Fake cases, real consequences: the High Court on AI-invented citations

The first case, Ayinde v London Borough of Haringey, was a legal-aid housing judicial review. The written grounds, settled by a pupil barrister instructed by Haringey Law Centre, cited five cases that do not exist, including one attributed to the Court of Appeal. When the council pointed this out, the solicitors described the citations as “cosmetic errors”. In April 2025 Mr Justice Ritchie called that “a grossly unprofessional categorisation”, found the conduct improper, unreasonable and negligent, and ordered wasted costs of £2,000 against the barrister and £2,000 against the law centre. He made no finding that AI had been used; the barrister denied it.

The second, Al-Haroun v Qatar National Bank, was a damages claim for £89.4 million. A witness statement filed by the claimant's solicitor cited forty five authorities. Eighteen did not exist. Many of the rest did not contain the quotations attributed to them or had nothing to do with the point. The client said he had produced the research using “publicly available artificial intelligence tools, legal search engines and online sources”. The solicitor accepted he had relied on it without checking. The court found it “extraordinary that the lawyer was relying on the client for the accuracy of their legal research, rather than the other way around”.

Both cases were referred to a Divisional Court, Dame Victoria Sharp, President of the King's Bench Division, and Mr Justice Johnson, under the court's Hamid jurisdiction, which exists to enforce the duties lawyers owe to the court. Judgment was handed down on 6 June 2025 as [2025] EWHC 1383 (Admin).

On the Ayinde barrister, the court said the threshold for contempt proceedings was met: either she had deliberately included fake citations, or she had used AI and her sworn denial was untrue. It decided not to start proceedings, citing among other reasons that she was “an extremely junior lawyer who was apparently operating outside her level of competence”, and said plainly that the decision “is not a precedent”. She was referred to the Bar Standards Board. The head of the law centre was referred to the Solicitors Regulation Authority for an inadequate response. In Al-Haroun the court accepted the solicitor had not realised the citations were false, so the contempt threshold was not met, and referred him to the SRA, to which he had already reported himself.

The court's general statements matter more than the individual outcomes. At paragraph 6: “Freely available generative artificial intelligence tools, trained on a large language model such as ChatGPT are not capable of conducting reliable legal research. Such tools can produce apparently coherent and plausible responses to prompts, but those coherent and plausible responses may turn out to be entirely incorrect.” At paragraph 7, those who use such tools “have a professional duty therefore to check the accuracy of such research by reference to authoritative sources, before using it in the course of their professional work”. And at paragraph 9, heads of chambers and managing partners were told that “practical and effective measures must now be taken”, and that in future the court would ask whether they had been.

The court's appendix set out a selection of similar incidents from England and Wales, the United States, Australia, New Zealand and Canada, and said many more could be given. The best known is Mata v Avianca in New York, where in June 2023 a judge found that two lawyers had filed opinions invented by ChatGPT, then stood by them after the court questioned their existence, and imposed a $5,000 penalty on the lawyers and their firm jointly. The Bar Council had warned in January 2024 that such tools' outputs should not be taken “on trust and certainly not at face value”. The Divisional Court's view was that guidance on its own had proved insufficient.

The pattern is the same each time. A language model produces text that reads like the real thing. The person using it is under pressure, trusts the fluency, and skips the check. Then the output passes through one or more other people who assume someone earlier has verified it. In Al-Haroun the chain ran from the client to the solicitor to the court with nobody opening the cases.

It has not stopped. The Law Society Gazette reported in August 2026 that the SRA had received 42 reports of potential AI misuse between July 2025 and July 2026, and had issued a warning notice. The Bar Council updated its generative AI guidance in November 2025, and the Bar Standards Board's guidance on AI, in force from May 2026, cites the Ayinde judgment directly.

This is not legal advice, and most readers are not lawyers. But the court described a failure mode that applies to any business putting AI-generated text in front of someone who will rely on it: a quote, a policy answer, a compliance statement, a figure in a report. Four controls follow from the judgment.

The person who sends it owns it. The court was clear that relying on AI, a colleague or a client does not transfer responsibility. Decide who signs off each kind of AI-assisted output and make sure they know the signature is theirs.

Check against the source, not against the output. The failure was not using AI; it was not opening the cases. Any claim that can be verified against an authoritative record should be, before it leaves the building. For customer-facing assistants, that means answering only from your own current documents and citing them.

Keep a record of what was checked. The Ayinde barrister's list of cases was deleted and the court asked the regulator to look into why. A system that records what the AI produced, what a person checked and what was changed protects the business and the person.

Leadership is accountable for the process. The court addressed heads of chambers and managing partners, not only the individuals. If your team uses these tools, the questions are whether there is a written rule, whether people are trained on it, and whether anyone checks it is followed.

The AI workflows we build for businesses are grounded in the business's own sources, so an assistant answers from the current price list or policy rather than from memory, and says when it cannot. Anywhere a mistake would matter, a named person approves before anything is sent, and every action is recorded so the check can be shown afterwards. We test against real cases, including the ones the system should refuse.

We also apply it to ourselves. Slipmark, our open-source check for AI coding agents, exists because an agent's own account of what it did is a claim, not evidence; it binds each approval to the exact files inspected and refuses to ship if they change. The High Court reached the same conclusion about a different kind of claim. The tool is useful. The check is the work.