AI in UK Policing & Public Safety

Why this briefing

The machinery being built is national. PoliceAI, the national centre for artificial intelligence in policing, is funded at £75m over three years and hosted by the College of Policing. A national facial recognition system has £26m behind it. The Police Reform White Paper proposes buying technology "once rather than 43 different ways." The direction is settled and the money is committed.

The accountability has not moved with it. There are forty-three territorial forces in England and Wales, and each chief officer answers to a locally elected commissioner for what happens in their force area. Retrospective facial recognition already runs at more than 25,000 searches a month against the Police National Database. Twenty-seven forces use robotic process automation. The capability arrives centrally; the consequence lands locally.

The sequence is the second problem, and the Government is the source for both halves of it. Its own consultation states that the current legal basis for police facial recognition rests on "common law powers and a patchwork of equalities, human rights and data protection laws" and "does not give the police sufficient confidence to use it at significantly greater scale." That consultation closed on 12 February 2026. On 7 July 2026 the Home Secretary told a Lords committee the framework was "still being refined." Forty more live facial recognition units are funded in the meantime.

The human moment

The stop, and the custody decision. Deaths in or following police custody rose to 27 in 2025/26, from 18 the year before, against an eleven-year average of 19. Not one of the thirteen forces inspected under the PEEL 2025–27 programme was graded good or outstanding for investigating crime; none was graded good or outstanding for the custody environment either. That is the record any model would be trained on, and the process any model would be inserted into. It is also, for the person being stopped or held, the only part of policing that is not abstract.

The case that carries it

In November 2025 West Midlands Police gave a Safety Advisory Group a written risk report containing a claimed previous fixture between Maccabi Tel Aviv and West Ham that had never happened. The claim came from an unauthorised Microsoft Copilot query, and it was not initially disclosed. Away fans were banned from Villa Park, the decision became a national political argument, the Chief Constable retired, and the IOPC opened an independent investigation that includes a voluntary conduct referral about the force's use of AI.

The Home Affairs Committee established all of this on the record in February 2026. It made no recommendation about police use of AI at all. That absence is the briefing's subject as much as the incident is.

What's inside

Twenty pages, written for a sceptical chief officer team:

• A note on the numbers — what is included, what was excluded, and why
• The money, the settlement, and two published figures for the same AI centre that nobody has reconciled
• The record the technology is being fitted to — investigation, custody, charge rates and workforce
• Where AI actually is today, across facial recognition, automation and redaction
• What two oversight bodies are counting, and what the ICO's facial recognition audits found
• The whole lifecycle, from public contact to disclosure and the case file
• The overlooked legal change — and why it is not section 80
• The threat side, running at the evidential and accountability machinery rather than at the force
• Where the accountability actually sits when a nationally procured tool gets something wrong
• An action plan, and a first engagement that does not wait for the framework

The discipline behind it

Twelve figures were found, were usable, would have improved the argument, and were excluded. The widely repeated "6 million hours of work annually, equivalent to freeing 3,000 officers from desk duties" has no published methodology, no baseline and no definition of an hour released; it appears only as an attributed government ambition. The inspectorate's own "up to 15 million police hours" is a primary source without a published derivation, and is quoted as a stated view rather than used as a number. A reported figure for how many forces still use Microsoft Copilot was excluded because its method could not be established — which is a real loss, because it would have measured what changed after the failure.

The stop and search disproportionality ratio was dropped rather than used, because the official dataset still shows April 2022 to March 2023. Using a three-year-old ratio in a 2026 briefing is the date trap this series has been caught by before. Where two government figures for the same thing disagree — £115m and £75m for the AI centre, published five months apart — the briefing says the number moved, rather than choosing one.

One finding points the other way and is reported as found: the charge rate for victim-based offences rose, from 6.3 to 7.3 per cent. The briefing says so, rather than reaching for a decline that is not there.

Who it's for

A chief officer team, a police and crime commissioner's office, a force digital or data lead, or anyone on a board being asked to approve an AI capability that was procured somewhere else. It assumes policing is well understood and AI is not.