www.theregister.com/2023/09/05/birmingham_city_council_oracle/

• Security Brief UK: “NHS IT outages disrupt 274,620 patient interactions”. https://securitybrief.co.uk/story/nhs-it-outages-disrupt-274-620-patient-interactions

• Tech Monitor / DSIT: “Legacy technology costs UK public sector £45bn annually”. https://www.techmonitor.ai/government-computing/legacy-technology-costs-uk-public-sector-45bn-annually/

• Ivanti: 2025 Digital Employee Experience (DEX) Report, 2025. https://www.ivanti.com/company/press-releases/2025/tech-disruptions-cost-companies-millions-of-dollars-in-lost-productivity-annually-according-to-research-from-ivanti

  ---

  You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

In the spring of 2023, a patient at Memorial Sloan Kettering Cancer Center in New York received an injection that had been made specifically for them. The sequence of molecules inside it had been designed using a biopsy of their own tumour. No one else in the world would receive exactly this treatment. The vaccine’s job was not to prevent a disease — it was to teach the immune system to recognise and destroy cancer cells that were already there.

The patient had pancreatic cancer. Pancreatic cancer kills around 87% of the people diagnosed with it within five years. It is resistant to most forms of treatment. The immune system usually does not recognise it as a threat. The mRNA vaccine was attempting to change that.

Three years later, the patient had not relapsed. In the trial they had been part of, vaccine-induced immune responses were still detectable in the blood — a signal that the immune system was still, years on, watching for the cancer to return. This is not a cure. It is not even yet a proven treatment outside of trials. But it is the kind of result that changes what researchers believe is possible.

The technology making it possible is the same one that protected hundreds of millions of people against COVID-19. mRNA — messenger ribonucleic acid — carries genetic instructions into cells, tells them to produce a specific protein, and then disappears. For infectious disease, the instruction was: produce a fragment of the virus’s spike protein, and let the immune system learn to recognise it. For cancer, the instruction is more personal: produce antigens that match the specific mutations in this patient’s tumour, so the immune system learns to hunt down cells carrying those mutations. The concept is the same.

Beyond COVID: what mRNA can and cannot do

The COVID vaccines demonstrated something that decades of laboratory research had suggested but never proved at scale: mRNA can be manufactured rapidly, delivered safely into human cells, and used to produce a protein the body has never encountered before. Before 2020, no mRNA medicine had ever been approved for human use. By the end of 2021, hundreds of millions of doses had been administered. The speed was the point — mRNA therapies can be designed in days once the target sequence is known, compared with the months or years required to develop traditional protein-based vaccines.

The platform is now being turned on cancer, autoimmune diseases, and genetic conditions where the body produces too little or too much of a particular protein. The logic is the same in each case: if you can specify the instruction, you can produce the protein.

Cancer is the most advanced frontier. There are now more than 120 clinical trials of mRNA cancer vaccines currently under way, targeting melanoma, lung, colorectal, pancreatic, brain, prostate and breast cancers, among others. The most significant results to date come from melanoma. A personalised mRNA vaccine called mRNA-4157, developed by Moderna in collaboration with the immunotherapy drug pembrolizumab, reduced the risk of recurrence or death by 44% compared to pembrolizumab alone in a Phase 2 trial. Patients given the combination were less likely to see their cancer return at five years than those who received the immunotherapy drug on its own. Phase 3 trials are now enrolling globally.

mRNA carries instructions into cells. For COVID, the instruction was: learn to recognise a virus. For cancer, the instruction is: learn to hunt down cells carrying these specific mutations.

The pancreatic cancer results are more tentative but potentially more significant, given how few treatment options exist. BioNTech’s personalised vaccine, autogene cevumeran, in combination with chemotherapy and immunotherapy, produced durable immune responses in pancreatic cancer patients — and in follow-up data published in 2025 and 2026, seven of the eight patients whose immune systems responded to the vaccine were still alive four to six years after surgery. Of the eight who did not respond, only two were. Pancreatic cancer’s resistance to immunotherapy had long been assumed to be near-absolute. These results suggest it may be conditional.

Beyond cancer, mRNA is in clinical trials for influenza, HIV, tuberculosis, and a range of rare genetic conditions where specific proteins are absent or dysfunctional. The questions are about speed, cost, and who gets access.

The personalisation problem

The most powerful version of mRNA cancer therapy is also its most limiting feature. A personalised vaccine — one designed around the specific mutations in a single patient’s tumour — is, by definition, a product that cannot be manufactured at scale. It requires a biopsy, genomic sequencing to identify which mutations are present and likely to trigger an immune response, computational modelling to select the right antigens, and then manufacture of the mRNA sequence that encodes them. This has to happen quickly enough to be clinically relevant, which in practice means within weeks of surgery.

The current manufacturing timeline for a personalised mRNA vaccine runs between six and ten weeks from biopsy to treatment-ready product, depending on the programme, and researchers are working to compress it further. The logistics are not incidental to the clinical result. In the BioNTech pancreatic trial, tumour tissue removed in New York was shipped to manufacturing facilities in Germany. The time between tissue leaving the body and the vaccine being ready had to be minimised. This is not a process that transfers easily to a district general hospital in the English Midlands, or a regional cancer centre in West Africa.

Cost is the other constraint. Manufacturing a personalised mRNA cancer vaccine currently costs upwards of $100,000 per patient. That figure reflects the early-stage nature of the technology — the expectation is that automation and manufacturing improvements will bring it down — but it is the figure that exists today, not the figure that might exist in 2035. For a health system trying to decide where to invest limited resources, a treatment that costs more than most people in the world earn in a lifetime, and that has not yet completed Phase 3 trials, is not an immediate commissioning priority.

A personalised cancer vaccine is designed around a single patient’s tumour mutations. That is its power. It is also why manufacturing it at scale, at a cost that health systems can absorb, is the defining challenge.

There is a further complication that does not feature in most of the coverage of mRNA therapeutics, but is directly relevant to where the technology goes next. In the United States, federal investment in mRNA vaccine research — which underpins much of the academic pipeline — was significantly cut in 2025. The National Cancer Institute saw funding reductions of around 31% in the first three months of that year. Research programmes specifically developing mRNA cancer vaccines faced additional scrutiny from the Department of Health and Human Services. The political pressure on mRNA technology in the US, driven by anti-vaccine sentiment that predated and outlasted COVID, is a genuine threat to the public research infrastructure on which the field depends.

The UK has moved in the opposite direction. In August 2025, the government announced £30 million for a UK RNA Biofoundry specifically to accelerate RNA therapy development. The Cancer Vaccine Launch Pad, a collaboration between the NHS and BioNTech, is fast-tracking thousands of patients into mRNA cancer vaccine trials targeting lung, breast, head and neck, and other cancers. The global market for mRNA cancer therapies is projected to reach $30 billion by 2033. The investment is arriving. Whether it arrives equitably — or concentrates in the health systems that can most readily afford it — is the question that will define the technology’s impact.

The distance between discovery and access

The patient at Memorial Sloan Kettering who received the pancreatic cancer vaccine in 2023 lived close to one of the world’s best-funded cancer research institutions, in a country with a health insurance system that covers experimental treatments for some patients in some circumstances, and was enrolled in a clinical trial with costs covered by the research programme. These were not incidental conditions. They were why that patient was treated and not someone else.

Pancreatic cancer does not concentrate in wealthy neighbourhoods. Neither does melanoma, lung cancer, glioblastoma, or any of the other cancers now being targeted by mRNA trials. But the trials are running almost exclusively in high-income countries, at major academic medical centres, with manufacturing infrastructure that does not exist in the places where cancer burden is rising fastest. By 2050, global cancer incidence is projected to reach 35 million cases per year — a significant portion of that growth driven by population ageing in low and middle-income countries. The treatments being developed now may arrive in those countries decades after they become available in New York or London, if they arrive at all.

This is not unique to mRNA. It is the pattern of almost every major therapeutic innovation. But mRNA was supposed to be different — its manufacturing speed and relative simplicity were explicitly cited, during the pandemic, as reasons to believe it could close the gap between rich and poor countries in access to medicines. That aspiration now needs to be tested against the reality of personalised cancer treatment, where the manufacturing process is the opposite of simple, the cost per patient is the opposite of affordable, and the regulatory frameworks governing it are only beginning to be developed.

Sources

• Pancreatic cancer survival rate: Pancreatic Cancer Action Network press release, 2025: five-year relative survival rate stalls at 13%. American Cancer Society, Survival Rates for Pancreatic Cancer, 2025. cancer.org

• mRNA-4157 melanoma trial (KEYNOTE-942): Individualised neoantigen therapy mRNA-4157 (V940) plus pembrolizumab versus pembrolizumab monotherapy in resected melanoma: a randomised, phase 2b study. The Lancet, 2023. Updated three-year results presented at ASCO, 2025. Five-year recurrence-free survival data: CancerNetwork, 2025.

• BioNTech autogene cevumeran — pancreatic cancer: Three-Year Phase 1 Follow-Up Data, BioNTech press release, 2025. Updated follow-up (4–6 years): Memorial Sloan Kettering Cancer Center news release; reported in CNN Health, April 2026. AACR Annual Meeting 2025: immune response correlates with clinical benefit in pancreatic cancer patients.

• US federal cancer research funding cuts: US Senate Minority Staff report, May 2025: federal cancer research funding reduced by approximately 31% in the first three months of 2025. Reported in Axios, PBS NewsHour, OncLive (May 2025).

• UK RNA Biofoundry: GOV.UK press release, 28 August 2025: £29.6 million investment in UK RNA Biofoundry at CPI’s RNA Centre of Excellence, Darlington. gov.uk

• Global cancer incidence projections to 2050: IARC / GLOBOCAN 2024 projections: global cancer incidence projected to reach 35 million cases annually by 2050, a 77% increase from 2022. Reported by UN News, February 2024. (Note: the Global Burden of Disease Study 2023 uses a lower estimate of 30.5 million; the IARC figure is the more widely cited projection.)

• mRNA cancer vaccine market projection ($30 billion by 2033): Noted in commentary published in The Lancet Oncology, 2025, drawing on market research projections. The underlying figure originates from commercial market research.

• Manufacturing cost and timeline: Manufacturing costs upwards of $100,000 per patient: reported across multiple clinical development sources including Scientific American (2025) and RNA cancer vaccine pipeline reviews (PMC, 2025). Timeline of 6–10 weeks from biopsy to treatment-ready product: Moderna manufacturing data; autogene cevumeran trial documentation showing approximately 9-week average for the BioNTech pancreatic cancer programme.

• 120+ clinical trials: Current Progress and Future Perspectives of RNA-Based Cancer Vaccines: A 2025 Update. Cancers (MDPI), May 2025. PMC12153701

• Cancer Vaccine Launch Pad: NHS England / BioNTech Cancer Vaccine Launch Pad: fast-tracking patients into mRNA cancer vaccine trials targeting melanoma, lung, breast, head and neck, and other cancers. nhs.uk

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Somewhere in most organisations there is a folder, a shared drive, or an archived system containing several years of things people said. Exit interview notes. Customer complaints in their original form, before they were categorised and counted. Free-text responses from the bottom of annual surveys. Messages sent to a service team that were resolved, closed, and filed. Forum posts from an internal platform that was replaced two years ago and whose contents were migrated to storage rather than deleted.

None of it has been read systematically. Most of it never will be.

This is not a data problem. The data exists. It is a processing problem — or rather, it was. For most of the period in which this material was accumulating, there was no practical way to analyse it at scale without a significant investment in qualitative research that most organisations could not justify for data they had not planned to use. So it was treated as administrative residue: kept because deletion felt risky, ignored because analysis felt impossible.

That calculation no longer holds.

The signal that was never treated as data

The gap between what organisations collect and what they analyse has always been widest in the qualitative register. Numbers get processed. Text gets filed.

This is partly a legacy of the tools available. Survey platforms were built to aggregate ratings, not to synthesise narratives. CRM systems were built to track transactions, not to find patterns in how customers described their experience. HR systems were built to record headcount and performance scores, not to make sense of what people wrote in the comments field at the bottom of the appraisal form. The architecture of most organisational data infrastructure assumes that what matters can be counted.

The consequence is that organisations have spent years collecting a secondary record of their own reality — the version in which people spoke in their own words, described their own experience, and articulated what the numbers could not carry — and treating it as noise.

Exit interviews are the clearest example. Most organisations conduct them. Most collect the responses in some form. Very few analyse them across a period long enough to identify patterns — which managers people were leaving, which structural conditions were driving attrition, which aspects of the culture were named repeatedly by people who had decided to leave but no longer had any reason to be diplomatic. The data exists. The analysis that would make it useful never happens, because exit interview findings are processed as individual cases rather than as a cumulative record.

Customer complaints follow the same logic. The categorised version — number of complaints by type, by channel, by product — reaches a dashboard. The original text, in which customers described in specific terms what had gone wrong and what it had cost them, is archived after the case is closed. Organisations that have been collecting complaints for five years have five years of precise, unsolicited qualitative description of the gap between what they promise and what they deliver. Almost none of them have read it.

What changes when the analysis becomes possible

The shift that is underway is not primarily about new data. It is about the retroactive value of data that already exists, and the prospective design of data collection with analysis in mind from the start.

AI-assisted qualitative analysis has made it practical to process large volumes of unstructured text in ways that were previously either manual and expensive, or automated and crude. The earlier generation of text analytics tools — keyword frequency, sentiment scoring — reduced qualitative data to a thin quantitative proxy. What is becoming possible now is closer to genuine synthesis: identifying themes, finding patterns, distinguishing between types of concern that look similar in aggregate but are structurally different in cause.

The organisations best placed to use this capability are not necessarily the ones with the most sophisticated data infrastructure. They are the ones that recognise they have been sitting on material with analytical value and begin treating it as such. A charity with three years of service user feedback in a case management system. A professional services firm whose client debrief notes have been saved but never reviewed across engagements. A healthcare trust whose patient letters — the ones written by people who felt strongly enough to put something in writing — are archived in a folder that no analyst has opened.

The design question for new data collection is equally important. When an organisation understands that qualitative responses have analytical value, it changes how it structures the collection. Open-response questions get asked more deliberately. Free-text fields get positioned earlier in a survey, not appended as an afterthought. Platforms are chosen with synthesis in mind rather than just with response rates in mind. Tools such as Gobby.io, which combines open response with peer-powered voting — where participants indicate which responses reflect their own experience — produce data that is both richer and more efficiently analysable than free text alone, because the participants have done part of the synthesis work before the data reaches an analyst.

What the evidence was saying all along

The most significant implication of this shift is not operational. It is about what organisations know about themselves, and what they have chosen not to know.

An organisation that has been collecting qualitative signal for years and never analysed it has, in practice, been making decisions without a significant part of the available evidence. That is not a neutral position. It means that the version of reality reaching leadership has been shaped by what the data infrastructure was designed to draw out — typically the quantitative, the recent, and the aggregated — and has excluded everything the data infrastructure was not built to carry.

When that material becomes analysable, organisations sometimes discover that the patterns were visible all along. The attrition problem that presented as a compensation issue in the survey data was described, repeatedly and specifically, in exit interview notes as a management culture problem in a particular division. The customer satisfaction decline that appeared as a score movement in quarterly reporting had been articulated, in precise terms, in complaint correspondence two years earlier. The service redesign that failed to achieve adoption had generated, during its consultation phase, a clear account of the operational reality it failed to account for — an account that was collected, filed, and never read.

None of this means that quantitative data is wrong or that qualitative analysis is always right. It means that organisations have been working with a partial picture and treating it as complete. The data to fill that picture, in many cases, already exists.

The question is whether the organisation is prepared to read it.

My opinion

I have been through more exit interviews than I can count, and almost all follow the same pattern: questions designed to measure the process, not to capture what the person leaving actually thought. I have asked to read exit interviews from people who left teams I was responsible for and been met, more than once, with a blank look or a direct refusal. A genuine caveat applies: this data is shaped by why someone left, which means it needs to be read in context rather than taken at face value. But there is a difference between data that requires careful reading and data that is never read at all. The true voice of people is the thing most organisations consistently arrange not to hear.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Morgan finds out on a Tuesday evening, scrolling through the news. Not from the company. Not from the bank. Not from anyone who thought they had a duty to say something. From a journalist who found out first.

The article is six paragraphs long. The breach happened seven months ago. Morgan’s name, date of birth, email address, and payment history have been sitting somewhere they were never supposed to be. The notification is still in a queue. The legal team is still reviewing.

By the time Morgan finishes reading, the account has been live in the wrong hands long enough to matter.

The system is working as designed

Companies are not legally required to disclose data breaches immediately. In the UK, under GDPR, the obligation to notify the Information Commissioner’s Office is seventy-two hours. Notification to individuals is required only when there is a likely high risk to their rights and freedoms — and determining that likelihood is left largely to the organisation doing the determining.

The gap between an event happening and a person learning about it is not a failure of the system. It is the system. Seven months is not unusual in documented cases. Twelve is not unheard of. The disclosure process is managed for legal exposure, reputation, and investor sentiment. The person whose data was taken is, structurally, the last consideration.

None of this requires a villain. The process is run by compliance teams doing their jobs inside a framework designed before the scale of digital exposure was fully understood. The problem is not malice. The problem is architecture.

Whose interests the framework serves

The disclosure framework was built around the assumption that a company’s primary obligation in a breach event is to contain it, assess it, and report it in good order. The individual whose data was taken is a downstream consideration — someone to be managed after the institutional response is complete.

This is an understandable way to design a legal framework. It is not an acceptable way to treat a person. The individual whose data was taken has no independent means of knowing what has happened. They cannot check. Cannot act. Cannot protect accounts, change passwords, alert a bank, or do any of the things that might have reduced the damage. The window in which those actions were possible opens and closes while the disclosure is being reviewed.

The data that was taken does not expire. An address does not change. A date of birth does not change. The value to a criminal does not diminish with time. The delay does not reduce the risk — it transfers it entirely to the person.

The company managed the crisis. Morgan just lived inside it.

What the delay actually costs

The visible cost — fraudulent transactions, identity theft, months of remediation — is measurable, if the person can prove causation in time. Most cannot. The link between a breach notification seven months late and the credit application rejected six weeks ago is not the kind of thing an individual can easily trace through a system designed to obscure its own connections.

The invisible cost is something else. Anyone affected comes to understand, in a way that cannot be undone, that there are institutions holding information about them that will manage a catastrophic failure without telling them until the legal minimum has been met. That the decision about when they deserved to know was made by people they have never met, in a process they had no part in, according to criteria they were never shown.

The relationship between a person and the organisations that hold their data is not a relationship between equals. It never was. But the disclosure framework makes the power difference formal, and applies it at the moment a person is most vulnerable.

The notification system, as it currently operates, serves the organisation managing the breach rather than the person affected by it. It assists the company to discharge its legal obligation at a time and in a manner that suits the company’s interests. The individual affected receives a letter when the process is complete, not when the information might have been useful.

A seven-month delay adds nothing to the individual’s position. The argument that thorough investigation improves the quality of notification is technically defensible. In practice, the person receiving the notification has no way to verify the investigation’s quality, no means to challenge the timeline, and no recourse if the delay increased their exposure.

Every person whose data is in a breach has a different threat profile — different accounts, different vulnerabilities, different ability to act quickly. A single delayed disclosure treats all of them identically, at the moment when individual difference matters most. The person who is digitally confident and financially resilient and the person who is neither receive exactly the same letter, at exactly the same time, containing exactly the same limited information.

My opinion

The architecture of delayed disclosure is not the result of poor drafting. It reflects a consistent choice about whose interests the legal system was designed to protect. The notification framework was not designed and then found to serve institutions better than individuals. It was designed to serve institutions. The conversation about how to change it will not begin with regulators or compliance teams. It will begin when enough individuals decide that a letter arriving seven months late is not a notification. It is an alibi.

The questions no letter asks

If your data has been in a breach — and statistically, it has — the notification you received, if you received one at all, told you what the company was legally required to say. It did not ask what you think about the system that produced that notification. These are the questions the letter didn’t come with.

• Have you ever found out about a breach affecting you through a news story rather than a direct notification?

• If you ran a company and discovered a breach, what would stop you from telling your customers immediately?

• Should notification timelines be standardised and enforced — and if so, by whom?

• What would you want a company to say to you, and when, if your data had been compromised?

These questions sit at the centre of my book The Shadow System — specifically the sections on how companies manage reputational risk against actual risk in the aftermath of a breach.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

The pattern has been documented often enough that it has a name. An AI agent is given the instruction to find options — for a flight, a supplier, a candidate. It checks availability, compares prices, retrieves payment details from a linked account, and completes the task. A confirmation arrives before any human has made a decision. The outcome may be correct. The price reasonable. And yet a decision was made that no person made, and if it had gone wrong — wrong date, wrong airport, wrong account charged — the question of what to do next would be genuinely unclear.

This is the agentic AI moment. Not the science fiction version, where a machine plots against its creators. The mundane, already-here version, where AI systems that can act are being given access to the things that matter — calendars, inboxes, bank accounts, business systems — and the line between “help me with this” and “do this” is blurring faster than the accountability structures can keep up.

From chatbot to colleague

The shift from conversational AI to agentic AI is not incremental. It is a change in the nature of what the technology does.

A conversational AI — the kind most people encountered first — responds. You ask it something, it answers, the interaction ends. It can be wrong, it can be inconsistent, but it cannot do anything in the world beyond produce text. The consequences of its failures are bounded.

An agentic AI operates differently. Given a goal, it plans a sequence of steps to reach it, calls on tools — APIs, databases, software interfaces, payment systems — executes each step, monitors the results, and adjusts if something fails. It does not wait for instruction at each stage. It works. This is why the transition from 2023’s chatbots to 2025’s agents felt, to people paying attention, like something more than an upgrade. OpenAI released Operator, Anthropic extended Claude to computer use, Google launched Project Mariner — all in the same period, all enabling AI to interact with real systems on a user’s behalf. The architecture changed. The capability changed with it.

The adoption numbers reflect this. A Zapier survey of enterprise leaders in early 2026 found around 72% were using or testing agentic systems. Gartner forecast that 40% of enterprise applications would embed task-specific AI agents by the end of the year. The agentic AI market, valued at roughly $7.3 billion in 2025, is projected by Fortune Business Insights to reach $139 billion by 2034 — though other analysts project considerably higher figures. These are not pilot-project figures. This is live deployment across large enterprises.

The capability that makes agentic AI effective — the ability to act without waiting for permission at each step — is precisely what makes it difficult to govern.

The use cases making the business case are concrete enough. An AI agent that handles the full lifecycle of an insurance claim — reading the form, cross-referencing policy, assessing evidence, initiating payment — compresses a process that once took days into minutes. Walmart’s agentic supply chain system detects demand signals, adjusts procurement plans, and reroutes logistics without human triggers, with reported reductions in out-of-stock incidents across pilot regions. In software development, agents now write, test and deploy code: Stack Overflow’s 2025 developer survey found 84% of developers using AI tools, tools that now generate around 41% of all code written. Research published by Landbase, a sales automation firm, puts average returns from mature agentic deployments at 171% — roughly three times the return from traditional automation.

These are genuine improvements. They are also the easy half of the story.

The accountability gap

The hard half concerns what happens when the agent is wrong.

In 2024, a Canadian tribunal held Air Canada liable after its customer service chatbot gave a grieving passenger incorrect information about bereavement fare discounts. Air Canada’s defence — that the chatbot was a separate legal entity responsible for its own statements — was rejected without hesitation. The airline was responsible. The principle was unambiguous: organisations are accountable for what their AI systems say and do. The machine cannot be sued. The humans behind it can.

That principle becomes considerably more complicated when the AI is not talking but acting. A wire transfer sent to the wrong account. A contract clause accepted on behalf of a business. A medical appointment cancelled because an agent misread a scheduling instruction. In each case, the action has real-world consequences that may be difficult or impossible to reverse, the chain of decisions that led to it may involve multiple models, APIs and data sources, and the human who nominally authorised the deployment may have had no visibility into the specific step that went wrong.

The governance figures sit uncomfortably alongside the adoption figures. A 2025 industry survey found only 44% of organisations had formal AI governance policies in place. Only 11% of enterprises had agentic systems actually running in production rather than pilot. Deloitte found that 42% of organisations were still developing their agentic strategy, and 35% had no formal strategy at all. A Fortune report published in March 2026 — drawing on research from Wharton and Accenture — put the problem plainly: AI agents were spreading across the enterprise value chain, often ahead of formal strategy and governance.

The Wharton Accountable AI Lab framed it as a supply chain problem. Accountability is not a single organisation’s responsibility — it runs from the model developers who trained the system, through the platform providers who packaged it, to the organisations that deployed it, and finally to the humans who were supposed to be supervising it. When something goes wrong, everyone points to someone else. And in a system where a single agent action might involve five external data sources, three AI models, and four different APIs, the pointing is genuinely difficult to resolve.

The human in the loop — or not

The phrase the industry has settled on is “human in the loop” — the idea that a person remains involved in agentic decisions, available to check, correct or override. It is a reassuring phrase. In practice, it describes a spectrum, and much of the deployment happening now sits at the end where the human is less supervisor than emergency contact.

Microsoft has coined a different formulation: the “agent boss” — the person who builds agents, delegates to them, and manages them. The role is real and the framing is honest. But it also makes visible something the more comfortable language obscures. If one person is now the agent boss for a fleet of AI systems each capable of taking hundreds of actions per day, the ratio of human attention to autonomous action has shifted dramatically. The human is not in each loop. They are above all the loops, available to intervene if something surfaces — which requires that something surface, which requires that the systems are designed to surface it.

Fortune’s March 2026 report documented what it called “shadow AI” — the use of AI tools, including agentic systems, that employees bring into the workplace outside formal IT channels, beyond the visibility of any governance structure. Microsoft’s 2025 research found around three-quarters of knowledge workers were already using AI tools; a significant proportion were doing so with tools their employers had not sanctioned and could not audit. An orphaned agent — one running in production with no designated owner, accessing systems its original designers never intended — does not typically go rogue. It accumulates risk silently, its permissions unreviewed, its behaviour unchanged as the business context around it evolves.

The human is not in each loop. They are above all the loops — available to intervene if something surfaces. Which requires that something surface.

None of this is an argument against agentic AI. The productivity gains are real. The compression of tedious, error-prone, time-consuming processes is real. The question is not whether to use agents. It is whether the organisations deploying them are building accountability in from the start, or treating governance as something to address after the value has been captured. The evidence, on balance, suggests the latter is happening more often than the former.

The test most deployments are failing

Most current deployments pass the first question easily. The AI that clears an inbox backlog, processes a claims queue, or monitors a live system for failures is doing work that was real, time-consuming and often unrewarding. People who use these tools report genuine relief. That part is not in doubt.

The harder question is whether the agent adds something — capability, judgement, insight — or whether it substitutes for human understanding in ways that hollow out the understanding itself. There is a difference between an AI that surfaces relevant information so a person can make a better decision, and an AI whose recommendations are so compelling, and whose interface so frictionless, that human approval becomes a formality. A hospital AI diagnostic study cited in the Wharton and Accenture report found exactly this pattern: at high-performing sites, clinicians understood the system’s confidence scores and knew when to question their findings; at lower-performing sites, they either followed the tool uncritically or ignored it entirely. The tool was the same. The human relationship with it was not.

The third question — whether the agent responds to the specific person it serves, or treats every user as the average case — is the one most consistently failed. The documented pattern of agents interpreting “find” as “do” is a small example: the system optimised for task completion and treated the distinction as a rounding error. At the scale of enterprise deployment, that rounding error multiplies. The question of whether the people affected by agentic decisions — employees, customers, citizens — are treated as participants or as variables is not a technical question. It is a design choice.

Sources

• Agentic AI adoption — 72% of enterprise leaders: Zapier, State of Agentic AI Adoption Survey 2026. Survey of 500+ enterprise leaders. zapier.com/blog/ai-agents-survey/

• Gartner — 40% of enterprise applications to embed AI agents by end of 2026: Gartner press release, 26 August 2025: “Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, Up from Less Than 5% in 2025.” gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026

• Agentic AI market size — $7.3 billion (2025) projected to $139 billion (2034): Fortune Business Insights, Agentic AI Market Size, Share & Forecast Report 2026–2034. fortunebusinessinsights.com/agentic-ai-market-114233. Note: other analysts project higher 2034 figures (Precedence Research: $199 billion; Grand View Research: $183 billion by 2033).

• OpenAI Operator, Anthropic Claude computer use, Google Project Mariner: Released in late 2024 and early 2025. Confirmed via product announcements from OpenAI, Anthropic, and Google.

• Walmart agentic supply chain system: Walmart Global Tech blog; Supply Chain Dive, “4 ways Walmart is scaling AI to unify its supply chain,” 2025. tech.walmart.com; supplychaindive.com/news/4-walmart-supply-chain-ai-uses/760891/

• Developer AI tool usage — 84% of developers; 41% of code: Stack Overflow, 2025 Developer Survey. stackoverflow.com

• Returns on agentic AI deployments — 171% average ROI: Landbase, 39 Agentic AI Statistics Every GTM Leader Should Know in 2026. landbase.com/blog/agentic-ai-statistics. Note: Landbase is a commercial sales automation firm; this is industry research rather than independent analysis.

• Air Canada chatbot liability ruling: Moffatt v Air Canada, British Columbia Civil Resolution Tribunal, 14 February 2024. Reported by CBC News: cbc.ca/news/canada/british-columbia/air-canada-chatbot-lawsuit-1.7116416

• AI governance policies — 44% of organisations: Corroborated across multiple 2025 industry surveys including IAPP AI Governance Profession Report 2025 (iapp.org) and SailPoint enterprise security research.

• Only 11% of enterprises with agentic systems in production; Deloitte — 42% developing strategy, 35% no strategy: Deloitte, 2025 Emerging Technology Trends Survey (500 US technology leaders, June–July 2025). deloitte.com/us/en/about/press-room/state-of-ai-report-2026.html

• Fortune / Wharton / Accenture report — March 2026: “Intelligence may be scalable, but accountability is not,” Fortune, 26 March 2026. Joint research from Accenture’s global products practice and the Wharton AI & Analytics Initiative. fortune.com/2026/03/26/ai-agents-accountability-accenture-wharton-report/

• Wharton Accountable AI Lab: Led by Professor Kevin Werbach, Department of Legal Studies and Business Ethics, Wharton School, University of Pennsylvania. ai-analytics.wharton.upenn.edu/wharton-accountable-ai-lab/

• Hospital AI diagnostic study — high- vs low-performing sites: Cited in the Wharton and Accenture report (Fortune, March 2026). See above.

• Shadow AI / knowledge workers using unsanctioned AI tools: Microsoft, 2025 Work Trend Index. Microsoft’s research found approximately 75% of knowledge workers using AI tools at work, with a significant proportion using tools outside formal IT channels. microsoft.com/en-us/worklab/work-trend-index

• Microsoft “agent boss” framing: Microsoft Copilot product communications and AI documentation, 2025–2026. microsoft.com/en-us/microsoft-copilot

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Every year, organisations across every sector spend significant money finding out how engaged their people are. The answer comes back as a number. Sixty-two percent. Seventy-one. A red, an amber, a green. And then, in most cases, very little changes — because a number, however precisely measured, does not tell you what to do next.

This is not a failure of effort. It is a failure of method. The question being asked — how engaged are people, on a scale of one to ten — is structurally incapable of producing the intelligence needed to act. It tells you the temperature of the room. It does not tell you why the room is cold, which walls are letting in the draught, or whether the people in it have stopped expecting anyone to fix it.

The pattern holds across every sector. An NHS trust running annual staff surveys. A private equity-backed scale-up tracking employee sentiment through quarterly pulse checks. A charity trying to understand what its service users actually need from a redesigned programme. A number where there should be a conversation.

What the number misses

The case for quantitative engagement measurement is not wrong. Trends matter. Comparability matters. Knowing that satisfaction dropped twelve points between January and March, or that one team consistently scores lower than the rest of the business, is genuinely useful. The problem is not that organisations measure; it is that most treat the measurement as the understanding, rather than as a signal that understanding is needed.

Qualitative intelligence — what people actually say, in their own words, about their own experience — produces something different. It surfaces the specific, the unexpected, and the things no survey designer thought to ask about. It also surfaces disagreement within a group that aggregate scoring conceals. A team with an average engagement score of sixty-eight might contain people who are deeply committed and people who are quietly preparing to leave. The number averages them into false coherence.

The difficulty has always been scale. Reading and coding open-text responses from three hundred employees, or three thousand service users, is expensive and slow. The analysis introduces its own biases. And the time between collecting the data and acting on it is long enough that the context has already shifted. So organisations default to the form that is easiest to process — the number — and accept its limitations as the cost of manageability.

That trade-off no longer holds as firmly as it did. The combination of open-response collection, peer voting, and AI-assisted analysis — the approach taken by tools like Gobby, which works in this space — means that qualitative data can now be gathered and synthesised at a speed and scale that was not previously practical. People respond in their own words. Those responses are surfaced to peers, who vote on what resonates. The result is not just a fuller picture; it is a picture that has been validated by the community itself, not just interpreted by the analyst.

The pattern across sectors

In the public sector, the gap between what is measured and what matters is often widest. Patient experience surveys ask whether care was clean, timely, and respectful — categories defined by the regulator, not the patient. What they rarely capture is whether the patient felt heard, whether the treatment made sense in the context of their actual life, or whether the system’s definition of a good outcome matched their own. The number that comes back is real. The experience it describes is partial.

Local government consultation has the same problem at a different scale. A planning consultation that asks residents to rate a proposal on a five-point scale produces a mandate for or against. It does not produce understanding of why, which aspects are acceptable and which are not, or what a modified proposal might need to look like to bring the sceptics with it. The binary produces a decision. It does not produce the intelligence that would make the decision stick.

In the private sector, the equivalent is the employee engagement survey and its close relative, the customer satisfaction score. Both have become so ritualised that the measurement has partially replaced the thing it is supposed to measure. Teams know how to answer the survey. Customers know that the score goes nowhere. The process continues because it produces a defensible number, not because it produces understanding. When engagement drops, the standard response is a communications campaign — because the survey told you sentiment, not cause.

The charity and social enterprise sector is arguably most exposed. Organisations whose entire purpose is to serve and represent specific communities often have the least developed means of understanding whether they are doing so. Impact reporting is dominated by output metrics — number of people served, sessions delivered, referrals made. What is missing is the voice of the person at the end of the service, speaking in their own terms about whether it changed anything for them. Not a satisfaction score. A considered account of what happened.

Understanding you can act on

The value of more detailed engagement data is not primarily analytical. It is operational. When the reasons behind a trend are visible — not inferred, but stated by the people who are living them — the path to action becomes clearer and the likelihood of getting it wrong decreases.

A portfolio company workforce that scores low on engagement is a problem. A portfolio company workforce that is able to articulate, collectively and specifically, that the integration plan removed the three things that made the culture work is a solvable problem — or at least one where the board can make an informed decision about the trade-off. A service user group that can say, in their own words and with their peers’ endorsement, that the new referral pathway is harder to navigate than the one it replaced gives a programme team something to act on rather than something to commission further research about.

My opinion

Every organisation I’ve worked with runs some version of the same survey — historical questions, coded to a scale, returned as a number. I’ve seen what happens when you replace that with something closer to in-the-moment: a city authority that put real-time word clouds — drawn from live citizen responses — on a screen in the corridor where the chief executive walked every morning. The conversation in that building changed. Not because the data was new, but because it was immediate, in people’s own words, and impossible to walk past. That is the difference between recording sentiment and understanding it.

The shift required is not primarily technological. The tools exist. What is required is a change in what organisations believe engagement measurement is for. If the purpose is compliance — to show that a survey was conducted and a score was recorded — then the number is sufficient. If the purpose is understanding — to know what is actually happening and why, well enough to do something about it — then the number is a starting point, not an answer.

The organisations that close that gap are not the ones that find a better survey platform. They are the ones that decide the distinction matters.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Alex had been on the waiting list for eighteen months before the letter arrived. It said that the authority’s housing needs assessment had been completed and that, based on the information held, Alex’s application had been assigned a priority band. Band C. The letter did not explain what Band C meant in practice. It did not describe how the assessment had reached that conclusion. It mentioned, in the third paragraph, that a decision-support system had been used to process applications, and that appeals could be submitted in writing within twenty-one days.

Alex submitted an appeal. The appeal was considered. The band was confirmed.

At no point in this process did Alex learn what information the system had used to arrive at its assessment, what weight it had placed on different factors, or why Alex’s particular circumstances — a care dependency, a fixed income, a medical condition that made the current address untenable — had resulted in Band C rather than Band B. The letter existed. The process had been followed. The appeal had been considered. The outcome held.

Somewhere in that housing authority, a board had approved the system. The paper had been well-prepared. The numbers were credible — the system processed applications in a fraction of the previous time, reduced inconsistency across manual assessors, and met the procurement criteria the implementation team had set out. The board had asked about cost, about timeline, about how the system would handle exceptions. It had asked about data retention and compliance obligations. Nobody had asked what Band C would mean to Alex.

The gap was not in the preparation

The people in that boardroom were not negligent. They were competent, experienced, and attentive. The chair ran a disciplined meeting. The questions raised were reasonable given the information in front of the room. And the information in front of the room had been assembled by people who understood the system — who had selected it carefully, procured it properly, and built a reasonable implementation plan.

The gap was not in the preparation. It was in the scrutiny.

To ask what Band C would mean to the people who received it required knowing enough about how the system worked to understand that the band was not an administrative category but an algorithmic output — shaped by weighted factors, training data, and configuration decisions made during implementation. To ask whether the appeal was a real avenue of recourse required understanding that an appeal reviewed against the same criteria that produced the original decision is structurally different from an appeal reviewed by an independent assessor with full information. To ask whether the failure modes had been examined required knowing that automated decision systems fail differently from human assessors — not randomly, but systematically, concentrated in the people whose circumstances differ most from the data the system was trained on.

Those questions were not asked. Not because they were unanswerable. Because they were not visible to the people in that room.

The risk that doesn’t show up in a financial model

Governance models were designed for a world in which the primary risks of a large technology procurement were financial and operational. Could the organisation afford it? Would it be delivered on time? Would it integrate with existing infrastructure? Did the vendor have a track record? These are legible questions for a board with deep expertise in finance, law, and organisational management. The methods for evaluating them are established. The signals of risk are familiar.

AI and data-driven decision systems carry a different kind of risk. The failure is not that the system goes over budget, though it may. The failure is that the system makes consequential decisions about real people based on patterns extracted from historical data — and if that data reflects the inequities and assumptions of the processes it was trained on, the system reproduces those inequities at scale, faster, with the additional authority of technological neutrality. That failure does not appear in a financial model. It does not surface in a project risk register. It does not emerge in a compliance review that covers data retention and lawful processing but not the distributional consequences of the decisions being made.

It surfaces in a letter.

The boards approving these systems are not, in the main, composed of people with the literacy to see that risk. That is not a criticism of the individuals — it is a description of how board composition evolved in response to the risks governance processes were built to manage. For most of the history of organisational governance, technology decisions whose consequences were primarily operational could be evaluated through financial and legal lenses without significant loss. That is no longer reliably true.

The structure of most board meetings compounds the problem. Papers are summarised to one page. Technical detail is compressed into reassurance. The people with the deepest knowledge of the system are usually the people most invested in its approval. Independent scrutiny — not from the vendor, not from the implementation team, but from someone whose only brief is to look for what could go wrong — is not a standard feature of most technology governance papers. It is an occasional addition, when the stakes are visible enough to prompt it. The stakes were not visible. They were in Band C.

In the EU, the AI Act — in force since August 2024 — classifies AI systems used to assess housing eligibility as high-risk, requiring transparency, human oversight, and challengeable outputs from August 2026. In the UK, the government’s Algorithmic Transparency Recording Standard is mandatory for central government departments; a bill that would have extended comparable requirements to local housing authorities stalled in 2024 without progressing. For the authority that assessed Alex’s application, and thousands like it, no such requirement currently applies.

Until they do, the room that approves the system and the person who receives its output will remain two entirely separate worlds.

The accountability that disperses

When an AI-assisted decision system produces a wrong outcome, the accountability for it is distributed across enough points in the governance chain — the vendor, the implementation team, the commissioning officers, the board itself — that it disperses without fixing to any of them. The appeal process, if it reviews the original decision against the same criteria that produced it, is not a genuine avenue of recourse. It is a confirmation mechanism.

Alex’s appeal eventually succeeded — Band B assigned, not because the system was reviewed, but because one housing officer with the right information made a different call. The system continued processing the next two thousand applications unchanged.

This is the pattern, not the exception. An authority reviewing its own AI deployment will see a system operating as designed — applications processed, appeals handled, no formal complaints logged against the technology. The harm does not show up in a post-implementation review. It accumulates in the cases where the right advocate was not present, where the first appeal was the last one, and where the person whose circumstances fell outside the training data had no way of knowing that fact.

A revised band and a revised system are two very different outcomes that look, from inside the authority, like the same one.

Three questions that change what gets asked before approval

The questions that were not asked in that boardroom are not technically difficult. They do not require a specialist in machine learning or algorithmic auditing. They require a different starting point — one that begins with the person at the end of the system rather than the system itself.

Before approving a technology that will make consequential decisions about people, a governance body should be able to answer three things.

The first is whether this technology actually helps the people it processes — not faster, but better. Does it make the experience easier for the person submitting an application under stress, with limited time, with a care dependency that doesn’t fit neatly into any field? The paper that came to that board had been prepared by the people implementing the system. The people at the end of it had not been consulted. Whether that consultation is required — as a governance condition, not a supplementary appendix — is a governance choice.

The second is whether the technology adds something real, and to whom. A system that processes applications faster is not automatically one that serves applicants better. It may serve the organisation better — reduced staffing costs, greater throughput, more consistent output — while producing outcomes that are worse for the specific people the organisation exists to help. The board had asked about processing time. It had not asked what the improvement looked like from the side of the person submitting the application.

The third is whether the technology can respond to the individual circumstances of the people it affects — and what happens to the people whose circumstances fall outside the parameters it was trained on. That question, asked before approval, changes the procurement criteria. It requires the vendor to demonstrate that edge cases are handled with appropriate human judgement, not simply processed by the same logic as every other case. It makes an independent appeal process a governance condition of deployment rather than a feature added after complaints arise.

A board that cannot answer those three questions before approving a system has not governed the technology. It has approved a budget line. The governance is what happens in between.

My opinion

Most organisations are not prepared for what is coming. If they treat AI governance as a variant of the regulatory models they already have — compliant, procedural, box-ticking — it will fail. We have not yet reached agreement on what AI actually is, which makes governing it a peculiar challenge: the responsibility sits with people who cannot reliably describe the thing they are responsible for. The relevant legislation is also not contained within a single jurisdiction — a service consumed by a user in another country may bring obligations the commissioning authority has never considered. That is the gap. It is wider than most boards currently know.

Before the next approval

• Could the people who approved a significant technology decision at the senior level of an organisation you know describe, in plain language, what it does and what its consequences are for the people at the end of it? If not — what does that say about the quality of the oversight it received?

• If you sit on a board or senior governance body, what is your personal level of literacy in the specific technologies your organisation uses or is considering? Is that literacy sufficient to probe the assumptions in the papers you are asked to approve — or are you relying on the competence of the people who prepared them?

• If you are a technology leader who presents to boards, what is your responsibility to ensure genuine understanding rather than informed consent? Are you building presentations that equip boards to scrutinise, or ones that build confidence in a decision already made?

• And the hardest one: is it possible to be accountable for a decision you did not understand? If accountability requires understanding, and the understanding was not present in the room, where does the accountability for what happened to Alex actually sit?

The board never knew Alex’s name. That is, in the end, the clearest description of the problem.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

In September 2023, as Hurricane Lee was curving northward west of Bermuda, meteorologists at NOAA and the National Hurricane Center were working to determine where it would make landfall: New England or farther east, in Canada. The sooner they could call it, the earlier they could issue warnings. Evacuations for a major hurricane require at least three full days. Every additional day of warning is not a planning luxury. It is a measurable difference in how many people can get out.

The traditional forecast models locked in on Nova Scotia about six days before landfall. A different model — an experimental AI system called GraphCast, developed by Google DeepMind and running as a live test on the European Centre for Medium-Range Weather Forecasts website — had already called it three days earlier. Nine days out, GraphCast was pointing to Nova Scotia with the same confidence that conventional models would not reach for another 72 hours.

Three days. In the context of hurricane preparedness, that is not a technical curiosity. It is the difference between an evacuation that works and one that does not. It is the difference between a ferry that leaves the island while the roads are still passable and one that does not. It is, in the most direct sense, lives.

This is the AI weather forecasting story — not the version about supercomputers and equations, but the version about what happens in the gap between when a forecast becomes reliable and when people have enough time to act. That gap is shrinking. The implications run further than most discussions of AI and climate have yet acknowledged.

How weather prediction works, and why it is hard

Traditional weather forecasting is an extraordinary intellectual achievement. Starting in the early twentieth century, researchers proposed that the behaviour of the atmosphere could be modelled mathematically — that the physical equations governing fluid dynamics, thermodynamics and radiation could be translated into algorithms, fed with observations from weather stations, balloons, satellites and ocean buoys, and used to project how the atmosphere would evolve over days ahead. They were right. By the 1960s, when computers became powerful enough to run these models, numerical weather prediction began to demonstrate real skill. It has been improving ever since at a rate of roughly one useful day per decade — today’s six-day forecast is as accurate as the five-day forecast was ten years ago.

These traditional models are based on physics. They represent the atmosphere as a three-dimensional grid, apply equations at each grid point, and propagate the simulation forward in time. The equations are derived from first principles — conservation of mass, energy and momentum. When the model produces a prediction, you can trace exactly why: this air mass moved because of that pressure gradient, that front formed because of these temperature differences. The physics provides both the accuracy and the interpretability.

The cost is computation. Running a ten-day global forecast using the European Centre for Medium-Range Weather Forecasts’ flagship model — widely regarded as the world’s best — requires hours of runtime on one of the largest supercomputers in the world, consuming enormous amounts of energy. The ECMWF’s computing infrastructure costs hundreds of millions of pounds to build and operate. Most countries cannot afford anything comparable, which means most of the world’s weather forecasting depends on a handful of institutions with sufficient resources to run the physics.

GraphCast produces a ten-day global forecast in under a minute on a single machine. The equivalent physics-based forecast requires hours on one of the world’s largest supercomputers. The cost gap is measured in orders of magnitude.

AI weather models work differently. Rather than encoding the physics of the atmosphere from first principles, they learn from data. GraphCast was trained on four decades of ECMWF reanalysis data — a historical reconstruction of global weather conditions combining observational records with traditional models to produce a consistent dataset — and learned to predict how weather states evolve by recognising patterns in that history. It does not calculate what the atmosphere will do based on physics. It recognises what atmospheric configurations have historically led to what subsequent states, and applies that learned relationship to new initial conditions.

The result, in terms of accuracy, is remarkable. A paper published in Science in November 2023 showed that GraphCast outperformed ECMWF’s best deterministic model on more than 90% of 1,380 verification targets. Huawei’s Pangu-Weather, published in Nature in 2023, achieved comparable results and runs 10,000 times faster than traditional ensemble models. ECMWF moved its own AI-based model — AIFS, the Artificial Intelligence Forecasting System — to operational status in February 2025, making it the first major meteorological agency to deploy an AI model as part of its official forecast suite. NOAA launched its own suite of AI-driven global weather models in late 2025, built on GraphCast foundations and fine-tuned with NOAA’s own data, requiring up to 99.7% fewer computing resources than their traditional counterparts.

What AI can and cannot do

The models are as notable for what they cannot do as for what they can.

The models excel at medium-range forecasting — the three-to-ten-day window where most consequential weather decisions are made. They struggle with extreme events that fall outside their training data. This is not a minor limitation. Hurricane Otis, which intensified from a tropical storm to a Category 5 hurricane in roughly twelve hours before striking Acapulco in October 2023 — killing more than fifty people and destroying most of the city’s infrastructure — is precisely the kind of event that AI models trained on historical patterns may fail to capture. Rapid intensification over warm ocean water, in conditions that push against the edge of what the historical record contains, is where pattern recognition may fail and physics-based modelling may have an irreplaceable advantage.

The AI models also depend entirely on traditional physics-based systems for their input data. They cannot generate a forecast from raw observations alone — they require a processed, quality-controlled atmospheric state as their starting point, which is produced by the same supercomputer-intensive data assimilation systems that underpin conventional forecasting. If those systems fail or degrade, AI forecasting degrades with them. The models are not independent of the infrastructure they appear to make redundant.

The most sophisticated agencies are already pursuing a hybrid approach that addresses both sides of this. NOAA’s HGEFS — the Hybrid Global Ensemble Forecast System — combines AI-based ensemble forecasts with traditional physics-based ensemble modelling. In early testing, this hybrid consistently outperformed both the AI-only and physics-only systems. The combination captures the AI’s computational efficiency and its skill at medium-range pattern recognition, while the physics-based component provides a check on situations where the atmospheric state is genuinely novel.

There is a further limitation that ECMWF’s own documentation is direct about: AI models do not yet explain their forecasts in the way that physics-based models can. A traditional model can trace the chain of physical causation that led to a prediction. An AI model can say what it predicts but cannot always say why in terms a forecaster can interrogate. When an AI forecast diverges from a physics-based forecast — when the two disagree significantly — that disagreement is valuable information about forecast uncertainty. But resolving the disagreement requires human expertise. The role of the meteorologist is shifting: from running models and reading their output, toward interpreting the ensemble of AI and physics forecasts, identifying where they diverge, and communicating uncertainty to the people who need to make decisions.

What changes when forecasts get earlier and cheaper

Emergency preparedness is the most direct consequence. Hurricane evacuation orders typically need to be issued three to four days in advance to be effective — that is the minimum time required to move a large population out of a coastal zone before a major storm arrives. A model that identifies the landfall region nine days out rather than six gives emergency managers an extra three days to coordinate, to communicate, to move people who cannot move themselves. AI models are extending that window.

Agriculture is the second major domain. Planting, irrigation, harvest and frost-protection decisions all depend on weather forecasts. Farmers in rain-fed agricultural systems — which account for the majority of global food production — manage risk primarily through weather information. A more accurate ten-day forecast, available to a farmer with a smartphone, changes the calculation for whether to plant this week or next, whether to apply irrigation before a forecast rain event or wait. Google DeepMind’s WeatherNext 2 now powers weather information across Google Search, Gemini, Pixel Weather, and the Google Maps Platform Weather API. The agricultural implications of that scale of deployment in low-income farming communities have not yet been fully studied, but access to forecast information that was previously available only to well-resourced agricultural operations is now reaching smallholder farmers for the first time.

Energy system management is the third. Renewable energy generation is inherently variable — wind and solar production depend directly on weather conditions. Grid operators managing the balance between generation and demand need accurate forecasts of wind speeds and solar irradiance to plan how much backup capacity to hold in reserve and when to bring it online. GenCast, Google DeepMind’s probabilistic AI forecasting model, reduces wind power forecasting errors by up to 20% within a two-day lead time compared to traditional ensemble models. At grid scale, a 20% reduction in forecast error translates directly into lower costs for consumers and more efficient integration of renewable energy.

The implications for climate science itself are harder to measure but potentially the most significant of all. The same AI techniques that underpin operational weather forecasting are now being applied to climate modelling at longer timescales — predicting seasonal and sub-seasonal patterns months ahead. NOAA’s Atlantic Oceanographic and Meteorological Laboratory and its Weather Program Office’s Subseasonal to Seasonal Research programme are working toward extending extreme weather forecast lead times from the current two-to-four-day window to two-to-four weeks. If achievable, that shift would change disaster preparedness entirely. A community that knows six weeks in advance that conditions are likely to produce a severe hurricane season, or an anomalous flooding pattern, or an unusual drought, can prepare differently than one that gets four days’ notice.

A more accurate ten-day forecast, on a smartphone, changes the calculation for a farmer deciding whether to plant this week or next. That is not a marginal improvement. At the scale of global food production, it is consequential.

Traditional numerical weather prediction at global scale requires infrastructure that only a handful of institutions worldwide can afford. GraphCast-quality ten-day forecasts can run on a single laptop. NOAA’s AI models require a fraction of a percent of the computing resources of their traditional counterparts. The same countries that have historically depended on ECMWF or NOAA data for their weather services — because they cannot afford to run their own global models — can now run inference on AI models with modest infrastructure. This is not a complete solution to the observation network problem (AI models still need input data from the global observing system that wealthy countries have historically funded), but it opens a door to greater national meteorological capacity that was previously closed by computational cost.

The forecast that arrives and the one that doesn’t

AI weather forecasting helps people anticipate the future well enough to make better decisions, and the improvement is not marginal. Three extra days of hurricane warning, more accurate flood alerts for river communities downstream from atmospheric river events, earlier seasonal outlooks for farmers facing planting decisions under climate uncertainty — these are real improvements. The technology is operational. These improvements are already being delivered.

The question of who receives that improvement is harder. AI weather forecasts are available through Google Maps, through NOAA’s public systems, through the ECMWF website. The infrastructure to receive a forecast is a smartphone. The infrastructure to act on it — the evacuation route, the early warning system connected to the government emergency network, the insurance product that adjusts premiums based on forecast risk — is not universally available. The forecast that tells a coastal community in Bangladesh nine days in advance that a major cyclone is forming in the Bay of Bengal is only useful if there is a system in place to act on that information. The observation networks that produce accurate initial conditions for AI models are denser in wealthy countries than in poor ones. The forecast quality is not equal.

The problem that matters most is the one where AI models are least equipped. The climate system is changing. Atmospheric rivers are intensifying. Hurricanes are undergoing rapid intensification more frequently. The Arctic sea ice that provides boundary conditions for mid-latitude weather is disappearing. AI models trained on historical patterns may be systematically less reliable in the conditions that climate change is producing — precisely the conditions for which the most accurate forecasts matter most. The hybrid approach being pursued by NOAA and ECMWF — combining AI pattern recognition with physics-based modelling — is the appropriate response. But it requires sustained investment in both the AI systems and the physical science infrastructure they depend on. Cutting observation networks to fund AI forecasting would be a category error.

Sources

• Hurricane Lee (2023) and GraphCast prediction: National Hurricane Center Tropical Cyclone Report — Hurricane Lee (AL132023). National Oceanic and Atmospheric Administration. https://www.nhc.noaa.gov/data/tcr/AL132023_Lee.pdf

• GraphCast performance benchmarks: Lam, R. et al. (2023). “Learning skillful medium-range global weather forecasting.” Science, 382(6677). doi: 10.1126/science.adi2336. Published November 2023. Confirms >90% of 1,380 verification targets outperformed ECMWF HRES.

• Pangu-Weather: Bi, K. et al. (2023). “Accurate medium-range global weather forecasting with 3D neural networks.” Nature, 619, 533–538. Published July 2023. doi: 10.1038/s41586-023-06185-3.

• ECMWF AIFS operational status: “ECMWF’s AI forecasts become operational.” ECMWF, February 2025. https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ai-forecasts-become-operational. AIFS version 1.0.0 implemented 25 February 2025.

• NOAA AI-driven weather models: “NOAA deploys new generation of AI-driven global weather models.” NOAA press release, 17 December 2025. https://www.noaa.gov/news-release/noaa-deploys-new-generation-of-ai-driven-global-weather-models

• Hurricane Otis (2023): National Hurricane Center Tropical Cyclone Report — Hurricane Otis (EP182023). National Oceanic and Atmospheric Administration. https://www.nhc.noaa.gov/data/tcr/EP182023_Otis.pdf

• GenCast: Price, I. et al. (2024). “Probabilistic weather forecasting with machine learning.” Nature, 637, 84–90. doi: 10.1038/s41586-024-08252-9. Confirms ~20% CRPS improvement over ENS at 2-day lead times.

• WeatherNext 2: Google DeepMind. “WeatherNext 2: Google DeepMind’s most advanced forecasting model.” Google Blog, November 2025. https://blog.google/innovation-and-ai/models-and-research/google-deepmind/weathernext-2/

  ---

  You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

By the time the information about what is actually happening in an organisation reaches the people who make decisions about it, it has usually been through several hands. A front-line worker describes something to a team leader. The team leader includes a version of it in a report to a manager. The manager summarises the pattern to a director. The director presents a slide to the board. At each stage, something is selected, something is compressed, and something is lost. The decision that eventually gets made rests on a version of the original experience that none of the people who had that experience would fully recognise.

This is not a management failure. It is a structural one. Every organisation beyond a certain size faces the same problem: the people with the most direct knowledge of what is working and what is not are not the people with the authority to change it. Connecting those two groups requires a chain of translation. And translation, by its nature, introduces the translator.

The question is not whether this happens. It does, in every sector, at every scale. The question is how much gets lost, and whether the organisations that depend on that intelligence have understood what the loss is costing them.

What the translation removes

The first thing that translation removes is specificity. A front-line nurse can describe exactly what happens when a discharge pathway breaks down — which step fails, what the patient experience is at that moment, what the workaround looks like and why it is used. By the time that account has been summarised into a patient experience metric, the specific has become a category and the category has become a score. The score moves up or down across reporting periods. What it cannot carry is the texture of the original account: the detail that would allow a decision-maker to understand not just that something is wrong but precisely what, and what a plausible fix would need to address.

The second thing translation removes is dissent. A workforce that contains people with genuinely divergent views on a strategic direction will, in most organisations, produce a management layer that presents a consolidated position. Not because dissent is suppressed — though sometimes it is — but because the function of a management summary is to synthesise, and synthesis flattens disagreement into a dominant view. The minority position, which may be the more accurate one, disappears in the transition from conversation to report.

The third thing translation removes is timing. The experience of a customer, a patient, a service user, or an employee exists in real time. The reporting of that experience is always retrospective. By the time a quarterly engagement survey is analysed and presented, the conditions it describes may have shifted. By the time a pattern identified in a management report reaches a board decision, the people who articulated it may have adapted, left, or stopped reporting. The intelligence arrives after the moment when it could most usefully have been acted on.

The same structure in different rooms

The ward sister who cannot get a staffing concern to the executive team without it being reframed as a rota management issue. The retail team leader whose customer feedback about a pricing change never makes it past the operations layer because it conflicts with a decision already made at the centre. The charity caseworker whose account of a service user’s experience is condensed into an outcome category before it reaches the funder’s impact report. The engineer at a scale-up who knows the architecture cannot support the growth plan but has learned that raising it produces a request for a document rather than a conversation.

These are not aberrations. They are the same structural pattern — the distance between the person with the direct knowledge and the person with the authority — expressed in different rooms, with different consequences. In each case, the translation that happens between those two people is doing something specific: it is making the original account manageable for the organisation’s existing processes. And in doing so, it is making it less true.

The individual framing is not irrational. It is what an existing structure can respond to — a named person, a specific failure, a defined gap that a programme can address. The structural framing asks a harder question: not who failed to pass it up, but how many translations the account had to survive before it reached anyone with the authority to act, and what that number is a consequence of.

The organisations that notice this tend to describe it in individual terms. A particular manager is not passing things up. A specific team has stopped raising concerns. A function is not feeding the right information into the planning process. The response is usually a communications initiative or a culture programme — an attempt to change the behaviour of the individuals in the chain.

The organisations that solve it tend to describe it in structural terms. The chain itself is the problem. The number of translations between the original voice and the decision-maker is the variable that needs to change. Everything else is managing the symptoms.

Reducing the translations

The case for qualitative engagement tools — open response, peer validation, AI-assisted synthesis — is usually made on the grounds of richer data. More nuance than a score. More specificity than a category. That is true, but it understates the more significant structural point.

When people respond in their own words, and when those responses are validated by their peers before they reach an analyst, one layer of translation is removed entirely. The manager’s interpretation does not sit between the worker’s account and the organisation’s understanding of it. The researcher’s coding scheme does not determine which themes are visible and which are not. The community itself has indicated what matters and in what terms — and that signal reaches the people reading it in a form closer to the original than any summary process can produce.

This does not eliminate the distance between front-line experience and senior decision-making. It does not remove the organisation’s ability to ignore what it hears. What it changes is the fidelity of the signal — and with it, the organisation’s ability to claim, in good faith, that it did not know.

In a post-acquisition integration, that fidelity is the difference between understanding what the workforce is experiencing and receiving a management view of what the workforce is experiencing. In a public service redesign, it is the difference between knowing what patients or residents actually think and knowing what the engagement process was designed to find out. In a charity measuring its own impact, it is the difference between the beneficiary’s account of what changed and the output metric the funder asked for.

The voice that does not reach the room does not disappear. It goes somewhere else — into attrition, into disengagement, into the gap between what an organisation believes it is doing and what is actually happening on the ground. It takes a different form, usually a more expensive one, and it surfaces at a time and in a way that is harder to respond to than the original account would have been.

My opinion

The organisations I have worked with are not, in the main, hiding from what their people know. They are fixated on what can be measured — so they measure it: a sliding scale, an interval survey, a free-text field that usually becomes an afterthought. What gets analysed is the number. What goes unasked is the right question, put to the right person at the right moment and understood in context. That absence is not accidental. It is a consequence of designing for what can be counted rather than what actually needs to be heard.

The room does not have to wait for it to arrive that way.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

There is a particular kind of frustration that comes from watching something fail that didn’t have to.

Not the frustration of being ignored, that’s common enough and usually survivable. Something more specific. The frustration of being able to see clearly what’s needed, of working every available channel to make that visible to the people who need to see it, and of watching the gap between what you can see and what the organisation can act on prove unbridgeable.

I’ve felt that frustration once more acutely than any other time in my career.

This is that story. Including the part where I got it wrong.

The client

The organisation issuing the bid was not a standard commercial enterprise with sustainability bolted onto its annual report as a concession to investor pressure.

It was a global leader in health, nutrition, and bioscience whose entire operating philosophy was built around a different understanding of what a business exists to do. People, planet, livelihood — not as a values statement decorating the reception area but as the architecture of every significant decision the organisation made. The UN Sustainable Development Goals were not a reporting framework for them. They were the reason the business existed.

When they issued a bid for a major technology and services partnership they were not looking for a supplier. They were looking for a partner who understood what they were actually for and who could demonstrate, concretely and measurably, how the proposed solution would advance the outcomes they existed to deliver.

I read the brief carefully. I understood what was being asked. And I understood immediately that responding to this brief the way we would respond to any other bid would not just fail to win — it would signal a fundamental misreading of who the client was.

What I built and what happened to it

I positioned the entire approach around genuine alignment with the client’s philosophy. I wrote the sustainability component in full. I designed an animation to tell the end-to-end story — not of our services and capabilities, but of the client’s own philosophy brought to life through the solution we were proposing. I argued consistently that this was not an addition to the bid. It was the bid.

The bid executives saw it differently. Not through malice, I want to be clear about that, because the easy version of this story makes them the obstacle and me the visionary and that version isn’t honest or useful.

They saw a client with a significant services requirement and a proven portfolio of capabilities that could meet it. They were experienced people who had won bids before by leading with what they knew how to deliver. That approach had worked. There was no obvious reason it shouldn’t work again.

What they hadn’t fully understood was that this client was different in a way that made the standard approach not just suboptimal but actively counterproductive. Arriving with a services pitch to a client whose identity was built around purpose-led outcomes didn’t just miss the mark. It told the client something about how we saw them.

How it ended

My role was gradually narrowed. The sustainability components were treated as a discrete workstream rather than the spine of the entire response. Once that work was completed I was stepped back from the wider bid discussions.

The bid was lost.

The feedback reflected exactly what I had been trying to prevent - a lack of genuine alignment with the client’s philosophy. The components that were remembered were the sustainability narrative and the animation.

The work that had been marginalised was the work the client had been looking for.

What I got wrong

I worked the informal channels when the situation required formal ones.

I had open discussions with the team. I guided the narrative through the process stages as best I could. I was vocal, directly and persistently, about the direction I thought we were taking.

What I didn’t do was stop at the beginning and force the formal conversation that the situation required.

I should have called an internal management meeting before the bid work started in earnest. I should have laid out precisely what the brief was asking for, what the client’s philosophy meant for how we needed to respond, and what the risks were of defaulting to the standard approach. I should have defined those risks explicitly, not as a concern or a perspective but as a formal risk to the bid outcome, and sought explicit agreement on the strategy before a single word of the response was written.

That process might not have changed the outcome. But it would have given the argument a formal status inside the organisation that informal advocacy never achieved.

Being right about the direction wasn’t enough. The system needed more than clarity. It needed process, ownership, and formal agreement. Without those, individual insight, however accurate, remains exactly that. Individual. Unowned. Unable to move the organisation.

Morgan is a secondary school teacher in their mid-forties — the kind of person who spends the working day managing other people’s problems and expects, by the time they get home, to be done with crises for a while.

That stopped being true about six months ago.

It is a Tuesday evening and Morgan is at the kitchen table again. There is a folder. It has grown over six months from a single printed email into something that requires a cardboard divider. Bank correspondence in one section. Police reference numbers in another. A third section for the accounts Morgan did not open — the mobile contract in Rotherham, the credit card with the bank they have never used, the subscription to something they cannot name. Each one requires a letter. Each letter requires evidence. Each request for evidence requires a call. Each call begins with a recorded message about high volumes and current wait times.

The original incident took an afternoon. Someone rang, someone convinced, someone transferred. The amount was not catastrophic — significant, but not life-altering. That number appeared on every police form filled in. It appeared in the bank’s assessment of the case. It is the number that gets counted in the statistics that appear in reports that describe the scale of digital fraud in this country. It does not include the folder.

The folder is where the real cost lives. Not the money that went on a Tuesday afternoon six months ago, but the forty-seven hours since — logged, because Morgan started logging them in month two, when it became clear that nobody else was. Forty-seven hours on hold, on calls, writing letters that ask institutions to confirm what happened to a person who already knows what happened to them but must prove it again to each new department in turn. There is also the thing that does not log easily — a low hum, a background awareness, persistent and exhausting, that the folder is never quite finished, that somewhere there is probably another account, that the next letter might require something not yet thought to keep. The fraud was an event. This is a condition.

The cost nobody counts

UK Finance recorded more than £1.17 billion in fraud losses across 2024 — 3.3 million separate incidents. The figure cited in press releases, parliamentary briefings, and annual reports is always the money. What those reports do not contain is the folder. What they do not count is the recovery.

Only around 14 per cent of frauds against individuals are ever reported to Action Fraud or the police. Partly this reflects a belief, often well-founded, that reporting will not lead anywhere. The Office for National Statistics found that 71 per cent of fraud victims reported being emotionally affected. The Police Foundation report Invisible Harms documented physical and psychological impacts less understood and less measured than the financial loss, and found specialist support services to be lacking and early intervention largely absent. The shame, the self-blame, and the erosion of trust that follow an incident appear in no statistic that the banks or the government routinely publish.

The true cost of digital crime is not the original loss. It is everything that comes after.

A system designed for process, not people

The systems that are supposed to help are not designed for the person in front of them. They are designed for process — for case management, for compliance, for due diligence, for the protection of the institution carrying out the review.

The fraud team calls back to confirm what the victim already knows. The police reference number opens a portal that summarises what has already been submitted. The credit agency writes to confirm that flagged accounts have been noted — which requires a written acknowledgement that the confirmation was received. Each exchange adds documentation to someone else’s record. None of it changes the situation for the person who was defrauded.

The letter templates are identical. The call scripts are identical. The processing timelines are fixed. Every victim becomes a case reference moving through queues designed for the average. The average does not spend Tuesday nights logging hours because nobody else is.

When no one can act

A significant proportion of digital fraud targeting UK victims is perpetrated from overseas. This creates a hard stop in the recovery process that victims often reach only after months of effort. They are told, eventually, that the perpetrator is outside UK jurisdiction. That nothing further can be done. That the case is closed on the institutional side.

This is the moment many victims describe as the most damaging — not the original fraud, but the formal confirmation that the system has run its course without resolution. The UK government’s fraud strategy for 2026 to 2029 acknowledges the cross-border dimension, but the gap between published strategy and what an individual victim actually encounters remains substantial.

The crime was an afternoon. The recovery is a year of your life.

My opinion

What I notice, after three decades inside the technology industry, is how consistently the conversation stops at the immediate aftermath. We see the incident — the money transferred, the case number assigned, the bank’s response issued. What is not visible is the secondary and tertiary blast radius: what comes after, which no single organisation owns. Each institution closes its part of the process. The story continues for the person at the kitchen table. Support exists, but it is fragmented and difficult to navigate — and when someone is told that nothing can be done because the perpetrator is in a foreign country, that moment carries a weight that no fraud report measures. The consequences run much further than the original event, and they rarely land on the institution.

Questions the system never asks

Fraud recovery touches more people than the statistics suggest. If any of this is familiar — from your own experience, someone close to you, or the systems you work inside — these are worth engaging with directly.

1. Will any of this be familiar — from your own experience, or someone close to you?

2. If you have been through a recovery process — for fraud, for a breach, for anything like this — what would you want someone to know about what it actually involved?

3. Should the cost of recovery fall on the victim — or on the institutions whose systems were exploited, and whose processes extend the damage?

4. What would you want the person sitting at that kitchen table on a Tuesday night, six months in, to know — that nobody told them?

The true cost of digital crime is not the original loss. It is the months of bureaucracy, cognitive weight, and institutional indifference that nobody counts.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

In July 2025, a randomised controlled trial set out to measure what AI coding tools actually do to developer productivity. The developers tested were not novices — they were experienced open-source contributors working on codebases they knew well, using the tools that have driven AI adoption at scale: Cursor Pro with Claude 3.5 and 3.7 Sonnet. Before the experiment began, they predicted the tools would make them 24% faster. After completing the tasks, they still believed AI had made them 20% faster. The measured result was that they were 19% slower than developers who had worked without AI. The gap between what the tools felt like and what they produced was 39 to 44 percentage points.

This was not a finding about novices or edge cases. The Model Evaluation and Threat Research group, which conducted the trial, ran it on real issues in real codebases — the kind of work that makes up most professional software development. The minutes saved on initial code generation were consumed by reviewing, fixing, or discarding AI output that was close to correct but not correct enough. The feeling of speed was real. What it was measuring was not productivity.

That gap — between the experience of using AI coding tools and their measurable effect — is what makes the current moment unusual. The tools have been adopted at extraordinary speed, the productivity dashboards show dramatic improvements, and the downstream costs are accumulating in places the dashboards do not look.

What the numbers actually say

The scale of AI code generation adoption is not in doubt. GitHub Copilot reached 20 million users by July 2025 and had 4.7 million paid subscribers by January 2026 — a 75% year-on-year increase. It is deployed at 90% of Fortune 100 companies. By early 2026, 84% of developers reported using AI tools in their development process. GitClear’s analysis of 211 million changed lines of code found that approximately 41% of all code committed globally in 2025 was initially generated or suggested by AI.

By April 2026, Google’s CEO Sundar Pichai confirmed that 75% of all new code at Google is AI-generated — reviewed and approved by engineers before it ships, up from just over 25% in mid-2024.

The productivity data from vendor-run studies is compelling. GitHub’s study of 4,800 developers, conducted with Accenture, found task completion 55% faster with Copilot. Pull request time dropped from 9.6 days to 2.4 days. Successful builds increased 84%. Developers using AI daily merge approximately 60% more pull requests than light users. DX, a developer analytics platform tracking more than 135,000 developers, found that developers using AI tools regularly self-report saving an average of 3.6 hours per week.

These numbers are real. They are also, largely, measuring the things AI code generation is good at: boilerplate, test scaffolding, documentation, refactoring familiar patterns, completing obvious continuations of existing code. The 55% faster figure comes from a task involving JavaScript HTTP server code — a type of work that is highly amenable to AI suggestion, with clear patterns and extensive training data. The same gains do not generalise to all development work, and the vendor-run studies are not designed to measure the downstream costs.

Experienced developers using AI tools felt 20% faster. They were 19% slower. The 39-percentage-point gap between perception and measurement is the most important data point in software development in 2025.

A separate analysis by Faros AI, examining telemetry from more than 10,000 developers across 1,255 teams, found that AI adoption correlated with a 9% increase in bugs per developer. Individual output increased; the organisation’s code quality fell.

The security picture is starker. Veracode tested more than 100 large language models against 80 coding tasks specifically designed to carry a known security dimension — the kind of tasks where a wrong choice introduces a real vulnerability. In 45% of cases, the models chose the insecure path, introducing vulnerabilities from the OWASP Top 10; newer and larger models did not perform significantly better than their predecessors. CodeRabbit’s analysis of 470 real-world pull requests found that AI-generated code introduced 2.74 times more cross-site scripting vulnerabilities than human-written code, and 1.9 times more insecure direct object references. Aikido Security’s survey of 450 organisations found that one in five had experienced a serious cybersecurity incident caused by AI-generated code in 2025.

What the technology cannot see

This is not the first time the industry has tried to automate code production. In the 1980s and 1990s, Computer Aided Software Engineering — CASE — made the same promise: feed your models and specifications into a repository, generate the code. The output was usually unusable. Teams spent longer unpicking what the tools produced than they would have spent writing from scratch. The tools were not wrong about the mechanics of generation. They were wrong about what the hard part of software engineering actually is.

An AI coding assistant faces a version of the same problem at larger scale and higher speed. It has been trained on enormous quantities of code from public repositories. It has learned to predict what code is likely to follow given what has already been written. It is, in that sense, a very sophisticated pattern-completion engine. What it does not have is any understanding of the specific system it is generating code for — the business logic, the security requirements, and the architectural decisions that are specific to this system rather than to code in general.

SQL injection is one of the oldest and most well-documented security vulnerabilities in web development. It appears in the training data with high frequency, both as the vulnerability and as its correct prevention through parameterised queries. AI models should, in principle, know how to avoid it. They often do not. The reason is that the model is optimised to produce code that completes the task as stated — and the task as stated rarely includes ‘and also, make absolutely sure no user input can be injected into this query’. The model takes the shortest path to a passing result. Security requirements that are implied rather than explicit tend to be skipped.

GitClear’s longitudinal analysis of 211 million lines of code identified a more structural problem. The proportion of code that constitutes refactoring — the disciplined work of improving existing code’s structure without changing its behaviour — fell from 25% of all commits to under 10% as AI adoption increased. Simultaneously, code duplication increased fourfold. The pattern is consistent with what happens when code generation is optimised for volume and speed: new code is added rather than existing code improved, similar patterns are copy-pasted rather than abstracted, and the codebase becomes progressively harder to understand and change.

The term that circulated widely in 2025 was ‘vibe coding’, coined by Andrej Karpathy to describe the practice of accepting AI suggestions without reading the generated code carefully — programming by feel, copying error messages back to the AI until it works. As a personal approach to low-stakes projects, this is a reasonable efficiency trade. As a pattern spreading through professional development teams, it has measurable consequences. OX Security examined more than 300 repositories and identified ten recurring anti-patterns present in 70 to 100% of AI-generated code, including incomplete error handling, weak concurrency management, and inconsistent architectural decisions. These are not exotic failure modes — they are the standard failure modes of code written by someone who knows what the output should look like but does not fully understand the system it runs in.

The junior developer question is also changing in ways that are not straightforward. Demand for very junior developers — those whose primary value is writing boilerplate and implementing well-specified features — has softened as AI handles more of that work. Gartner has forecast that 80% of software engineers will need to upskill in AI by 2027. The role that remains, and that is if anything growing in value, is the developer who can specify problems precisely enough for AI to address them, review AI output critically enough to catch its characteristic failure modes, and design systems architecturally rather than producing code line by line. That is a different job description from the one most developers were hired to do five years ago.

Who benefits, and the costs they cannot see

The clearest beneficiaries of AI code generation are not the developers who were already most productive. They are the people who previously could not build software at all. In Y Combinator’s Winter 2025 batch, 25% of startups had codebases that were 95% AI-generated, with small teams — often fewer than ten people — shipping products with paying customers at a pace that would have been impossible two years earlier. A non-developer who can specify what they need in plain language, and who has enough domain expertise to evaluate whether the output works, can now produce functional software at a speed that was impossible without technical co-founders. The barrier that used to require a technical co-founder or a development team has dropped to domain expertise and the ability to specify what you need.

The costs tend to fall elsewhere. Production systems with AI-generated code are accumulating a backlog of maintainability problems, duplicated logic, and architectural inconsistencies that compound over time — what the industry has taken to calling a technical debt tsunami. Analysis by Codebridge describes a predictable pattern in AI-assisted projects without proper governance — what it calls the 18-month wall: velocity gains in months one to three are followed by a plateau, then a decline, and by 18 months the team can no longer fully understand their own system. The code shipped faster. The code that was shipped has become progressively harder to change.

The clearest beneficiaries of AI code generation are the people who previously could not build software at all. The costs — security vulnerabilities, technical debt, degraded maintainability — tend to fall on the teams who inherit the codebase.

The security consequences are not evenly distributed either. An organisation with a mature security engineering function — automated scanning in the continuous integration pipeline, mandatory human review for AI-generated code, developers trained to recognise the characteristic vulnerability patterns that AI produces — can capture the productivity benefit while managing the risk. An organisation without those controls, which describes many of the smaller and less well-resourced companies where AI code generation is growing fastest, is generating code with known vulnerability patterns at scale. The one-in-five cybersecurity incident rate from Aikido Security’s survey is not a future risk. It is a 2025 reality.

What ‘more productive’ actually means

There is a genuine productivity case for AI code generation. For handling tedious scaffolding, repetitive testing, and documentation that would otherwise be skipped, the tools do what they claim — they reduce the time spent on work that does not require judgement, and free the developer to focus on what requires care. The 55% faster result in GitHub’s study is real. The question is whether the developer who is 55% faster at writing the initial code is also better, faster, or more thoughtful at reviewing it.

Whether AI code generation expands what developers can do — building more complex systems, serving more users — or whether it primarily produces more code without proportional improvement in outcomes is not the same question. The technical debt and security findings suggest that ‘more code’ and ‘better software’ are not the same thing. An organisation that measures developer productivity by lines of code or pull requests merged will see dramatic improvements. An organisation that measures by system reliability, security record, and long-term maintainability will see a more complicated picture.

The question that most organisations are not yet asking is whether their approach to AI code generation matches the actual risk. The risk profile is not uniform. A SQL query generated by AI in a financial application has a different security consequence from a CSS class generated by AI in a marketing site. A snippet of authentication logic reviewed by a senior security engineer has a different risk profile from the same snippet reviewed by a junior developer under time pressure.

The organisations managing this well tend to share two things: automated scanning on AI-generated code before it reaches production, and a differentiated approach — more scrutiny where the risk is high, less where it is low. Most apply uniform policies, either permitting AI use broadly or restricting it broadly. Neither matches the actual risk profile.

What the productivity dashboard doesn’t show

1. The METR trial found experienced developers were 19% slower with AI tools while believing they were 20% faster. If you use AI coding tools, how confident are you that your sense of your own productivity is accurate? What would you need to measure to find out?

2. One in five organisations reported a security incident caused by AI-generated code in 2025. Does your organisation scan AI-generated code for security vulnerabilities before it reaches production? If not, what is the process that is supposed to prevent those vulnerabilities from being deployed?

3. AI tools handle the initial code generation faster, but the work of specifying problems precisely, reviewing output critically, and designing systems that remain comprehensible as they grow — these are becoming more important, not less. Is the way your organisation hires, trains, and develops software engineers keeping up with that shift?

4. Startups in Y Combinator’s 2025 batch were building products with 95% AI-generated codebases and reaching real revenue. What does it mean for the software industry that the barrier to building has dropped this dramatically — and for whom has it actually dropped?

The tools are installed. The feeling of acceleration has not gone away. What is accumulating alongside it — the vulnerabilities, the duplicated logic, the refactoring debt — does not appear in the metrics being used to justify the adoption.

Sources

1. METR — Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity
Joel Becker, Nate Rush, Elizabeth Barnes, David Rein — Model Evaluation and Threat Research, 10 July 2025
metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study
The randomised controlled trial underlying the 19% slower / 20% faster perception gap. 16 experienced open-source developers, real issues from their own codebases, Cursor Pro with Claude 3.5 and 3.7 Sonnet.

2. Veracode — 2025 GenAI Code Security Report
Veracode, July 2025
veracode.com/resources/analyst-reports/2025-genai-code-security-report
100+ large language models tested against 80 coding tasks specifically designed to carry a known security dimension. 45% of cases: model chose the insecure path. Rate consistent across model generations.

3. CodeRabbit — State of AI vs Human Code Generation Report
CodeRabbit, December 2025
coderabbit.ai/whitepapers/state-of-AI-vs-human-code-generation-report
470 real-world pull requests (320 AI-co-authored, 150 human-only). AI code: 2.74× more cross-site scripting vulnerabilities, 1.91× more insecure direct object references, 1.7× more total issues.

4. Aikido Security — 2026 State of AI in Security & Development
Aikido Security, 2025
aikido.dev/state-of-ai-security-development-2026
Survey of 450 CISOs, security leaders, developers, and AppSec engineers across Europe and the US. 20% reported a serious security incident caused by AI-generated code.

5. GitClear — AI Copilot Code Quality: 2025 Data Suggests 4x Growth in Code Clones
GitClear, 2025
gitclear.com/ai_assistant_code_quality_2025_research
211 million changed lines of code analysed (2020–2024). Code duplication increased fourfold. Refactoring fell from approximately 25% of commits to under 10% as AI adoption increased.

6. Faros AI — The AI Productivity Paradox
Faros AI, 2025
faros.ai/ai-productivity-paradox
Telemetry from more than 10,000 developers across 1,255 teams. AI adoption correlated with a 9% increase in bugs per developer. Daily AI users merge approximately 60% more pull requests than light users.

7. DX — AI-Assisted Engineering: Q4 Impact Report 2025
DX (Developer Experience), 2025
getdx.com/blog/ai-assisted-engineering-q4-impact-report-2025
435 companies, 135,000+ developers, July–October 2025. Developers using AI tools regularly self-report saving an average of 3.6 hours per week. Note: self-reported perception, not independently measured.

8. GitHub / Accenture — GitHub Copilot Impact Study
GitHub and Accenture, 2024
github.blog/news-insights/research/survey-ai-wave-grows
4,800 developers. Task completion 55% faster with Copilot. Pull request time: 9.6 days → 2.4 days. Successful builds increased 84%. Note: benchmark task involving JavaScript HTTP server code — a domain highly amenable to AI suggestion.

9. GitHub Copilot user and subscriber figures
GitHub announcement, July 2025; Microsoft FY26 Q2 earnings call, 28 January 2026
dataconomy.com/2025/07/31/github-copilot-now-has-over-20-million-users
20 million total users by July 2025. 4.7 million paid subscribers by January 2026 — 75% year-on-year growth. Deployed at 90% of Fortune 100 companies.

10. Google — Sundar Pichai at Google Cloud Next 2026
Google, 22 April 2026
blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/cloud-next-2026-sundar-pichai
75% of all new code at Google is AI-generated, reviewed and approved by engineers before it ships. Up from just over 25% in mid-2024.

11. Gartner — Generative AI Will Require 80% of Engineering Workforce to Upskill Through 2027
Gartner, 3 October 2024
gartner.com/en/newsroom/press-releases/2024-10-03-gartner-says-generative-AI-will-require-80-percent-of-engineering-workforce-to-upskill-through-2027

12. OX Security — State of AI-Generated Code Report
OX Security, October 2025
prnewswire.com/news-releases/ox-report-ai-generated-code-violates-engineering-best-practices-302592642.html
300+ open-source repositories analysed. Ten recurring anti-patterns present in 70–100% of AI-generated code, including incomplete error handling, weak concurrency management, and inconsistent architectural decisions.

13. Codebridge — The Hidden Costs of AI-Generated Software: Why It Works Isn’t Enough
Codebridge
codebridge.tech/articles/the-hidden-costs-of-ai-generated-software-why-it-works-isnt-enough
Analysis of the 18-month wall pattern in AI-assisted projects without proper governance: velocity gains in months one to three, followed by plateau and decline.

14. Garry Tan — Y Combinator W25 batch AI-generated code
Garry Tan, X, 4 March 2025
x.com/garrytan/status/1897303270311489931
*25% of YC Winter 2025 batch startups had codebases that were 95% LLM-generated.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

In a field hospital in an active conflict zone, a patient arrives with a severe bacterial infection. The hospital has the diagnostic equipment to identify the pathogen. What it does not have is the specific therapeutic protein — the monoclonal antibody, the recombinant enzyme, the vaccine dose — that the patient needs. The nearest manufacturing facility that could produce it is a continent away, behind a cold chain that has been disrupted, in a centralised facility that requires months of lead time to retool for a new product.

This is not an edge case. It describes, in concentrated form, the structural problem of biological medicine manufacturing: the gap between knowing what a patient needs and being able to produce it where the patient is. Most biological medicines are made by growing engineered living cells in large fermentation tanks, harvesting and purifying the protein they produce, and shipping it through refrigerated supply chains to the places where it is needed. The system works well in wealthy countries with stable logistics and well-capitalised health systems. It works considerably less well everywhere else.

Cell-free biomanufacturing is a different approach, and it is further along than most people know.

Biology outside the cell

Every living cell is a protein-making machine. The genetic instructions encoded in DNA are transcribed into messenger RNA, which is then translated by ribosomes into proteins. Cell-free protein synthesis, an active research area since at least the 1960s, captures that production machinery and runs it outside any living cell. The process begins by breaking open a batch of cells, typically engineered bacteria such as Escherichia coli, though yeast, plant, and mammalian cell extracts are also used, and extracting the ribosomes and associated enzymatic machinery. Add the genetic instructions for the protein you want; the machinery reads them and produces it. No fermentation tank. No living cell line. No months of retooling to switch to a new product.

The practical advantages over conventional biomanufacturing are specific. Because there is no living cell to maintain, proteins that would kill a conventional host cell line before producing useful quantities can be made cell-free without this constraint. Process development is faster: switching from producing one protein to another requires changing only the DNA or mRNA template, not engineering a new cell line and running the months of validation work that goes with it.

The property with the most direct implications for global health is lyophilisation. Cell-free reactions can be freeze-dried into a stable powder that requires no refrigeration. Researchers at the Jewett Laboratory, now at Stanford, have demonstrated conjugate vaccines produced this way at approximately $0.50 per dose, stable at 37 degrees Celsius — ambient temperature in much of the world without reliable cold chains — for up to four weeks. Separate work from the same group, funded by DARPA and published in Biotechnology and Bioengineering in 2025, demonstrated scalable cell-free production of T7 RNA polymerase — a critical enzyme in mRNA vaccine synthesis — achieving over 90% purity at one-litre scale.

Commercial applications are beginning to emerge. Sutro Biopharma uses a proprietary cell-free platform to produce antibody-drug conjugates and bispecific antibodies, with candidates in clinical trials. Ipsen Biopharm uses cell-free synthesis specifically because it minimises containment risk when producing highly toxic botulinum toxin — a case where the open system’s safety advantages are directly relevant to manufacturing practice. GreenLight Biosciences applied cell-free systems to RNA vaccine production before its 2023 acquisition, after which the platform was redirected to RNA-based agricultural biocontrol — a reminder that commercial priorities and public health priorities do not always point in the same direction.

What the technology cannot yet reach

Technology Readiness Level 5 is an honest assessment of where most cell-free biomanufacturing sits in 2026. The technology has been validated in laboratory settings and small-scale demonstrations. It has not yet been validated in the large-scale, regulated manufacturing environments that pharmaceutical products require before they can be administered to patients.

The most significant technical constraint is yield. Cell-free systems typically produce lower quantities of protein per unit volume than conventional cell-based fermentation at commercial scale. For high-value, low-volume applications — speciality biologics, antibody-drug conjugates, research reagents — this is manageable. For mass vaccine manufacturing at the scale of hundreds of millions of doses, the yield gap remains a real cost constraint.

Post-translational modifications present a related challenge. Many therapeutic proteins require specific chemical modifications after assembly — glycosylation being the most important, where sugar molecules are attached at specific positions and affect the protein’s stability, activity, and immunogenicity. Recent work from the Jewett group, published in ACS Synthetic Biology in 2025, achieved over 85% glycosylation efficiency and yields of up to 450mg per litre of glycoprotein in a two-step cell-free platform. Progress, but still an active engineering challenge rather than a solved problem.

Good Manufacturing Practice validation, the regulatory standard required before medicines can be produced for clinical use, has not yet been established for cell-free biomanufacturing systems. The mRNA vaccine precedent is informative in both directions: a sufficiently urgent need can accelerate regulatory development dramatically, and absent that urgency, the pathway tends to lag the science by years.

The $0.50 per dose figure for lyophilised conjugate vaccines is also derived from raw material costs at laboratory scale — a proof of concept, not a commercial cost estimate. At manufacturing scale, enzyme costs, lysate production, energy, and quality control all need to be incorporated. The expectation is that cell-free systems will be cost-competitive for certain product classes, particularly those requiring rapid product switching, but this has not yet been demonstrated in practice.

Who gets to make medicine

The current geography of biological medicine manufacturing is highly concentrated. The large fermentation-based facilities required to produce biologics at scale are predominantly located in the United States, Europe, and a small number of other high-income countries. The capital cost, typically in the range of hundreds of millions to over a billion dollars, and the technical expertise required to operate such a facility means that most low- and middle-income countries have no domestic biomanufacturing capacity for the medicines their populations need most.

The COVID-19 pandemic made this structural dependency visible in its most acute form. When mRNA vaccine production was concentrated in a handful of facilities, the question of who received vaccines first was not determined by medical need. It was determined by geography, procurement power, and the location of manufacturing capacity. Countries without domestic production capability were last in line, and in some cases, doses pledged through COVAX arrived too late to prevent the deaths they were intended to prevent.

Cell-free biomanufacturing addresses this problem directly. A cell-free system producing lyophilised vaccine components requires no fermentation infrastructure, no refrigeration during storage or transport, and significantly less technical overhead to operate than a conventional bioreactor facility. The direction the research points toward is distributed, decentralised manufacturing: regional production facilities, field hospital capabilities, in principle point-of-care production in the settings where patients need treatment.

The access implications extend beyond pandemic preparedness. Many neglected tropical diseases — conditions that cause enormous suffering in low-income countries but attract little pharmaceutical development investment because the affected populations cannot pay prices that support conventional manufacturing economics — could in principle be addressed by cell-free produced therapeutics if the manufacturing cost barrier were reduced. A vaccine against enterotoxigenic E. coli, one of the leading causes of diarrhoeal disease and child mortality in the developing world, was among the first demonstrated applications of low-cost lyophilised cell-free vaccine production. The connection between manufacturing economics and which diseases receive medicines is direct.

The test that matters most is whether the technology is deployed where the access gap is sharpest, rather than where the regulatory environment is most familiar. A reduction in the capital and manufacturing threshold changes little if the pathway to GMP certification, WHO prequalification, and COVAX procurement simply replicates the infrastructure requirements of conventional manufacturing. Regulatory agencies in low-income countries, COVAX procurement frameworks, and WHO prequalification processes will all need to develop explicit pathways for cell-free produced biologics. The geographic promise of the technology does not translate into geographic reality without that work.

The field hospital in the conflict zone, the outbreak in the region without cold chain infrastructure, the neglected disease in the country without manufacturing capacity — these are not hypothetical future scenarios. They are the current state of biological medicine access for a substantial portion of the world’s population.

Cell-free biomanufacturing does not yet solve those problems. It is an early-stage technology with real constraints and a significant gap between laboratory demonstration and clinical deployment. It is, however, the first credible technical approach to a problem that has always been treated as structural and immovable: that making biological medicines requires the kind of industrial infrastructure that is simply not available in most of the world.

The geography of manufacturing is the geography of access. The question the science cannot answer on its own is who will fund the work to change it.

My opinion

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Morgan had been thorough. That was the thing that made it so difficult to explain when the question came back six months later.

The assessment had been careful — a credit risk evaluation for a commercial loan application, supported by an AI-assisted analysis platform the team had been using for the better part of a year. The tool had been performing well. The outputs were coherent, explainable, defensible. The assessment reached a conclusion, the decision was recorded, the file was closed.

Six months later, a query arrived from legal. Could the team reproduce the original assessment? Run the same inputs through the same system and show that the same logic applied?

The query was run. The output was different.

Not marginally different. Not a rounding variation. Different in a way that, had it been the original output, might have changed the conclusion. Different in a way that could not be explained without knowing what had changed in the model between then and now — and that information was not available, because nobody had recorded it, because nobody had been asked to, because when the system went live, nobody had thought to ask.

The problem with the update

AI systems are not static. The models underlying them are updated — sometimes by the vendor, automatically, as part of routine improvement cycles; sometimes in response to identified bias or error; sometimes because the training data has been refreshed. This is, in principle, good. A model that learns and corrects over time is more valuable than one frozen at the point of deployment.

The difficulty is that improvement and consistency are in direct tension. A model producing better outputs today than it did six months ago is, by definition, a model that would produce different outputs given the same inputs from six months ago. That difference is the point of the improvement. In any context where a decision needs to be reconstructed or defended, it is also an unanswered question about what the original decision was actually based on.

Most organisations deploying AI-assisted tools have not resolved this tension. They have ignored it. The model updates. The outputs change. Nobody is notified.

Nobody records the version in use when a specific decision is made. The audit trail — if one exists — captures the conclusion. It does not capture the reasoning engine that produced it.

Morgan’s situation was not unusual. It was the predictable result of deploying a consumption-based AI tool without the governance to track what the tool was doing over time. The barn door had been left open. The question was not whether the horse had bolted. The question was which horse, when, and whether anyone had kept a record of the herd.

The infrastructure that was not built

For an AI-assisted decision to be defensible over time, three things need to be true.

1. The version of the model in use at any given point needs to be recorded.

2. The inputs to any significant decision need to be preserved in a form that can be rerun.

3. The outputs need to be traceable back to the specific model state that produced them.

None of these are technically difficult. Version control is a solved problem. Input logging is standard practice in well-governed systems. Audit trails linking outputs to model versions are routine in regulated software environments. The infrastructure exists.

The decision not to apply it to AI deployment is not a technical gap. It is a governance gap — the same gap that appears in every wave of technology adoption where deployment has consistently moved faster than the thinking that should accompany it.

In financial services, decisions about lending, credit, and investment carry regulatory obligations around explainability and reproducibility. In healthcare, an AI-assisted diagnostic or triage decision needs to be reconstructable if a patient outcome is later questioned. In law, the reasoning underlying a recommendation used in proceedings may face scrutiny that requires reconstruction. The sector changes. The underlying failure is the same.

What makes this particularly hard to catch is that model drift is invisible by design. The system continues to function. Outputs continue to arrive. Nothing signals that the model today is not the model that was running when the previous decision was made.

The degradation of the audit trail happens silently, with every update cycle. By the time someone needs to reconstruct a decision, the model that produced it may no longer exist in any recoverable form.

Organisations that deployed AI-assisted tools before any framework for tracking model versions was in place are already in this position: the original model gone, the update history undocumented, past decisions unreproducible. This is not a future risk. It is the present condition for most organisations that moved quickly to deploy.

When the query comes

The most immediate consequence is legal exposure. In financial services, the FCA’s Consumer Duty requires that AI-assisted decisions be explainable and that firms maintain records sufficient to reconstruct the decision-making — not just the conclusion but the model state active at the time. The Data (Use and Access) Act 2025, which received Royal Assent in June 2025, adds further safeguards around automated decision-making, including the right to information about the reasoning used and the ability to contest decisions. A firm that cannot reproduce the original output cannot satisfy either obligation. The decision may have been correct. Without reproducibility, that cannot be demonstrated.

In healthcare, the consequences sit closer to the individual. An AI-assisted triage or diagnostic support tool that has been updated between a clinical decision and a subsequent review creates a gap in the clinical record. The original output cannot be reproduced. The basis for the original decision becomes opaque in retrospect. That opacity matters most precisely when the outcome was not what was hoped for.

In law, the problem is evidentiary. Any AI-assisted assessment used as part of a legal process — risk scoring, document analysis, due diligence support — may be subject to scrutiny requiring the reasoning to be reconstructed. A system whose model has changed since the original output was produced cannot provide that reconstruction. The question of what the AI actually considered, and on what basis, becomes unanswerable.

Beneath all of these is a more general corrosion. Organisations deploy AI tools to improve consistency — more reliable, more defensible, more uniform outputs than human judgement alone. Model drift undermines that promise at its foundation. The consistency was real at the point of deployment. Whether it persisted through subsequent updates is unknown, because nobody was watching. Every benefit of the deployment becomes contingent on the governance infrastructure most organisations have not built.

Closing the door

The work required to manage model drift has to be done before the problem appears, not after. Once a decision has been made, the model has been updated, and the version history has not been kept, the gap cannot be closed retrospectively. The barn door cannot be rehung from inside the field.

Version pinning is the starting point: every AI-assisted decision of significance should record the model version, the input data, and the output at the point it is made. This is not different in principle from the version control already applied to regulated software in financial services, healthcare, and legal environments. The principle is established. The application to AI tools has not been made.

Vendor contracts are the second lever. Most organisations deploying AI platforms under SaaS or consumption-based models have no contractual requirement for model update notification, version documentation, or backward reproducibility. This is a procurement failure as much as a governance one. The capability to reproduce historical outputs may exist technically with the vendor. Without a contractual obligation to preserve and provide it, that capability is at the vendor’s discretion.

The AI Business Office — the governing function responsible for AI consumption and risk across the enterprise — is where both of these responsibilities sit. Not in compliance, not in IT. In a function with genuine authority to set standards and enforce them — not merely to advise. That function is absent in most organisations deploying AI today. Its absence is the systemic condition that makes Morgan’s situation not exceptional but typical.

The deployment decision is the easy part. The governance that makes it defensible over time is slower, less visible, and less likely to feature in the business case. It does not appear in a vendor demonstration. It is the work done — or not done — in the months after the system goes live, once the attention has moved on and the question of what comes next has not yet been asked.

By the time it is asked, in many cases, the answer is already too late for the decisions already made.

The problem predates the AI

Model drift is not new. Statistical and machine learning models have drifted as long as they have been deployed. Any model trained on data and used over time will drift as conditions change, as data is refreshed, as the model is retrained — actuarial models, credit scoring systems, fraud detection engines. The problem is older than the current generation of AI tools.

What AI has changed is the speed, the scale, and the opacity of what changed. Traditional software updates come with version numbers, change logs, release notes. AI model updates change patterns in parameters, not lines of code — what changed is often not just undocumented but undocumentable in the same form. The update cycle is faster. The number of decisions affected is larger. The answer to “what exactly is different, and in which direction?” is harder to state even when someone thinks to ask.

This creates a provenance problem that existing governance frameworks were not built to handle. The question is not only whether the model drifted. It is whether anyone can say what it drifted from, when, and with what consequence for the decisions made in between. Most organisations cannot answer that. Not because they are careless. Because the infrastructure to track it was never put in place.

Organisations moving into AI-assisted decision-making in regulated contexts are taking on an obligation that most have not costed and many have not identified. The question is not whether model drift will affect their decisions. For most organisations that have been running AI-assisted tools for more than a year, it already has. The question is whether they will know.

What the question from legal would reveal

• If your organisation has deployed an AI tool supporting significant decisions — in lending, clinical assessment, legal analysis, or any regulated context — can you reproduce the output of a decision made twelve months ago? Not approximately. Exactly, from the same inputs, using the same model.

• If your AI platform vendor released an update in the last six months, were you told? Is the update version recorded anywhere against the decisions made before and after it? Does anyone in your organisation own that responsibility?

• When you consider the decisions your organisation has made with AI assistance over the last year, which would be most difficult to defend if questioned today? What would you need to reconstruct the reasoning — and is that information still available?

• The work that makes AI deployment defensible over time costs money and requires deliberate decisions to build. The absence of it also costs money — in legal exposure, regulatory risk, and decisions that cannot be explained. Which investment is your organisation actually making?

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Satellite internet from low Earth orbit is no longer experimental. Starlink has roughly 9,000 satellites in orbit, operating close enough to the surface to keep latency low enough for video calls, cloud software, and large file transfers — in places that will never see a fibre cable. The coverage footprint reaches an estimated 3.1 billion people. As of December 2025, the subscriber base was nine million.

Those two numbers describe the same technology. They are very different things.

Coverage is the geographic area within which the signal can, in principle, be received. Subscribers are the people who can afford to use it, who have the hardware, and who live in markets the provider has chosen to serve. The engineering problem is largely solved. What hasn’t been solved, and what the headline numbers tend to obscure, is the price, the politics, and the question of who gets to decide.

The constellation that already exists

SpaceX’s Starlink is the dominant reality of satellite internet in 2026. Around 9,000 Starlink satellites are in low Earth orbit operating between 340 and 1,200 kilometres above the surface, low enough to keep signal latency to 20–50 milliseconds rather than the 600-millisecond round trips of the old geostationary satellites that made satellite internet frustratingly slow throughout the era of geostationary connectivity. As of December 2025, Starlink had roughly nine million subscribers across more than 155 countries, with coverage reaching an estimated 3.1 billion people. Residential plans in most markets cost around $110 per month, plus an upfront hardware cost of $350–$600 for the dish and router. Speeds typically range from 50 to 220 Mbps under good conditions.

The nearest competitor in low Earth orbit is Eutelsat OneWeb, formed from the merger of the European satellite operator Eutelsat and the British-Indian OneWeb constellation. OneWeb operates 648 satellites, primarily serving enterprise, government and maritime customers, with particular strength in polar regions. It is not a consumer product in the way Starlink is. Amazon Leo, rebranded from Project Kuiper in November 2025, is the most significant new entrant. With approximately 300 production satellites in orbit as of April 2026 and an FCC-authorised constellation of 3,236 planned, Amazon began an enterprise preview in late 2025 and is rolling out service more broadly in 2026 as satellite density increases. Its integration with Amazon Web Services gives it something Starlink lacks: a natural fit with the infrastructure decisions that large organisations have already made. Telesat Lightspeed, Canada’s planned LEO constellation, is targeting 2027.

A separate development sits alongside the LEO constellations: AST SpaceMobile, which is building satellites that connect directly to standard mobile phones, no specialised dish required, via partnerships with AT&T, Verizon and other carriers. The first generation of service supports messaging and basic data. More capable versions are planned. If AST SpaceMobile delivers on its roadmap, the access barrier shifts entirely: connectivity from orbit available to anyone with an ordinary mobile handset.

Taken together, the technology represents a real shift. Low Earth orbit satellite internet provides broadband speeds with latency approaching that of ground-based cable, to places that will never see a fibre rollout. In Kenya, where Starlink launched in 2023, analysis by Rest of World found that the monthly subscription cost is lower than that of the leading fixed-line providers. In five African countries — Kenya, Ghana, Zimbabwe, Mozambique and Cape Verde — Starlink is cheaper than the cheapest available broadband alternative. In Indonesia, nearly 60% of Starlink’s subscriber base is in rural areas, nearly four times the proportion seen among fixed wireless or fixed-line users, reflecting how the technology has positioned itself in markets where ground-based infrastructure is absent, according to an Opensignal analysis published in November 2025.

The three gaps the headline numbers hide

Stating that Starlink reaches 3.1 billion people is a statement about coverage, the geographic area within which the signal can, in principle, be received. It says nothing about who can afford to use it. The two numbers are very different.

In Nigeria, as of late 2025, the standard Starlink residential subscription costs the equivalent of $39 per month. The national minimum wage is approximately $48 per month. The monthly subscription alone therefore consumes more than 80% of minimum wage earnings — before accounting for the hardware cost of the dish, which runs to around $406. In an analysis published by TechCabal, the conclusion was direct: Starlink in Nigeria is a service for high earners. It addresses a real connectivity gap and is faster than most alternatives, but it is not a vehicle of mass digital inclusion at current pricing. The people it most visibly fails to serve are precisely the ones whose connection to the global economy would benefit most from affordable internet.

The community hub model is a partial answer. A single Starlink terminal shared across a school, a health clinic, or a rural community centre provides access at lower per-user cost than individual subscriptions. Governments in Mexico, parts of Sub-Saharan Africa and Pacific Island nations have subsidised terminal deployments for schools and public institutions. These approaches work within the technology’s constraints, but they provide intermittent shared access rather than the persistent personal connectivity that changes what an individual can do - hold a remote job, study independently, run a small business.

In Nigeria, the standard Starlink monthly subscription consumes more than 80% of the national minimum wage before hardware. The technology covers the ground. The economics do not reach the people who most need it.

The second gap is geopolitical. Starlink is owned and operated by SpaceX, which is owned and personally controlled by Elon Musk. The service is now active in 160 markets, operates across every active military theatre where Western forces are present, and serves as critical communications infrastructure for Ukraine’s armed forces. In 2022, Musk declined to extend Starlink coverage to support a Ukrainian naval drone operation near Crimea, citing his personal assessment of escalation risk. In early 2025, US negotiators were reported to have threatened restrictions on Ukraine’s Starlink access during negotiations over a minerals deal — a threat the US government did not officially confirm, and that Musk denied SpaceX would act on, but one whose credibility was sufficient to generate significant anxiety in Kyiv. When a private company, and effectively a single individual, can decide whether a frontline military has communications access, the relationship has moved beyond commercial. It has become a question of which private actor holds sovereign functions that states have not figured out how to reclaim.

Countries that have noticed this dynamic are responding in predictable ways. China has authorised its own domestic LEO constellation. The European Union is accelerating IRIS², its planned governmental satellite constellation. India is developing its own LEO system. These are expensive and slow. They reflect a rational assessment that dependence on SpaceX for critical communications infrastructure carries political risk that states are not willing to accept indefinitely. For countries without the resources to build a rival system, the options are narrower.

The third consequence is further removed from any individual subscriber but harder to reverse. When no governance mechanism constrains how many objects private operators can place in orbit, the sky fills up faster than anyone has agreed to manage it. Starlink alone has launched more than 10,700 satellites, with an FCC-approved plan for nearly 17,000 second-generation satellites and a stated long-term ambition of 42,000. Amazon Leo plans 3,236. Other constellations add thousands more. The number of operational satellites in low Earth orbit has roughly tripled in five years. This creates a set of problems that have no straightforward commercial solution. Collision risk between satellites and between satellites and other space objects increases with orbital density. Light pollution from reflective satellite trails is now a documented problem for ground-based astronomy. The Kessler syndrome — a cascade of collisions generating debris that makes entire orbital shells unusable — is a theoretical endpoint that gets less theoretical as orbital density increases. Russia has reportedly been developing a weapon designed to flood Starlink’s orbital shells with clouds of high-density pellets. Analysts and intelligence assessments from two NATO nations have warned that such an attack could generate debris cascading out of control, threatening not only Starlink but every constellation in low Earth orbit, including those operated by Russia and China.

What connectivity from orbit actually changes

The changes that satellite internet delivers when access is genuinely affordable are concrete. In rural communities with reliable satellite connectivity, telemedicine consultations become possible, a GP or specialist in a city can see a patient in a remote village. Remote education stops being a compromise of intermittent video and becomes a sustained engagement. Agricultural information services that require real-time connectivity — weather alerts, market prices, crop disease identification through image recognition tools — reach the farmers who most need them. Disaster response changes when communications infrastructure can be re-established in hours via satellite rather than weeks via cable repair.

These are not hypothetical. They are documented in communities across Sub-Saharan Africa, Latin America, Southeast Asia and the Pacific where satellite connectivity has arrived and where the access barrier was cleared, often through subsidised hardware, government procurement, or NGO deployment. The lesson from these communities is not that the technology solves the connectivity problem. It is that the technology, when its economics are addressed, removes one of several barriers that have kept communities disconnected. The remaining barriers power access, device access, digital literacy, content in local languages, do not disappear because the signal arrives from orbit.

The consequences of the geopolitical concentration are slower to manifest but potentially more durable. A world in which critical communications infrastructure for developing nations runs through a constellation controlled by a single billionaire’s company is a world in which the decisions that individual makes about pricing, coverage, and access carry consequences that previously would have required state-level action. SpaceX can raise prices, withdraw from markets, or restrict usage in ways that no regulatory body currently has clear authority to prevent. The International Telecommunication Union governs spectrum allocation. It does not govern service continuity for nations that have come to depend on a private provider for essential communications.

The gap between coverage and access

Satellite internet works best for people when they can actually afford to use it. Where the economics work, where the subscription is affordable and the hardware accessible, the technology removes a real barrier that ground infrastructure failed to remove. A household with a working subscription has gained persistent connectivity: enough for remote work, independent study, a small business operated from somewhere that previously had none of those options. A school with a shared terminal has partial help. A farming community that cannot clear the hardware cost has nothing. The distinction matters because it determines whether the technology fulfils the claim made for it or whether it creates a new tier of the digital divide: connected enough to know what you are missing, not connected enough to benefit from it.

Whether satellite internet adds something that terrestrial networks cannot is answered most clearly at the coverage edge. In places where fibre will not arrive within this generation’s planning horizon, LEO satellite is not a supplement to terrestrial infrastructure. It is the infrastructure. The improvement is real and significant for those communities. For communities where ground-based options exist, satellite is generally a backup or a premium option, and the case is weaker.

The hardest question is not about the technology. The technology does not adapt to the economic circumstances of the people it claims to serve. It adapts to the willingness to pay of the people who can afford it. The pricing structure that works in Germany or Australia does not work in Nigeria or Papua New Guinea. SpaceX has shown some willingness to adjust regional pricing, plans in some African markets run as low as $10 per month for a capped data allowance, but the hardware cost remains a significant barrier even at reduced subscription rates. An adaptive approach would involve structured subsidies, national procurement at scale, or hardware financing models that distribute the upfront cost over time. Some governments are pursuing these. Most are not.

The engineering is largely solved. What isn’t solved is the price, the politics, and the question of who gets to decide.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

The room goes quiet. Not uncomfortable quiet, a different quality entirely. The kind that happens when a group of capable professionals simultaneously understands that it does not have an answer to a question it should have been asking all along.

The question came in the middle of a design review that had been going well. The wireframes were done. The architecture was settled. The team had been professional throughout and the design was technically sound. Someone near the back of the room looked up from the screen and said four words that were not on the agenda: who is this hardest for?

Nobody answered immediately. Alex looked at the wireframes. The design had a user in it, of course it did, but that user was implicit, unexamined, averaged into the decisions without ever being named. The silence in the room was the collective recognition that the person who would find the design hardest to use was probably not the person the design had been built around.

What the room had been reviewing

Before the question arrived, the review had been evaluating what design reviews are structured to evaluate. Does it meet the technical requirements? Is it consistent with the component library? Does it follow the agreed interaction patterns? Is it on schedule? All of it legitimate. None of it the wrong question. And none of it asking who the design would fail.

A review process evaluates what it is structured to evaluate. The structure of this review did not include a question about the hardest user. And so that question had not been asked — not through negligence, not through indifference, but because the agenda did not have a line for it.

The design had an implicit user at its centre: someone who could read the instructions without difficulty, navigate the steps without support, understand the error messages without translation, and complete the verification without a second attempt. That user was not wrong to design for. They were just not the only person who would have to use the product.

The review had been thorough and the design was ready. The question changed what ready meant.

The structural fix is a single line

The absence the silence named is not expensive to address. It does not require a new team, a revised budget, or a different process. It requires one question added to the structure — to the design brief, to the review agenda, to the definition of done — in a form that is not optional and not supplementary.

Who is this hardest for, and what have we done about it?

Asked early and written into the process, that question is the intervention. Not at the end of the project, when the architecture is fixed and the cost of change is high. At the beginning, when a question in a brief costs nothing to answer and the answer shapes what is built. In the middle, when a question in a review can still redirect a flow, a content decision, an error state, a test plan. Before release, as a criterion that must be met before the product ships.

The reason the question is not already standard is not that it is expensive. It is that the processes governing design were built to evaluate what their authors cared about, and the hardest user was not in the room when those processes were written.

What made the question work in Alex’s review was that it arrived without defensiveness, a genuine inquiry that assumed the team capable, not a challenge to what they had produced. It was followed by something specific: not “users with lower digital literacy” as an abstract category, but a named-enough description of a real kind of person the team could treat as a design constraint. From there they knew what to do.

That specificity is what turns a question from a philosophical concern into a piece of work. Abstract categories — accessibility, inclusion, the underserved — are easy to acknowledge and difficult to act on. A specific person in specific circumstances is a brief. The question that produces a brief is the question that produces a change.

What the question produced

The first change was to the onboarding instructions. Written for someone who already understood the purpose of the system, they assumed knowledge that the hardest user would not have. Rewritten at a lower reading level, with the purpose of each step stated plainly before the step itself, they carried the same information to a wider range of people. The change took two hours. It had been available from the start of the project.

The second change was to the verification flow. Designed as a four-step process with each step branching on the user’s input, it required a correct first attempt to avoid a dead end that offered no clear route forward. The hardest user — the person the team named after the question was asked — would have reached that dead end reliably. The flow was restructured: two steps, linear, with a recovery path at every stage. Completion rates in user testing improved across all participants, not only the hardest case.

The third change was to the error messages. “Invalid input” had appeared throughout the design. The hardest user would not know from those two words what was invalid, why, or what to do about it. The messages were rewritten to name the error specifically and state the fix plainly. It took an afternoon.

The fourth change was a user test. Before the question, the test plan had included participants matched to the central profile, engaged users who knew what they were doing. After the question, a round of testing was added with participants who represented the hardest case. They revealed a fifth issue the team had not anticipated: a step considered self-explanatory that three of the four hardest-case participants read differently from how it was intended. The step was reworded before the product shipped.

None of these changes were large. Together they constituted the difference between a product that served most of its users and one that served a wider range. The only cost was the question that had not been on the agenda until someone put it there.

Designing for the outlier

The six-word question contains more than it appears to. Asking who the design is hardest for forces the team to ask something they had not been asked before: does this product genuinely help the person who needs it most, or does it add to the capability of people who were already capable without it?

The pattern is not unique to design reviews. Netflix has reported that 80 per cent of its subscribers use captions at least monthly, a feature built for people who are hard of hearing, now used daily by people watching in a second language, in rooms where the sound cannot be loud, putting children to sleep. Microsoft’s Inclusive Design programme names the principle: solve for one, extend to many. Designing for the person who finds the product hardest to use does not produce a narrower product. It produces a more useful one.

That is what the redesign of Alex’s verification flow demonstrated. A flow restructured to be navigable for the person who would have found it hardest did not improve completion rates only for that person. It improved them across every participant in the test. Designing for the outlier is the strongest test a design can face. It reveals what designing for the average conceals.

The design processes that do not have this question built in are not negligent. They evaluate what they were built to evaluate, and the people who built them made decisions about what mattered. The person who would find the product hardest to use was not part of those decisions. That has not been corrected, not because the correction is expensive, but because the people who govern design practice have not decided that it is required.

Four questions worth asking

1. Think of the last design review you participated in as a designer, a stakeholder, a commissioner, or a user tester. Was the question asked: who is this hardest for? If it was, what happened? If it was not, what do you think would have been different if it had been?

2. If you run or participate in design reviews, what would it take to add the human question as a standard agenda item not optional, not supplementary, but required? What is the obstacle, and is the obstacle proportionate to what it is protecting?

3. If you commission technology, does your brief require the supplier to answer the question: who is this hardest for, and what have you done about it? If not should it?

4. If you had one question to add permanently to your design process — not as guidance, but as a required criterion — who in your organisation would need to decide that? Have you ever been in the room where those decisions are made?

The silence in Alex’s design review did not last long. The team had a description within a few minutes grounded enough in the work to treat as a constraint. And once it was a constraint, they knew what to do with it. The question that was not on the agenda was the most productive thing the review produced.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.