18 of 45 citations did not exist. What a court said about AI drafting.

The facts are ordinary, which is what makes them useful. Nobody in either case was trying to deceive a court. In one, a barrister’s grounds for judicial review cited five cases. In the other, a witness statement supporting an application cited forty-five. The court checked.

What exactly went wrong?

In Ayinde, the grounds cited cases with plausible names, plausible citations and plausible propositions — El Gendi, Ibrahim, H v Ealing, KN v Barnet, Balogun — and none of them existed. The court’s finding at paragraph 36 is one line: “The case that is cited (El Gendi) does not exist.”

In Al-Haroun, the client did the research and the solicitor relied on it. The client accepted responsibility and said the citations “were generated using publicly available artificial intelligence tools, legal search engines and online sources”. The solicitor accepted that his witness statement contained citations of non-existent authorities and that he had relied on the client’s research “without independently verifying” it. Paragraph 74 records the arithmetic: forty-five citations, eighteen of which do not exist, and of those that did, “in many instances they did not contain the quotations that were attributed to them, did not support the propositions for which they were cited, and did not have any relevance to the subject matter of the application”.

What did the court say about the tools?

Paragraph 6, in full: “Freely available generative artificial intelligence tools, trained on a large language model such as ChatGPT are not capable of conducting reliable legal research.”

And paragraph 7, which is the sentence to keep: “Those who use artificial intelligence to conduct legal research notwithstanding these risks have a professional duty therefore to check the accuracy of such research by reference to authoritative sources.”

What were the consequences?

Wasted costs of £2,000 each against the barrister and the law centre in Ayinde. The barrister referred to her regulator; the solicitor and the law centre referred to the Solicitors Regulation Authority. In Al-Haroun, the solicitor and his firm referred to the SRA. The court listed the powers available in cases like these at paragraph 23: public admonishment, costs orders, wasted costs orders, striking out, referral to a regulator, contempt proceedings, and referral to the police.

Paragraph 9 says why it matters beyond the two cases: “There are serious implications for the administration of justice and public confidence in the justice system if artificial intelligence is misused.” The judgment was sent to the Bar Council, the Law Society and the Council of the Inns of Court, and it tells heads of chambers and managing partners that in future the court will ask whether their leadership responsibilities have been fulfilled. The appendix lists similar cases from England and Wales, the United States, Australia, New Zealand and Canada.

Why is this on a page for brokers and advisers?

Because nothing in the reasoning is specific to law. Substitute “suitability report” for “grounds”, “FCA Handbook” for “authorities”, and “file review” for “Divisional Court”, and every sentence still holds. An IFA who lets a tool draft a report that cites a rule which does not exist has the same problem as the barrister, in front of a different regulator. The tool is not capable of reliable research; the duty to check is the professional’s; the consequences follow the professional.

Three things follow for any regulated firm that uses AI to write:

1. Every claim the draft makes has to be traceable to a source the firm holds. Not “the model said”. The document, the page, the date. A draft that cannot show its sources is not a draft — it is a liability with formatting. That is how we build first-draft systems, and the rule is the same one this site applies to its own figures.
2. The human check has to leave a record. The court’s question in Al-Haroun was, in effect, “who verified this?”, and the answer was nobody. The defensible answer is a named person, a timestamp, and what they changed. In a regulated firm the audit trail is the product.
3. A system that cannot say “I do not have that” will invent it. The failure in both cases was fluency without retrieval. An answer with no citation should be returned as no answer, and the refusal should be logged, because the refusal is the evidence that the guardrail exists.

The judgment is public and short. If your firm has an AI policy that has not been rewritten since it was handed down, it was written for a different world.