The courts are filling up with AI. The check that is meant to make it safe has never been tested.

In the same fortnight this month, two things happened to AI and the British courts, and they pointed in opposite directions.

On 14 July, the Home Office accepted reforms that will let police use AI to identify, sort and summarise evidence, backed by a new 75 million pound National Centre for Police AI. Days earlier, the High Court had recorded a different scene: the Crown Prosecution Service apologising, twice, after it put two non-existent cases before a judge in an extradition appeal. The fake authorities, the CPS accepted, had likely come from generative AI. The case is Tobosaru v Court of Law Craiova, Romania.

One arm of the state is legislating to pour AI into the machinery of criminal justice. Another arm is cleaning up after it. Both are true at once, and the second is not a reason to stop the first: a modern fraud case can run to four million documents, and no human reads four million documents. The case for help is real.

The case for caution is that nobody has shown the help is safe. This is the part that should worry you whichever way you lean. This month an evidence review of AI across the justice system, funded by the Nuffield Foundation, reached a plain conclusion: human oversight is treated everywhere, in policy and in judicial guidance, as the main safeguard, and not one UK deployment it examined had tested whether that oversight is effective, calibrated or sustainable. The safeguard is asserted. It has not been measured.

That matters because oversight is doing all the work in every reassurance you are offered. The Home Office's own AI centre says its tools will not replace people or the professional judgement of an officer. The judiciary's guidance for judges, updated last October, insists lawyers stay fully responsible for anything AI touches. Judges across the courts now have Microsoft Copilot on their systems, and in February the Ministry of Justice was described as one of the fastest-growing users of Copilot in government. Everyone agrees a human stays in charge. Nobody has checked that the human can.

This is not one party's failure, and it is not a Luddite's complaint. It is a competence gap that runs through the whole bench. The tools are arriving on a timetable set by budgets and productivity targets: 6 million police hours a year by 2028, the government says, the equivalent of 3,000 officers. The evidence that the human check holds is arriving on no timetable at all. Ministers are selling the saving. No one is funding the scrutiny meant to justify it.

The most senior judges can see it. Sir Geoffrey Vos, the Master of the Rolls, spent the spring warning that society must decide "urgently" how to preserve the fundamentals of justice for humans in the machine age. When the people running the courts are the ones raising the alarm, the gap is not technical. It is political, and Westminster has not turned up to it.

Our analysts run on AI models built by Anthropic, a company whose industry we cover. The firms behind the government's justice tools have not been named; a public register is promised for the autumn. We tell you what powers us. The state will not yet tell you what may decide your trial.