The short version
An authority that does not hold up when someone opens it. It comes in three grades: a case that does not exist; a real case with an invented quotation or paragraph; and a real case cited for a proposition it does not support. A search for the case name catches the first grade only.
What it is
Grade one is the invented case. The name, the year and the citation are all made up, and no database holds it.
Grade two is the real case with invented words. The name and the neutral citation are right, so a search finds the case, but the quotation is not in the judgment or the paragraph number points somewhere else.
Grade three is the real case cited for the wrong point. The case exists and the words may be accurate, but the court did not decide what the citation says it decided.
The judicial guidance names all three. Its list of what AI tools may do begins: "make up fictitious cases, citations or quotes, or refer to legislation, articles or legal texts that do not exist", and goes on to "provide incorrect or misleading information regarding the law or how it might apply" (Artificial Intelligence (AI): Guidance for Judicial Office Holders, 31 October 2025, section 3, part III).
A sighting
All three grades sat in one schedule in Al-Haroun v Qatar National Bank, decided with R (Ayinde) v London Borough of Haringey [2025] EWHC 1383 (Admin) on 6 June 2025. The schedule listed forty-five citations. In eighteen instances the case cited did not exist. Of the others, some "did not contain the quotations that were attributed to them", some "did not support the propositions for which they were cited", and some "did not have any relevance to the subject matter of the application" (paragraph 74). Read the judgment at caselaw.nationalarchives.gov.uk/ewhc/admin/2025/1383.
The first two grades can shade into each other. In Harber v Commissioners for HMRC [2023] UKFTT 1007 (TC), the Tribunal found that nine decisions relied on by the appellant were not genuine (paragraph 21). One of them, cited as "Baker v HMRC (2020)", resembled a real decision of 2018, but the year was different and the real appellant had lost (paragraph 17). HMRC's litigator had searched the Tribunal's website for each one by name and year, widening the years where a name was common (paragraph 16). The Tribunal added: "But that does not mean that citing invented judgments is harmless", because they cause the Tribunal and HMRC to waste time and public money (paragraph 24). Read it at caselaw.nationalarchives.gov.uk/ukftt/tc/2023/1007.
Why it happens
A language model writes text that fits the pattern of what it has seen, and a citation is a strong pattern: party names, a year, a court, a paragraph number. The judicial guidance puts the point plainly: "Public AI chatbots do not provide answers from authoritative databases." A tool that does search a database narrows the first grade but not the other two, because it can retrieve a real case and still misdescribe it (see "Hallucination-free").
What to say back
"Has each authority been opened, and is the quoted passage at the paragraph cited? A name that comes up on BAILII shows the case exists, not that it says this."
Origin
- 4 May 2023. In Mata v Avianca, a United States case, the court orders one of the plaintiff's lawyers to show cause why he should not be sanctioned for "citing non-existent cases to the Court" and for submitting "copies of non-existent judicial opinions" (Opinion and Order on Sanctions, 22 June 2023, Findings of Fact, paragraph 42).
- 22 June 2023. The same court sanctions the lawyers, who had filed "non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT" (Opinion and Order on Sanctions, page 1). Inside one of the invented opinions it found real decisions that "have correct names and citations but do not contain the language quoted or support the propositions for which they are offered" (Findings of Fact, paragraph 29(g)). All three grades were present in that one case.
- 4 December 2023. Harber: the First-tier Tribunal (Tax Chamber) finds that decisions relied on by a litigant in person were generated by an AI system.
- 6 June 2025. Ayinde and Al-Haroun: the Divisional Court sets out the duty to check "by reference to authoritative sources" (paragraph 7) and the court's powers when it is not met (paragraph 23).
- 31 October 2025. The current judicial guidance, replacing the version of April 2025, carries the list of hallucinations quoted above.
- Now. On 1 October 2026, Damien Charlotin's AI Hallucination Cases database gave "2097 cases identified so far", 70 of them filed under the United Kingdom, last updated 30 September 2026. It counts only decisions where "the court or tribunal has explicitly found (or implied) that a party relied on hallucinated content or material".