AI in UK Central Government Services

Why this briefing

DWP describes its Universal Credit advances model as "the only fraud and error machine learning model currently deployed at scale in live service". Its own July 2026 assessment shows the model refers non-UK nationals at 3.35 times the rate of UK nationals — up from 2.27 a year earlier — while those referrals turn out to be correct at the same rate. DWP's own words: "the model is not working as effectively as it could for this cohort".

DWP's answer is the human review. "There is always a human intervention and decision." The caseworker is not told the model made the referral, and model referrals are mixed with random ones "to mitigate against human bias". That is a careful design, and it protects the review from the model. Nothing published shows whether the review catches what the model gets wrong.

Since 5 February 2026 the law has turned on whether human involvement in a decision is "meaningful". And under regulations in force since May 2026, the ICO is writing a statutory code on AI and automated decision-making that — through a provision few have noticed — courts and tribunals must take into account.

The human moment

The referral a caseworker never knows the model made. A Universal Credit claimant asks for an advance; a model flags the request; a caseworker decides, without being told why the case is in front of them. The review that demonstrably changes outcomes comes later: between April and June 2026, 60 per cent of benefit appeals decided at hearing were overturned in the claimant's favour.

What's inside

Fifteen pages, written for a senior official who knows government well and AI less well:

• A note on the numbers — what is included, what was excluded, and why, including government's own £45bn estimate
• Decisions at a scale no other sector makes, and a transparency register that has grown almost fivefold
• The pressure point — the tribunal is the one review whose effect is published
• DWP's own fairness figures for its advances model, read in full
• Four risks, none of them about the technology failing
• The threat side — AI on both sides of the same decision
• The UK legal position, and the ICO code that will be evidence
• Beyond the UK — benefit eligibility and asylum are both high-risk in the EU
• Seven things that can be done now, none needing new law
• Three levels defined by where accountability sits, and three horizons
• An eight-workstream way in

The discipline behind it

Every figure carries a source: a department's own published assessment, the National Audit Office, the Public Accounts Committee, the Ministry of Justice's tribunal statistics, a tribunal decision, or the statute book. Where the briefing reports an absence — such as the asylum AI tools missing from the transparency register — every record on the register was read.

The most quotable number in the sector is excluded on the record: government's own £45bn a year from "full potential digitisation", whose methodology does not assess whether automation is feasible. A widely reported tribunal decision on a Home Office asylum refusal is used only as an illustration, attributed and stated as unproven, because it was not on the tribunal's published database.

Who it's for

A senior civil servant who has just picked up a department's AI programme, a member of an audit and risk assurance committee, or a digital leader asked to show that a department's human review is meaningful. It assumes government is well understood and AI is not.