UK police told to stop using Copilot for court statements — and 'just review the output' is the part that breaks

The order

According to the FT report (paywalled to our fetch; summarized via the thread), a senior UK policing figure told forces in England and Wales to stop using commercially available AI tools — Microsoft Copilot among them — to produce court and witness statements until those tools are properly assessed. The deployment, in other words, ran ahead of the evaluation. The stated mitigation is a policy requiring officers to "check everything that it produces."

Why the safeguard collapses where it counts

The thread didn't argue AI is useless for police paperwork. It argued the safeguard is misapplied:

"Not only are they using AI before they've properly assessed them, they also end up using Copilot which must be one of the worse AIs currently available … they hope to be able to rely on 'Please review the outputs' which obviously isn't an ac[ceptable safeguard]." — a developer building local-first software

The sharpest distinction in the thread was between two very different kinds of text:

"Subjective information should be input by the police, because I would assert the specific wording matters. It matters that the words used to describe what the policeman saw comes out of the policeman's brain." — a commenter separating boilerplate from sworn evidence

That's the crux. A model can format an address block. It should not be the originating author of "the suspect appeared aggressive" in a document a court treats as the officer's own recollection — because a human skimming a fluent paragraph for errors is exactly the review least likely to catch a plausible-sounding fabrication.

A better default

One commenter pointed at how the evidentiary system already solved this:

"Courts prefer to have live witness testimony for a good reason. Detectives prefer to have statements made with the events as fresh as possible for a good reason … Where text is needed, auto-transcribe." — a commenter proposing transcription over generation

Transcription captures what was actually said; generation invents what might plausibly have been said. The halt isn't anti-AI Luddism. It's a recognition that "review the output" is a control designed for low-stakes drafting, quietly load-bearing in a setting where the precise words can put someone in prison.