AI in UK Justice, Courts & Probation

Why this briefing

Government's Algorithmic Transparency Recording Standard is mandatory across central government, and it is written to catch exactly the tools this sector runs. A record is required wherever a tool has "a significant influence on a decision-making process with public effect" — and significant influence expressly "includes where an algorithmic tool meaningfully assists, supplements, or fully automates a decision-making process." There is no justice carve-out, and no deadline.

The justice family is not a reluctant filer. Twelve records sit on the register across the Ministry of Justice, HMPPS, the Probation Service, HM Courts and Tribunals Service, the Office of the Public Guardian and Cafcass. HMPPS's record for its Violence in Prisons Estimator is a model of the form: model type, training data, a published bias and ethics assessment, an eighteen-month review cycle, and a plain statement that a human decision-maker is involved in every process where the estimates are used.

What a search of the register does not return is OASys, recidivism or parole — the actuarial instruments that have informed assessments about recall, categorisation and release for two decades. The department's own 2024 revalidation of those instruments reports "large miscalibration" by ethnicity on the Risk of Serious Recidivism tool, and nothing was published beside it.

The human moment

The risk assessment score. Recall, parole, categorisation and licence conditions each turn partly on a number produced about a person. In 2024, probation practitioners adequately assessed risk of harm in 28 per cent of cases, against 60 per cent in 2018-19 — the inspectorate's own finding. Since 5 February 2026, whether the human involvement in a decision like that was meaningful turns, by statute, partly on how far the decision was reached by profiling.

What's inside

Fifteen pages, written for a senior official who knows justice well and AI not at all:

• A note on the numbers — what is included, what was excluded, and why
• The system this sits inside, and why recall joins the two ends of it
• The pressure point — risk assessment quality halved, against a 25 per cent workload target
• What is on the transparency register, and what a search of it does not return
• Four risks specific to the sector, none of them about the technology working
• The threat side, running at the evidence rather than at the organisation
• The UK legal position — section 80(3), and the subsection nobody quotes
• Beyond the UK, stated as a comparator rather than an obligation
• Seven things that can be done now, none needing new law
• Three levels defined by where accountability sits, and three horizons
• Benefit realisation, and the measure missing from the sector's own criteria
• An eight-workstream way in

The discipline behind it

Every figure carries a source, and every absence was tested. Where the briefing says the register returns nothing for the actuarial instruments, a control search was run beside it — the same register returns seven records for "probation" — so the claim is what was actually tested, not what was assumed.

Two figures that would have made the briefing louder are excluded on the record. A widely repeated figure for how many people are profiled each day by a Ministry of Justice system reaches the public record only through an advocacy organisation and could not be traced to a departmental source. Subgroup accuracy figures for one sexual-offending predictor rest, in the smallest case, on three recorded reoffences in a group of 153 people, and the department's own report cautions against reading subgroups of that size. The briefing leads instead on the full-cohort finding.

Who it's for

A Ministry of Justice or HMPPS senior official who owns a delivery portfolio, a probation or prisons leadership team, or a board member being asked to approve AI in a decision about a person's liberty. It assumes the justice system is well understood and AI is not.