AI-Fabricated Authorities in UK Courts: The Cases and Verification Rules, 2023–2026

A High Court judge referred a case to the Divisional Court in which one of the authorities cited was a judgment she was said to have decided. It did not exist. Neither did seventeen others filed in the same claim. Since late 2023, courts and tribunals across the United Kingdom have been confronting invented law cited as real — and in answering it, they have begun to define the standards that lawyers, supervisors and litigants in person are expected to observe before citing authority. This article maps the principal reported decisions, quoted and linked to the paragraph level of the judgments themselves.

Disclosure: This article is published by Search the Law, a commercial legal research platform. Where the article references legal research tools, readers should be aware of that connection. The legal analysis in this article has been independently verified against the primary sources cited.

Last verified: 23 July 2026. This area is developing at speed; propositions below are accurate to that date.

A Body of Law Nobody Planned

Three years ago, the citation of a fabricated authority in a UK court was a curiosity. The first reported instance — Harber v HMRC [2023] UKFTT 1007 (TC) — involved a taxpayer facing a £3,265 penalty who relied on nine invented tribunal decisions, described by the tribunal as “plausible but incorrect,” supplied to her by a friend. She had never heard of the tribunal’s own website. The decision was reported, briefly noted, and treated as a cautionary oddity about litigants in person and chatbots.

By July 2026, a practitioner-maintained tracker had recorded sixty-seven UK judgments or incidents involving confirmed or suspected AI-related false citations; a global database maintained by Damien Charlotin records well over 1,700 worldwide. Those trackers include suspected as well as confirmed cases, and they are not official statistics — but the reported decisions alone now run from the First-tier Tribunal to a Divisional Court, and they have hardened from judicial tutting into something like doctrine: a leading case, a ladder of consequences applied in practice, a supervision principle, and — in one chamber so far — a procedural definition of verification itself. Most of it was built in eighteen months.

What follows is a map of the principal reported decisions: what each one added, and what they now collectively require.

The Leading Case: Ayinde and Al-Haroun

The framework decision is R (Ayinde) v London Borough of Haringey; Al-Haroun v Qatar National Bank QPSC [2025] EWHC 1383 (Admin), a judgment of a Divisional Court comprising Dame Victoria Sharp P and Johnson J, handed down on 6 June 2025 under the court’s Hamid jurisdiction — the High Court’s inherent power to police the conduct of those who appear before it. Two unrelated cases arrived within weeks of each other, and together they gave the court both ends of the problem: the junior lawyer, and the lay client whose lawyer stopped checking.

In Ayinde, a pupil barrister settled grounds in a homelessness judicial review citing five authorities that did not exist. What converted an error into a scandal was the response. When the defendant’s solicitors asked for copies of the cases, the reply — in writing — dismissed the request as an attempt to “barter our client’s legal position for cosmetic errors” (quoted at [42]). Fake law, filed in the High Court, recharacterised as a formatting slip. At the wasted costs hearing, Ritchie J described the “cosmetic errors” characterisation as “grossly unprofessional,” ordered £2,000 in wasted costs against counsel and the instructing law centre alike, and referred both to their regulators.

In Al-Haroun, the claimant in an £89.4 million claim filed witness statements citing forty-five authorities. Eighteen did not exist; many of the rest said nothing like what they were cited for. The judicial assistant tasked with checking them reported that “the vast majority of the authorities are made up or misunderstood” ([74]). The citations had been generated by the client, using publicly available AI tools; his solicitor adopted them wholesale. 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” ([79]). One of the invented authorities was attributed to Dias J — the judge who referred the matter.

From those facts the Divisional Court drew the propositions that now govern. Freely available generative AI tools “are not capable of conducting reliable legal research” and “may cite sources that do not exist” ([6]). Accordingly, those who use AI in legal research “have a professional duty… to check the accuracy of such research by reference to authoritative sources, before using it in the course of their professional work” ([7]) — and the court named what counts as authoritative: the Government’s legislation database, the National Archives’ case law service, the official Law Reports, and the databases of reputable legal publishers. Where fake material nonetheless reaches a court, “save in exceptional circumstances, admonishment alone is unlikely to be a sufficient response” ([31]).

Then the passage everyone should read twice. On either version of the facts before it concerning the pupil barrister, the court held that “the threshold for initiating contempt proceedings is met” ([68]). It declined — on the particular facts, against a very junior barrister already under regulatory investigation — to initiate them. And it immediately closed the door it had just declined to walk through: that decision “is not a precedent. Lawyers who do not comply with their professional obligations in this respect risk severe sanction” ([69]).

The mercy was personal. The threshold finding is what remains on the books.

The Consequence Ladder, As Actually Applied

Ayinde set out the full menu of judicial responses at [23]–[32]: public admonition; costs orders; wasted costs orders under section 51 of the Senior Courts Act 1981; striking out; regulatory referral; contempt proceedings; and — at the top — referral to the police, because deliberately placing false material before a court with intent to interfere with the administration of justice is the common law offence of perverting the course of justice, carrying a maximum of life imprisonment.

What makes 2025–26 remarkable is that the courts have now used almost every rung:

Strike-out for abuse of process. In Bandla v Solicitors Regulation Authority [2025] EWHC 1167 (Admin) (13 May 2025), a struck-off former solicitor’s appeal cited twenty-five cases that did not exist. Fordham J struck out the grounds of the court’s own motion, holding that the court “needs to take decisive action to protect the integrity of its processes against any citation of fake authority.”

Wasted costs against the firm. In Ndaryiyumvire v Birmingham City University (Birmingham County Court, judgment 14 October 2025, published the following month), two fabricated cases reached a signed court document via the AI research feature built into the firm’s practice-management software. The wasted costs order was made against the firm, the judge describing the episode as, in substance, a failure of management — with the order automatically reportable to the SRA.

Referral of the supervisors. In Rodney v Gee’z Micro Bar and Pitstop (County Court, 23 April 2026), a paralegal used the same software family’s “legal research assistant” tools; the supervising solicitor signed the statement of truth without checking. Applying Ayinde, the judge held admonishment alone insufficient, referred both solicitors to the SRA, and directed that the judgment be published.

Referral for attempting to mislead. In MS (professional conduct: AI generated documents) [2025] UKUT 305 (IAC) (promulgated 12 August 2025), a barrister whose ChatGPT-drafted grounds cited a fictitious Court of Appeal authority compounded the problem during the hearing itself — producing further unverified AI output over the short adjournment — and was found to have “attempted to mislead” the tribunal ([12]). He was referred to the Bar Standards Board.

The warning shot for everyone else. In Rafique v HMRC [2026] UKFTT 673 (TC) (6 May 2026), the tribunal found a litigant in person’s submissions were AI-generated, with hallucinated authorities. It acknowledged sympathy for the perception that AI offers the unrepresented “a more level playing field” — and then set the boundary for all litigants, represented or not: “no person has immunity from sanction” ([100]), and “it is a contempt of court for fabricated authorities to be cited to a court or tribunal as if they are genuine” ([102]).

That warning must be read with Ayinde at [26]. Contempt is not established by negligence alone: the person must know that the material is false, or lack an honest belief that it is true.

One rung remains unused. We have found no reported UK decision, as at 23 July 2026, committing a person for contempt specifically for AI-fabricated citations. In Olsen v Finansiel Stabilitet A/S [2025] EWHC 42 (KB) (16 January 2025), Kerr J reached that restraint “narrowly and somewhat reluctantly” in the case of two elderly litigants in person. Several decisions have warned that future cases may attract more severe consequences — the Divisional Court’s “not a precedent” being the plainest.

No Longer a ChatGPT Story

The early incidents let the profession tell itself a comfortable story: this is about desperate litigants and lazy juniors misusing free chatbots. The 2026 cases have dismantled that comfort item by item.

The fabricated authorities in Ndaryiyumvire and Rodney did not come from ChatGPT. They came from the AI research features of mainstream practice-management software — tools firms already pay for, sitting inside workflows firms already trust. In UK and R (Munir) v Secretary of State for the Home Department [2026] UKUT 81 (IAC) (“UK and Munir”, promulgated 17 November 2025), a fictitious Court of Appeal authority called Horleston appears to have arrived through Google’s “AI Mode” during an ordinary search. The tribunal did not take that on trust: the panel ran the search itself and watched Google’s AI generate different, equally plausible, equally fictitious benches for the same non-existent case ([17]). A specialist tribunal, mid-hearing, reproducing a hallucination experimentally — from the search bar every lawyer in the country uses between meetings.

Nor is it only invented cases. In Cork v Smith [2026] EWHC 1199 (Ch) (22 May 2026), a letter from a major national firm set out the purported text of rule 12.37(5) of the Insolvency Rules 2016 in support of an application. The judge went looking for the quoted words, checked the surrounding chapter in case the paragraph reference was merely wrong, ran word searches for the phrases — and “was unable to find them in the IR 2016 at all,” recording that his “first thought was that this was an artificial intelligence… ‘hallucination’” arising during the firm’s research ([20]). Fabricated legislation, not fabricated case law — the same failure, one level deeper.

And it reaches institutions as well as firms. In Tobosaru and Tofan [2026] EWHC 1720 (Admin) (8 July 2026), two authorities cited in the CPS’s extradition submissions did not exist. The CPS accepted the citations were “likely to have originated from the use of artificial intelligence” — but identified “the critical failing as the fact that the reviewing lawyer did not properly check the accuracy of the document before it was filed and served” ([99]): “the operative cause was human error in the failure to verify the authorities relied upon” ([100]). Senior managers then reviewed seventy-eight other cases handled by the same lawyer and found no similar issue ([101]). That is the principal public prosecution service for England and Wales, stating on the record that the problem is not the software but the unverified workflow. The Family Court has met the same phenomenon in children proceedings: Re A, B, C, D (Extension of Assessment; Use of AI: Hallucinations) [2026] EWFC 71 (B) (5 March 2026) carries the subject in its very name.

Nor do the subscription platforms escape. Peer-reviewed research at Stanford — testing particular versions of the leading US legal AI research products at a particular time — found they produced incorrect or misgrounded answers on a significant minority of queries, roughly one in six for the best performer and one in three for another, even as marketing spoke the language of “hallucination-free” citations. Those figures are not permanent error rates for every current version, but they demolished the assumption that a professional subscription discharges the checking. We have examined that research in a separate analysis. The tribunals have likewise rejected the idea that sophisticated use is a defence: commenting on Zzaman v HMRC [2025] UKFTT 539 (TC) — itself a warning about the “dangers of reliance on AI tools without human checks” — the Divisional Court in Ayinde was blunt: “The critical safeguard is to check any output by reference to an authoritative source.”

The upshot is uncomfortable but clarifying. “I don’t use ChatGPT” now offers little reassurance, because AI is increasingly embedded in search engines, practice software and drafting tools. The duty attaches not to the tool but to the output: if it cites authority, the authority gets checked.

The Supervision Principle

If one line from this body of law belongs above every supervising partner’s desk, it is this, from the Upper Tribunal in UK and Munir: whether false authorities are “inserted by a hapless trainee or by ChatGPT is really neither here nor there; the point is that the qualified legal professional with conduct of the matter is expected to ensure that such documents are checked” ([37]).

The tribunal’s headnote goes further: a supervisor who fails to check a junior’s work is likely to be more culpable than a lawyer whose own work contains the hallucination. In that case a solicitor of fifteen years’ unblemished practice was referred to the SRA over grounds drafted by a junior member of staff. The tribunal also warned that uploading confidential documents to an open public AI tool places that information, in effect, in the public domain — and may breach confidentiality and waive privilege. Privilege was not a contested issue requiring determination, so that observation should be treated as a warning rather than a binding ruling; we examined its practical consequences in our analysis of UK and Munir.

There is a genuine debate to be had about where the blame has been allocated. The pupil barrister in Ayinde was named in a national judgment at the start of her career; the software vendor whose “research assistant” invented the authorities in two county court cases faced no judicial consequence at all. The courts have picked their answer — the signature carries the risk, and seniority deepens rather than dilutes it. Whether the profession regards that allocation as fair is a question worth debating openly, rather than one Hamid hearing at a time.

The People With the Least Protection

Running underneath all of this is an access-to-justice problem the judgments keep brushing against. The Master of the Rolls put it plainly in a speech this February: for the unrepresented, the first port of call used to be a lawyer, if one was available and affordable — “now the first port of call is ChatGPT or CoPilot.”

The sequence of litigant-in-person cases — Harber, Olsen, Zzaman, Gunnarsson [2025] UKUT 247 (TCC) (July 2025, where the Upper Tribunal warned that AI “should not be relied upon without checking”), Jeffs v Lewisham (where the AI invented the wording of the regulations themselves), Rafique — shows tribunals straining to be fair to people doing their honest best with the only research assistant they can afford. The sympathy is real. So is the direction: Rafique ends with an express judicial warning to a litigant in person about contempt.

Sit with the asymmetry. The people most likely to rely on a free chatbot are the people least able to afford the tools that would catch its inventions — and they are now the ones being warned about contempt. One 2025 study, testing three general-purpose models on matched problems, found higher hallucination rates on London-law questions than on Los Angeles-law questions — a caution against assuming that what holds for US law holds here. Whatever one’s view of the individual decisions, this is not a stable position for a justice system to rest in — and it will not be solved by guidance documents addressed to lawyers. We have written elsewhere about what this moment looks like from the unrepresented side.

Verification Now Has a Working Definition

The most practical development in the whole line of cases is procedural, and it came from the Upper Tribunal. Its amended judicial-review claim form requires a legal representative to confirm, by statement of truth, that every cited authority:

(a) exists; (b) can be located using the citation provided; and (c) supports the proposition of law for which it is cited.

The tribunal said that other forms and directions are to be similarly amended, and that a representative who signs such a statement in a case where false authorities appear “should ordinarily expect to be referred” to their regulator (UK and Munir, headnote). This is not yet a universal court rule — it is a requirement of one chamber — but it is the clearest judicial formulation to date of what meaningful citation verification entails, and it is fairly described as an emerging judicial framework rather than settled law of general application.

The regulators have been converging on the same standard from their own directions: the judiciary’s October 2025 guidance for judicial office holders, the Bar Council’s November 2025 update (large language models “are not a substitute for… diligent verification”), and the Bar Standards Board’s first dedicated AI guidance, published on 18 May 2026, which says free public tools will generally be too high-risk for legal or client-specific work. We have covered the judicial guidance and the BSB’s verification duty in detail separately. The Civil Justice Council’s consultation on AI in the preparation of court documents closed on 14 April 2026. In an update published on 30 June, the Council reported a high degree of alignment among respondents that no additional AI-specific formal requirements are presently necessary for pleadings, skeleton arguments and similar documents prepared by legal professionals, because existing professional-responsibility frameworks are considered sufficient. Its emerging position remains subject to finalisation: witness statements, expert evidence and the distinct problems facing litigants in person remain under consideration, with the final report expected later in 2026.

The three limbs deserve to be read as what they are: a working definition of verification. And the third limb is the one that will decide careers, because it is the failure mode no marketing department mentions. A fabricated citation is often exposed by the first existence check. A real case cited for a proposition it does not support — the “misgrounded” citation — may survive that check and follow the advocate into the hearing. Limbs (a) and (b) can often be tested by locating the authority from its citation. Limb (c) requires reading the cited passage in the context of the judgment.

What This Body of Law Actually Asks of You

Strip away the sanctions and the Latin, and the emerging law reduces to three working rules.

Verify at the paragraph, not the citation. The three limbs make citation-checking necessary but nowhere near sufficient. The question is always whether the source, read at the pinpoint and in context, supports the proposition advanced. That is the practical implication of Ayinde’s verification duty at [7], read with the three-limb confirmation introduced in UK and Munir.

Supervision is not delegation. After UK and Munir, Ndaryiyumvire and Tobosaru, the person who signs, and the organisation whose systems let the document out, carry the risk. A junior’s draft, a client’s research and a software feature’s output are all, in the courts’ eyes, the same thing: unverified input.

The tool matters less than the check. The cases now span free chatbots, search-engine AI features, the AI tools inside professional practice software, and — in Cork — the statute book itself. No tool’s branding transfers the duty, and no tool’s failure excuses it. The only thing that discharges it is reading the authority — which makes the practical question for any research workflow, human or machine, how quickly it lets you get from a proposition to the paragraph of the judgment said to support it.

Search the Law is designed to shorten that final verification step: locating the authority, opening the full judgment, and moving to the passage said to support the proposition. It does not replace professional judgment; it makes the checking process faster and more transparent.

That standard is the one this article has tried to hold itself to. Every quotation above is linked to the paragraph of the judgment it comes from, on the full text of the case. If a proposition here matters to your practice — or if you simply disagree with it — the paragraph is one click away. Go and read it. That habit, more than any tool, prompt or policy, is what the courts are now demanding of all of us.

Sources and Authorities

Case	Citation / reference	Court	Decision date	Relied on at
Harber v HMRC	[2023] UKFTT 1007 (TC)	FTT (Tax)	4 Dec 2023	[20]
Olsen v Finansiel Stabilitet A/S	[2025] EWHC 42 (KB)	High Court (KB)	16 Jan 2025	[113]
Zzaman v HMRC	[2025] UKFTT 539 (TC)	FTT (Tax)	3 Apr 2025	[29]
R (Ayinde) v Haringey (wasted costs)	[2025] EWHC 1040 (Admin)	High Court (Admin)	3 Apr 2025	[33], [46]
Bandla v SRA	[2025] EWHC 1167 (Admin)	High Court (Admin)	13 May 2025	[53]
R (Ayinde) v Haringey; Al-Haroun v QNB	[2025] EWHC 1383 (Admin)	Divisional Court	6 Jun 2025	[6]–[7], [23]–[32] (incl. [26], [31]), [42], [68]–[69], [74], [79]
HMRC v Gunnarsson	[2025] UKUT 247 (TCC)	Upper Tribunal (TCC)	23 Jul 2025	[112]
MS (professional conduct: AI generated documents)	[2025] UKUT 305 (IAC)	Upper Tribunal (IAC)	12 Aug 2025	[12]
Ndaryiyumvire v Birmingham City University	No neutral citation; published Judiciary transcript	County Court (Birmingham)	14 Oct 2025	transcript
UK and R (Munir) v SSHD	[2026] UKUT 81 (IAC)	Upper Tribunal (IAC)	17 Nov 2025	headnote, [17], [21], [37]
Re A, B, C, D (Extension of Assessment; Use of AI: Hallucinations)	[2026] EWFC 71 (B)	Family Court	5 Mar 2026	noted
David Jeffs v London Borough of Lewisham	LON/00AZ/LSC/2025/1032	FTT (Property Chamber)	16 Mar 2026	[15]–[16], [32]
Rodney v Gee’z Micro Bar and Pitstop	No neutral citation; published Judiciary transcript	County Court	23 Apr 2026	transcript
Rafique v HMRC	[2026] UKFTT 673 (TC)	FTT (Tax)	6 May 2026	[100], [102]
Cork v Smith	[2026] EWHC 1199 (Ch)	High Court (Ch)	22 May 2026	[18]–[21]
Andreea-Maria Tobosaru v Court of Law Craiolva, Romania; Ionut-Andrei Tofan v Secuava Local Court, Romania	[2026] EWHC 1720 (Admin)	High Court (Admin)	8 Jul 2026	[99]–[101]

Internal links open the full judgment text on Search the Law with paragraph anchors; external links are to the official judiciary.uk or GOV.UK publications for decisions not carried in the national case law archive.

Researching the Authorities

The case law in this area is developing quickly. Key search terms for tracking it include:

For the leading framework: “Ayinde v Haringey [2025] EWHC 1383”, “Hamid jurisdiction AI”, “wasted costs fake citations”, “professional duty to verify AI research.”

For the 2026 decisions: “UK and Munir [2026] UKUT 81”, “Cork v Smith Insolvency Rules AI”, “Tobosaru CPS artificial intelligence”, “Rafique fabricated authorities contempt”, “statement of truth cited authorities UTIAC.”

For litigants in person: “Harber v HMRC AI cases”, “litigant in person ChatGPT tribunal”, “Gunnarsson AI citations Upper Tribunal.”

For regulatory guidance: “BSB AI guidance 2026”, “Bar Council generative AI guidance”, “judiciary AI guidance October 2025”, “Civil Justice Council AI court documents.”

Search the Law is not a law firm and does not provide legal advice. The information in this article is for legal research purposes only. If you need advice about professional conduct or regulatory obligations, contact the Solicitors Regulation Authority, the Bar Standards Board, or the Law Society. Always read the full judgments linked above before relying on any proposition in this article.