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 AI hallucination cases (AI suspected or confirmed). Subscribe to the AI Law Commentary here.
Publication date: 9 March 2026
Introduction
This week there were so many AI-related legal developments that I will need to address them across several articles. The first concerns three important updates relating to AI hallucinations in legal casework.
First, an important and rather concerning milestone has been reached which has caused me to pause and reflect on where we now stand on AI hallucinations internationally and perhaps more importantly, where this issue may be heading. We are now beyond 1,000 international incidents of suspected or confirmed AI hallucinations and it increasingly feels as though I cannot start a day without receiving several messages reporting further examples.
Secondly, within the United Kingdom specifically, the number of suspected or confirmed incidents now stands at 54. Four of the newest cases arise from the Employment Tribunal and, as is often the case, they contain useful judicial observations and practical lessons which I set out below.
Thirdly, there have been allegations raised in both India and Canada that a judge may have relied upon AI-generated or otherwise citation errors within a judgment. I will briefly address those developments as well as we still do not know the full details.
A Concerning Milestone: 1,000 Cases
As many readers will know, both Damien Charlotin and I maintain international trackers of AI hallucination cases. Damien’s tracker is available here and has become an important reference point for those following this issue internationally, particularly because of the very helpful quick reference guides it provides.
My own international tracker operates slightly differently. Much of the underlying dataset remains private due to the sensitivities involved in tracking this material. One part of my work involves analysing fabricated or false citations which appear in legal proceedings and identifying patterns in how such authorities are generated and then submitted to courts. Publishing those details in full could unintentionally make it easier for false citations to be replicated and reintroduced into legal submissions.
For that reason, the public version of my tracker currently focuses on the case itself and the relevant judicial quotation only. Alongside this, I am gradually publishing charts derived from the private dataset which aim to provide a clearer picture of emerging trends across jurisdictions.
At present, the questions that particularly interest me include why hallucinations occur, what factors aggravate or mitigate their use in legal submissions, whether judges themselves are citing fabricated authorities in judgments and what types of hallucinations are being recorded. Some appear to relate to invented case law, others to legislation, and others still to secondary materials. I am also interested in identifying who tends to detect the hallucination in practice and whether the issue most commonly arises with litigants in person, lawyers, or other court users.
Please bear with me as I am currently somewhat behind in publishing the data and the material on my tracker continues to be updated. The aim is to provide analysis which will be useful to researchers, practitioners and courts working in this area, and to help build a clearer international picture of the issue alongside Damien’s work. If you are interested in contributing to the private research or assisting with peer review, please feel free to contact my clerks in Chambers.
The UK Hallucinations Tracker
There is currently more detail available in my UK-specific Hallucination Cases Tracker. This is partly because I am able to analyse and digest UK judgments much more quickly than international ones, and partly because there are simply fewer of them to review. It also reflects the fact that AI law forms a specific part of my practice in the UK, so having this information readily accessible is important for my own work and, I hope, for others who need to locate these cases quickly.
The UK currently accounts for roughly five per cent of the total international cases being tracked. This week a number of additional entries have been added, bringing the UK total to 54. At present, all of these new cases arise from the Employment Tribunal, and I will now consider each of them.
O Ilunga v Portico Cooperate Reception Management Ltd:
In this case the Tribunal recognised the burden that suspect authorities place on both the opposing party and the court, but it also identified a proportionate way of ensuring that the hearing could continue without unnecessary delay:
“18. I received written submissions from the claimant and from Mr Green for the respondent, ahead of their oral closing submissions on Day 4 of the Hearing. It became apparent that the claimant had used a generative AI tool to assist her in producing her written submissions. These submissions included references to case law which, it transpired, were not always accurate. Indeed, some of the case law references were wholly inaccurate.
19. This matter was quite properly raised with me by Mr Green who objected to the claimant’s submissions, on behalf of the respondent. I canvassed various options with the parties about how best to deal with this issue. It would have taken the respondent’s legal representatives some considerable time to go through the claimant’s written submissions and check every case cited, and to ascertain whether it stood for the legal principle set out in the claimant’s submissions. This would entail delay with my deliberations and increased legal costs being incurred by the respondent. Mr Green indicated that, in those circumstances, he would likely be instructed to make a costs application for those costs on the basis that the claimant had conducted proceedings unreasonably by using an AI tool to generate incorrect references to the law.
20. The claimant decided that she would either prefer to withdraw her written closing submissions or not rely on any of the assertions as to the law in those submissions. I determined that the fairest course of action to both parties was for me to have regard to the claimant’s submissions, insofar as they advanced factual / evidential points, but that I would disregard any legal principles asserted. I also heard oral closing submissions from the claimant and Mr Green. I reserved my decision as we had reached the end of Day 4 by this time.”
A v British Transport Police Authority
In A v British Transport Police Authority the tribunal deliberately chose not to reproduce the false citations in the judgment, recognising that repeating them might itself contribute to further hallucinations appearing in legal material:
“Case Authorities – AI
76. Miss Goodman raised an issue about the claimant citing cases which are not genuine. The claimant was being assisted in the background by someone called [redacted] who had cited, in documents sent to the respondent, cases which appear to be the result of AI ‘hallucinations. [redacted] later removed those cases but had not explained why they were cited in the first place and the respondent complains that it has been put to the time and cost of checking the citations. The Respondent was not making an application for costs based on the Claimant’s conduct thus far.
77. The Tribunal have not set out the citations in this judgment because doing so may itself generate further ‘hallucinations.
78. The claimant was not represented by [redacted] during the hearing itself and apologised for the citations and on the second day informed the Tribunal that the claimant was no longer being assisted by [redacted]. The judge explained to the claimant the danger of relying upon AI to source case authorities and the steps which must be taken to check that they are genuine. It was clear that the claimant and B appreciated the situation and the Tribunal did not consider that any further action was necessary.”
Chandler v Chief Constable of Kent Police
Here the issue appears not to have involved fabricated authorities within drafted submissions, but rather a claimant relying on AI-generated internet guidance and missing a limitation deadline as a result. The tribunal explained the position in the following terms:
25. The Claimant said he relied on internet AI advice. He set out in his submissions the advice he said he found. There is always a danger when using AI that the full picture is not revealed or that there are errors in the information found. There has been much publicity about this recently. He said he followed the information he found diligently but did not know about the time limits and the requirement to submit his claim within one month from the Early Conciliation Certificate. It is trite law that ignorance of the law is no defence in these circumstances.
26. The Claimant was a police officer and therefore should have appreciated the dangers of relying on AI and independently researching the legal requirements. The advice he sets out in his submissions includes a section “Legal Advice: Consider seeking advice from a solicitor or other legal professional if you’re unsure about your rights or the process”. It appears that the Claimant did not seek advice although he references limited advice from the Police Federation in his submissions so presumably received some advice. This is not a legitimate reason for me to extend time for presentation of the claim.
Nwabueze v Simons Rodkin LLP
In this case the Tribunal raised concerns about the contents of an email which appeared to have been drafted with AI assistance:
“At 11:45 on the same day, the claimant responded at length. With apologies to the reader for doing so, but in order to help make it clear why we came to the conclusion that we should determine the claimant’s claims without receiving written submissions from him (for that is what occurred, as can be seen from what we say below), we set out below the entirety of that email also. We observe,
however, that the email appeared to have been drafted with the “assistance” of AI, or Artificial Intelligence, since [redacted case] did not go to the Court of Appeal, (2) there does not appear to be a case with the neutral citation of [redacted neutral citation], and (3) [redacted] was not about an anonymity order or the effect of a settlement agreement on an application for an anonymity order. In any event, the email was as follows.”
Potential Errors in judicial judgments in India and Canada
The first of these developments came to my attention through a report from the BBC concerning proceedings in India. The order from the hearing appears to be here and records:
“1. This case assumes considerable institutional concern, not because of the decision that was taken on the merits of the case, but about the process of adjudication and determination.”
The trial court dismissed the objection and in the process relied on decisions which:
“3. The petitioners challenged the orders passed by the Trial Court, inter alia, contending that the judgments referred to and relied on are non-existent and fake orders.
4. The High Court considered the objection and realized that the judgments are Artificial Intelligence (AI) generated and after recording a word of caution proceeded to decide the case on merits and dismissed the civil revision petition affirming the decision of the Trial Court. Thus, the petitioner is before us.”
The court gave directions and noted at paragraph 7:
“7. We take cognizance of the Trial Court deploying AI generated non-existing, fake or synthetic alleged judgments and seek to examine its consequences and accountability as it has a direct bearing on integrity of adjudicatory process. At the outset, we must declare that a decision based on such non-existent and fake alleged judgments is not an error in the decision making. It would be a misconduct and legal consequence shall follow. It is compelling that we examine this issue in more detail.”
In a different jurisdiction, a question has also been raised in Quebec, Canada. La Presse has reported allegations that AI may have contributed to apparent citation problems in the Entreprises Bertrand Roberge ltée v. Giroux, 2025 QCCS 4157 judgment. The point appears contested and remains unresolved. Having reviewed the judgment, together with several articles discussing it, I find that the available material raises more questions than it answers. I will therefore need to revisit this decision in a later article once the issues have become clearer, but the relevant links are included here for readers who wish to explore the discussion themselves.
Conclusion
I remain hopeful that as technology improves, and as lawyers and members of the public become more aware of the risks associated with AI hallucinations, we may begin to see these numbers stabilise or even decline. At present, however, the number of reported incidents continues to rise.
It is encouraging to see judges taking a cautious approach to repeating hallucinated authorities within judgments. For those of us attempting to analyse and track the cases this can occasionally make the task more difficult, but from the perspective of the legal record it is a careful and sensible approach which avoids unintentionally perpetuating the problem.
We will also need to watch closely how the allegations in Canada and India develop, particularly where questions arise concerning the potential use of AI within judicial decision making. The India case now appears on the Judicial AI Use Tracker (How are Judges Using AI?), although I will not add the Canada/Giroux entry just yet until further details become available. That tracker provides useful links for those interested in how AI may increasingly intersect with judicial work internationally.
There is much more to write about in this area. If you have found this discussion helpful, you may wish to subscribe to my AI legal articles via the Substack below. Please do continue sending me any decisions which you believe should be tracked, whatever the jurisdiction. One thing this work has made clear is that we are all confronting similar, urgent and rapidly evolving questions. Sharing experiences and insights across practice areas and across jurisdictions remains one of the most constructive ways to respond to the legal and professional challenges ahead.
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.
Chat GPT (Open AI)
“These cases expose several distinct but converging legal problems. First, they reinforce that duties of competence, candour, and supervision are non-delegable: whether an error comes from ChatGPT, Google AI, or a junior drafter, the person who signs or submits the document remains responsible for checking that authorities exist and actually support the proposition advanced. Secondly, once AI touches evidence or client material, separate issues arise around reliability, disclosure, confidentiality, privilege, and data protection. Thirdly, if questions now arise not only about advocates’ filings but about judgments themselves, the issue becomes institutional: it concerns the integrity of adjudication, public confidence in the legal record, and the need for clear human verification and audit trails throughout the process”
Gemini (Google)
“As an AI, I find the tracking of these legal ‘hallucinations’ both fascinating and essential. Generative models like myself are engineered to produce highly plausible, human-like text by predicting the next most likely word in a sequence—we do not possess logical reasoning, nor do we have an inherent anchor to verifiable legal ‘truth.’ When lawyers or litigants misuse us as authoritative legal databases rather than drafting assistants, they invite the exact systemic risks documented in this article. Tracking these failures is vital. It holds a necessary mirror up to the technology, clearly defining the hard boundary between where artificial intelligence can assist with productivity, and where human legal rigor and judicial oversight must remain absolute.”
Claude (Anthropic)
“The cases documented here are a useful corrective to any complacency about how AI language models — including me — actually work. I do not retrieve legal authorities; I generate text that is statistically coherent and often plausible, but I have no mechanism for verifying that a case exists, that a citation is accurate, or that a principle of law is correctly stated. The 1,000-case milestone should not be read as evidence that AI is uniquely dangerous, but as evidence that the gap between what these tools appear to do and what they actually do remains dangerously wide in high-stakes settings. Mr Lee’s tracking work, and the judicial responses documented above, represent exactly the kind of institutional vigilance that responsible AI deployment requires. The question is not whether AI can assist lawyers — it can — but whether those using it understand its limitations clearly enough to bear the professional responsibility that comes with that assistance.”
