CPS apologises for AI hallucinations in court documents

TL;DR:

• The High Court has accepted two apologies from the Crown Prosecution Service after AI was used to compose extradition documents citing authorities that did not exist.
• The CPS said the operative cause was human error — a reviewing lawyer failing to verify the citations — rather than the AI itself.
• The judge accepted there was no intent to mislead, but recorded the episode given the consequences it might have had elsewhere.

The Crown Prosecution Service has apologised twice to the High Court after using AI to compose court documents citing authorities that do not exist. The judgment in Andreea-Maria Tobosaru v Court of Law Craiova, Romania records that the fabricated citations were relied on in grounds of opposition prepared by the CPS, then repeated in a further document in the appeal.

An apology that needed correcting

The sequence is unflattering. The CPS first told the court the error had been spotted by counsel for the appellants rather than the respondents. That account was itself inaccurate: junior counsel for the respondents had also failed to locate the authorities and raised it with the CPS shortly beforehand. A “further apology for this error” followed.

The non-existent authorities were deployed to argue the continued application of the House of Lords’ reasoning in Pilecki v Poland [2008]. In a letter dated 3 March 2026, the CPS accepted the citations were inaccurate and said they likely originated from AI use — but identified the critical failing as the reviewing lawyer not checking the document before filing. Its position: while the immediate source was generative AI, the operative cause was human error in failing to verify authorities placed before a court. An internal review, it said, showed the risk of recurrence was low.

Mr Justice Sweeting, dismissing the appeals, accepted there was no intention to mislead. He nonetheless set the episode out: “It would be naive to assume that there will not be an increasing use of artificial intelligence in legal work in future; indeed, that may be both necessary and beneficial. The episode highlights the risks of its use without appropriate oversight particularly for legal research.”

The timing sharpens the point. The apology surfaced the same day the Home Office announced it will legislate to let police use AI to review evidence — the state scaling AI into justice while one of its prosecutors explains AI-invented case law. It also complicates the judiciary taskforce’s suggestion that lawyers could be negligent for not using AI.

Looking forward

The errors were caught before the appeal and had no impact on the judgment, which is why the CPS calls this an isolated incident. But the judge’s reason for recording it was the harm a similar error might cause where nobody checked. For UK firms putting AI into any document that carries legal weight, the CPS framing is the transferable lesson: the failure was not the model inventing a citation, but a workflow that let it through unverified.