Key Takeaway
This legal article brings together eight recent UK cases of AI Hallucinated/False Citations (Confirmed or Suspected), together with one further reported incident for which I have not yet located the judgment. It now also includes the first case before the Solicitors Disciplinary Tribunal concerning a lawyer’s use of AI in legal proceedings. The UK tracker now contains 76 entries by Matthew Lee.
Ad/Marketing Communication
This legal article/report forms part of my ongoing legal commentary on the use of artificial intelligence within the justice system. It supports my work in teaching, lecturing, and writing about AI and the law and is published to promote my practice. Not legal advice. Not Direct/Public Access. All instructions via clerks at Doughty Street Chambers. This legal article concerns UK Cases of AI Hallucinated/False Citations. Subscribe to the AI Law Commentary here.
Introduction
I have been reading some interesting AI judgments that people continue to send me and I look forward to writing them up and sharing my thoughts. Please keep them coming.
Before doing that, I wanted to update the UK position of cases involving AI-hallucinated or false citations, either confirmed or suspected, to see where we are. My previous post on judicial AI hallucination cases attracted a great deal of attention, which suggests I am not the only person still oddly fascinated by this issue.
For newer readers, and there seem to be more of you every week, which I am hugely grateful for, I also keep a full UK tracker with the names and details of the cases, together with graphs showing where these incidents are appearing most often across the courts. I hope you enjoy reading that alongside the other trackers in my 15-tracker AI Law trackers here. If you can think of any other AI law issue I should be tracking, do let me know.
Just as I was preparing to circulate this article, the Solicitors Disciplinary Tribunal published its decision in SRA v Kumar. It is the first case before the Tribunal concerning a lawyer’s use of AI in legal proceedings. The decision is too important to deal with fully as a late addition to this round-up, so I have included a short account below with a link and will return to it in a dedicated article in due course.
SRA v Kumar
Case No. 12884-2026
25 August 2026
The Solicitors Disciplinary Tribunal considered submissions and correspondence containing non-existent cases and inaccurate or unsupported legal propositions arising from the Respondent’s admitted use of generative AI. He used AI again after the initial errors had been identified. Dishonesty was not alleged. The Tribunal was also dealing with a separate criminal conviction, but made clear that it would have imposed the same sanction had the AI allegation stood alone. It struck him off the Register of Foreign Lawyers.
Its central warning was: “A lawyer is always responsible for the accuracy of his or her output.” The Tribunal further observed:
28.12 Mistakes happen. Not every mistake will be a matter serious enough to engage the Regulator or this Tribunal. In his response to the Rule 14 Statement, the Respondent had categorised his wrongdoing as “technological negligence”. His insistence, in his Rebuttal Statement of 28 July and in oral submissions, that he ‘mistakenly believed’ that AI was a “lawful drafting aid” reflected a persistent lack of insight into his misconduct. He blamed having been out of legal practice and being, “unaware of recent
regulatory warnings or procedural directives regarding AI research tools”.
The Tribunal found that the Respondent had embarked on the use of AI without giving any proper thought to the consequences of so doing. He had failed to take proper steps to ensure his work output was accurate or to satisfy himself that the electronic tools he was seeking to use were reliable.Having done so, the Respondent’s culpability was very high because it was repeated after errors had been pointed out and explained. Rather than take active steps to rectify his mistakes, the Respondent repeated them. The Tribunal gave very substantial weight to the repeated nature of the Respondent’s misconduct.”
This first decision merits much fuller consideration, which I will provide in a separate article shortly.
Habib v SOSHD
UI-2026-000882 & UI-2026-000883, Upper Tribunal (IAC)
12 June 2026
The UTIAC dismissed the appellants’ challenge to the First-tier Tribunal’s refusal of their entry-clearance appeals, finding no material error of law. However, Deputy Upper Tribunal Judge Sternberg added a significant postscript concerning the appellants’ written submissions. The Grounds of Appeal and Supplementary Skeleton Argument contained a series of inaccurate or misleading authorities: cases were cited for propositions they did not support, citations led to entirely different cases, purported quotations did not appear in the authorities relied upon and some cases appeared not to exist at all. The Tribunal observed:
“50. This unhappy sequence of incorrect and misleading citations of case law and factually incorrect assertions in the Grounds of Appeal and Amended Skeleton Argument caused me substantial concern about the way in which both documents had been drafted and presented to the FTT (to obtain permission to appeal) and to this Tribunal on the appeal. It seems to me that an Artificial Intelligence Large Language Model may well have been used to draft or amend both documents.
51. I have not set out the names of the cases which appear to me to be likely to be hallucinations because this decision will be published online on the Tribunals Decisions website. The inclusion of cases which are, or may be, fake may perpetuate the loop of such cases being cited by AI if this decision is scraped by a Large Language Model.”
The Tribunal then considered whether to make a wasted-costs order or refer the matter to a Hamid judge, but first gave the solicitors and counsel an opportunity to explain what had happened:
53. Accordingly, I indicated at the close of the hearing that I was considering the question of wasted costs of my own volition and referral to a Hamid judge to consider whether either the solicitors and counsel acting for the Appellants should be referred to their respective regulators, independently of my conclusions on the substantive grounds of appeal.
54. The decision of Dove J and UTJ Lindsely in MS (Professional Conduct; AI Generated Documents) Bangladesh [2025] UKUT 305 (IAC) explains the likely consequences of the misuse of large language models. I bear that decision closely in mind in taking this course.
55. Before making any decision on wasted costs or on referral to a Hamid Judge I will give the Appellants’ solicitors and counsel an opportunity to respond in writing, if appropriate in the form of a document supported by a statement of truth. I require an explanation as to:
a. How the Grounds of Appeal and Supplementary Skeleton Argument came to be drafted;
b. What use if any was made of Artificial Intelligence to do so;
c. What checks were made of the citations in both documents before they were submitted to the FTT and the UT;
d. What supervision was in place by the person drafting those documents;
e. What checks [A’s Counsel] made of the citations in those documents in preparing his submissions for this appeal; and
f. Why [A’s Counsel] did not draw to the Tribunal’s attention the issues with the citation of authority in both documents until prompted to do so in answer to the Tribunal’s questions.
56. I now set those directions in the Notice to Show Cause below.
No wasted-costs order or referral was made in this decision. The Tribunal instead gave directions requiring written explanations before deciding whether any further action should follow.
Lodhia v Twelve Trees (Bromley-By-Bow)
24 July 2026
In this judgment, the High Court examined the claimant’s conduct in a discontinued defamation claim. The judge found that documents relied upon in support of an application to postpone a hearing had been fabricated or doctored and that witness statements contained false evidence. The claimant attributed the false material to AI, maintaining that it had generated the false documents without being instructed to produce false evidence. The full judgment sets out the evidence and findings in considerable detail and readers wanting the specifics can read it at the link above.
Mallon v Steer Energy Solutions Limited
15 June 2024
This is a slightly older case, dating from 2024, but I have decided to include it because there are relatively few reported UK cases from that period dealing expressly with the use of generative AI in litigation. In this case, the Employment Tribunal was reconsidering a costs order in a disability discrimination claim arising from a recruitment process. AI did not feature in the usual way. However, the Tribunal directly criticised the claimant’s use of AI-generated material, linked it to unreliable submissions and unnecessary cost:
“48. Around 15 emails followed over the next month, covering 200 pages in the reconsideration bundle, which were often difficult to follow, relying on several different AI generators of submissions about reasonableness, costs awards and disability discrimination generally and enclosing various attachments. They appeared to demonstrate an ability to create significant work for others when focussed on a goal of reducing the claimant’s financial exposure”
49. Whilst the claimant may believe that such an approach helps him, as he merely types in a question and then forwards the responses, the fact is those submissions simply generate a lot of work and confusion, and are often highly unreliable, literally creating non-existent case law and precedent.
50. Nonetheless, within them, it was able to be ascertained that the claimant was seeking a reconsideration on the grounds previously intimated, and it seemed to be in the interests of justice to allow him to put his case, not least in an effort to accommodate his disabilities.”
Mboungueng v Severn Trent Water Limited
30 September 2025
Here, the Employment Tribunal was considering the claimant’s unsuccessful application to set aside the dismissal of part of his discrimination claim following non-compliance with an unless order. The more notable aspect of Employment Judge Camp’s reasons, however, concerned the authorities relied upon by the claimant. The Judge identified a series of case citations that either appeared not to exist or did not support the propositions for which they had been cited and said he strongly suspected that AI had been used to prepare the submissions:
“29. The Claimant’s position may be that he did not use AI to help produce his application and submission. If he did not then what he did is, potentially, if anything, more blameworthy than if he did.
29.1 If he did use AI, then he evidently did not check the accuracy of what AI produced and was therefore reckless as to whether the information he gave the Tribunal was true and as to whether he was in danger of misleading the Tribunal.
29.2 If he did not use AI, then I struggle to think of any explanation for what he wrote other than that he simply made things up. This does not just apply to the non-existent case [case name redacted]. With the possible exception of [case name redacted], it is not credible that anyone could read reports of the other cases the Claimant has referred to and genuinely think they say what the Claimant has submitted they do.”
Employment Judge Camp further observed:
30. Either way, there was a risk that the Tribunal would be misled. Moreover, either way the Tribunal – I – have had to spend a great deal of time trying to find a nonexistent case in order to satisfy myself that it definitely does not exist and even more time reading (or re-reading) from beginning to end the other cases the Claimant has referred to in order to make sure that there was not something in them, tucked away somewhere, that I was not previously aware of and that supports the Claimant’s submissions about them. It is not usually necessary for the Tribunal to do this, because normally people refer to cases reasonably accurately. If they refer to one I am familiar with, I don’t need to look it up in order to know that the reference is accurate (for example, the Claimant’s references to [case names redacted] in his email of 18 August 2025). If they refer to one I am not familiar with, I can look it up and quickly see that it says what it is alleged to say.
31. The total time I have spent on this point is literally hours. It is a waste of my time and means I have less time to deal with the cases of other Tribunal users, which is not fair to them.
32. I have given some thought to what I should do about all this. What I have decided to do is not to make any orders on my own initiative. The main reason for my decision is that I think for me to take action at this stage – such as requiring the Claimant to explain himself and/or show cause why his claim should not be struck out for unreasonable conduct – would in practice not achieve anything other than unnecessarily to take up more of the time and resources of the Tribunal and the Respondent and to disrupt preparation for the DRA and final hearing, which has already been disrupted significantly.
33. I should make clear to the Claimant, however, that if in the future he makes further seriously inaccurate references to caselaw, or does anything else of a similar kind, this will be treated by the Tribunal with the utmost seriousness and could result in his whole case being struck out for scandalous and/or unreasonable conduct, in accordance with rule 38 of the Employment Tribunal Procedure Rules 2024.”
Chevalier-Hernandez v Poplar Dock Management Company Limited
LON/00BG/LSC/2025/0917
First-tier Tribunal, Property Chamber (Residential Property)
24 February 2026, amended 9 March 2026
The First-tier Tribunal was considering a leaseholder’s challenge to charges for pest control, carpet cleaning and an arrears fee. The more notable issue, for this article, was the applicant’s reliance on a purported authority that neither the Tribunal nor opposing counsel could locate. He withdrew the point and suggested that he might have found the case through Google or ChatGPT. The Tribunal did not determine that AI had generated the authority, but gave the following warning:
61. [applicant] accepted that he did not know how the case had come about but that he thought he had found it via Google or possibly another search engine such as Chat GPT.
62. The tribunal is concerned about the proliferation of AI generated “fake” legal authorities and reminds all parties, including litigants in person, of their duty to check that any legal authorities upon which they rely are genuine. Litigants in person can check on: https://onlinelibrary.london.ac.uk/resources/databases/bailii
Lockesfield Management Company Limited v Leasehold and Freehold Owners
First-tier Tribunal, Property Chamber (Residential Property)
LON/00BG/LVL/2024/0601
1 December 2025
The management company sought to vary 42 unenfranchised leases because the existing 1/91 service-charge formula no longer allowed recovery of the full expenditure after some leases had merged with the freehold. The Tribunal granted the variation, retrospectively. The Tribunal noted that two self-represented respondents relied on authorities which it considered irrelevant:
“19. At the commencement of the hearing, the Tribunal raised the issue that [redacted names] were seeking to rely upon legal authorities which appeared to have no relevance to these proceedings, and which did not support the propositions in respect of which they had been cited. [redacted] stated that he had obtained independent legal advice to the effect that the case law relied upon was not relevant and that statutory provisions had been cited which did not exist.
20. [redacted] confirmed that an internet search had been undertaken as part of her preparation for the hearing which could well have included in the use of generative artificial intelligence, and she accepted the Tribunal’s preliminary view that these legal authorities were not applicable. For these reasons, the legal authorities were disregarded. However, the Tribunal accepts that [redacted] was doing her best as a self-representing litigant who was unaware of the risks of “AI hallucinations” and makes no criticism of her.”
Brook House (Henley) RTM Company Ltd v Henley Brook Ltd
CAM/38UD/LRM/2025/0008
First-tier Tribunal, Property Chamber (Residential Property)
13 April 2026
The First-tier Tribunal held that the tenants’ RTM company was entitled to acquire the right to manage Brook House, rejecting the freeholder’s arguments that adjoining car-parking/access land defeated the claim or pushed the non-residential element above the statutory 50% threshold. AI featured when Counsel for the applicant argued that:
31…it appeared the Respondent had relied upon a large language model to prepare both the counter-notice and the statement of case. He noted that the Respondent had cited legislative provisions inaccurately, including [redacted]… Mr Castle submitted that these errors were indicative not of a litigant in person misunderstanding the law, but of material generated by artificial intelligence, confidently stating entirely untrue facts which could not reasonably be taken from the wording of the authorities named.
Counsel raised further issues:
“32…He argued that this was not merely a misreading or paraphrasing but further evidence that the Respondent had adopted unreviewed AI-generated content. This may not itself be conduct in the proceedings, but may (together with the shifting positions taken before and after the time limit for application to the tribunal to determine entitlement to the right to manage) be relevant if conduct in the proceedings is found to be unreasonable.”
Further:
“33. Mr Castle submitted that a party who presents arguments produced by AI, without reviewing and assessing their correctness, behaves unreasonably. He noted that the Tribunal is designed to accommodate litigants in person, many of whom are capable of presenting coherent cases without resorting to “machine-generated nonsense”. The Respondent does not seem small; it appears they paid £3.2 million for the Property and the car park in 2018. The use of such material, he argued, wastes the Tribunal’s time and imposes unnecessary costs on the opposing party. It makes early determination more difficult.
34. The Tribunal notes these submissions. However, as the alleged use of AI and the consequences said to flow from it were not pleaded, and as the Respondent did not attend the hearing, Rule 13(6) prevents the Tribunal from making any order for costs at this stage. If the Applicant wishes to pursue such an order, it may make a formal application within 28 days of the date of this decision…”
Guardian Report (Judgment not located)
In this Guardian report, a senior judge is said to have raised the possibility that the Home Office relied on AI-hallucinated information when refusing an asylum claim, after a country policy and information note cited in the refusal decision could not be found. The judge reportedly said that the refusal letter bore hallmarks consistent with AI use and that the reference may have resulted from an AI hallucination. I have not located the judgment, so I cannot independently verify the report or add anything beyond it. I would be grateful if any readers could point me to the decision.
Comment
There are many interesting issues arising from the cases above. I will focus on three, before returning briefly to SRA v Kumar.
Firstly, long-time readers will not be surprised to learn that I was encouraged by the Upper Tribunal’s decision not to name the suspected fake cases, thereby reducing the risk that their inclusion might “…perpetuate the loop of such cases being cited by AI if this decision is scraped by a Large Language Model.” I have been discussing this issue for some time. I put the point slightly differently in my article for Counsel Magazine which was subsequently cited in the Law Gazette:
“Well-intentioned judges often cite hallucinated cases and their erroneous legal principles in full within official judgments to show the extent of the problem to those reading. However, judges may be inadvertently exacerbating the issue because those AI-generated inaccuracies are being integrated into the established legal canon indirectly.”
Readers and colleagues have, quite reasonably, pushed back on my view several times on this particular issue. Some have suggested that, while the problem is theoretically possible, it is unlikely to arise to any significant extent in practice. Others have pointed to the public importance of judges setting out hallucinated authorities in full, so that what has happened can be properly understood, scrutinised and analysed.
My own research and discussions with lawyers have nevertheless suggested that there is at least a practical dimension to the concern. One example was shared by Sally McLaren, Assistant Librarian at Inner Temple Library, on LinkedIn (click to see the screenshot):
“The Ghosts of Ayinde. This is something that Matthew Lee and Paul Magrath have talked about. Fake citations now firmly in the public record. I am fascinated by this pollution. It is worrying for all the obvious reasons (see screenshot). It underlines the need to verify once, twice, thrice, and thrice again, to make up nine! But it also highlights the need for some conventions quick smart. A retrospective redacting perhaps? Or an encyclopedia of stop words? Although both approaches would totally spoil it for explorers from the future wanting to write up this postmodernist pickle.”
I still think there may be merit in maintaining a public database of hallucinated legal materials. That could allow judges to record that a false authority had been relied upon without necessarily reproducing the fictitious citation and legal proposition in full. A judgment might instead refer, for example, to “false case no xxx”, with a hyperlink to a separate database recording the material and explaining why it is false. That would preserve transparency and allow the problem to be studied, while reducing the risk of fabricated material being absorbed back into the information environment on which future AI systems may rely.
The idea has not, so far, gained much traction and I appreciate it may bring its own challenges. Any such database would need unmistakable human-readable and machine-readable labelling, together with a link to the judicial finding, so that the material could be studied without itself being mistaken for legal authority.
Secondly, I thought Employment Judge Camp’s observation that, if AI had not been used to produce the Claimant’s submissions, the position might actually be more blameworthy was particularly apt. The point, in brief, was that if AI was not responsible for the material, the judge struggled to identify any obvious explanation for what had been written other than that it had simply been made up.
That interests me for two reasons. First, it is striking that the use of AI, which might have been regarded as an aggravating feature, can, in this context, provide the less serious explanation for how fictitious material came to appear in a legal submission. Secondly, it raises a rather different concern. As awareness of AI hallucinations becomes more widespread, there is at least a risk that AI may sometimes be invoked as an explanation for material that was in fact invented by the person responsible. Courts may therefore increasingly have to distinguish between genuine AI hallucination and the rather older problem of human fabrication.
These cases may be an early warning of a much larger evidential problem. When I created and launched my International AI Deepfake Database and Tracker on 22 February 2026, which has since attracted considerable interest, I wrote that, serious though hallucinations are, “they do not come close to the evidential challenges posed by deepfake material”.
False citations can often be exposed by checking authoritative legal databases. Convincing AI-generated or manipulated documents, images, audio or video may be much harder to identify and authenticate. Deepfake evidence may therefore prove to be one of the most significant issues the profession will have to confront. I have a lot to update on my tracker and will be writing about this issue soon.
Finally, SRA v Kumar appears to be a significant development. The decision was published only as I was finishing this article, so I have confined myself to the brief account above rather than attempting a fuller analysis. I will look at the Tribunal’s reasons more closely and return to the case in a separate article.
If you have found it helpful, you may wish to subscribe to my AI law articles via the Substack below. Please do continue sending me decisions that you think should be tracked, whatever the jurisdiction. One of the clearest lessons from this work is that lawyers, judges and court users in different systems are often confronting closely related problems. Sharing decisions, experiences and reflections across practice areas and across jurisdictions remains one of the most constructive ways to respond.
Final Words from Popular LLMs
This is the point at which I invite an LLM to provide an independent comment on the preceding article. At times the observations are slightly off the mark, at other times they are perceptive, and occasionally they produce something entirely unexpected. Whatever the outcome, I find it a useful and revealing exercise to see what emerges.
ChatGPT (OpenAI)
“These cases show that AI-generated false citations are not confined to any one court, tribunal or type of participant, and that the central issue is responsibility for material put forward as genuine. Whether AI use is admitted, suspected or denied, judges increasingly expect authorities to be checked before they are relied upon. The first Solicitors Disciplinary Tribunal case concerning a lawyer’s use of AI in legal proceedings now demonstrates the possible professional consequences: repeated reliance on unverified AI-generated material was treated as sufficiently serious to justify strike-off even though dishonesty was not alleged. As generative AI becomes more widely used, the decisive questions will remain who adopted the material, what checks were undertaken and whether its provenance can be established.”
Gemini (Google)
“Reaching 76 recorded UK instances marks an unmistakable shift: AI-generated and fabricated citations can no longer be dismissed as teething trouble or isolated mishaps. Two evolving judicial developments in this milestone case are particularly striking. First is the judicial refusal to name suspected ‘ghost’ citations in published judgments—an active measure to prevent legal databases from poisoning the future training data of Large Language Models. Second, and perhaps most significant for professional ethics, is the court’s warning that fabricating authority without AI may carry even greater culpability than negligent machine reliance. Whether an error originates in algorithmic hallucination or deliberate human drafting, the duty to verify remains strictly non-delegable. The responsibility to safeguard the integrity of legal record rests squarely where it always has: on human oversight.”
Claude (Anthropic)
“What strikes me across these seven is how quickly AI hallucination has gone from being the live issue to being procedural background. Habib is the clearest example: the tribunal doesn’t need to establish that an LLM produced the fake citations before moving to sanction. Your database idea is the right response to the transparency gap that creates, but it only works if it’s built so a future model cannot cite it back as a real source in turn, robots-excluded, watermarked as negative data, that kind of thing.
