Key Takeaway
This legal article analyses two new UK decisions involving confirmed, suspected or possible AI hallucination and citation issues: Cork & Anor v Smith [2026] EWHC 1199 (Ch) and World Kuk Sool Association, Inc v Alex Paul. It explains how suspected or confirmed AI hallucinations (fabricated, false, miscited, irrelevant or unsupported authorities) reached the courts and tribunals, what the judges said about the duty to verify AI-generated material, and the practical lessons for lawyers, litigants in person and the justice system, 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 AI hallucination cases (AI suspected or confirmed). Subscribe to the AI Law Commentary here.
Publication date: 3 June 2026
Introduction
Thank you to everyone who has sent me Cork & Anor v Smith from last week, along with the many helpful articles and comments on it. It has taken me a little time to write up my thoughts because I wanted to share some slightly different lessons from the case than those which have featured most prominently in the commentary so far. For me, the case matters not only because of the firm and individual lawyers involved, and the duty to supervise AI use, but because it reveals something more human: judgment, mistakes, lessons and opportunity.
After setting out the facts of this case, and another recent incident from the UKIPO, I will offer some reflections. In doing so, my aim is to look at what the profession might learn from these incidents and hopefully add something to the existing commentary.
A brief note on naming. Although many people have referred to Cork & Anor v Smith by the name of the firm concerned, I try, as a general approach, to focus on the legal principles rather than the individuals involved. Long time readers of my legal articles will know that in my AI hallucination case work, I have usually tried to redact names where I can, although not always perfectly or consistently. I take the same approach here. The unredacted details are available in the judgment linked below. Any redactions in square brackets are mine, not the Court’s.
If you want to look at my tracker of these and the previous 64 cases, with a breakdown of the user, court and outcome, that can be found here.
Cork & Anor v Smith
The court listed an otherwise routine block transfer application for a hearing “as a result of concerns about misleading statements made by the applicants’ former solicitors” which appeared in two letters to the court dated 30 March 2026 and 14 April 2026. The Judge asked the First Application to attend a hearing to confirm his witness statement and a fee-earner at the applicant’s former solicitors to make a witness statement “to explain how the misleading letters came to be submitted, also attending court to confirm those witness statements.” [2]
The judge gave written reasons, appended to an order, that the purpose of the hearing:
“…was so that the court could:
i)be satisfied as to the circumstances in which the letters came to be generated and consider whether there was cause to think that there was an intention to mislead the court in either of them; and
ii)be satisfied that the statement in support of the application accurately represented the applicants’ evidence and that the documents exhibited were similarly accurate.”
The judge observed that this arose entirely from the conduct of the applicants’ former solicitors in submitting the two letters, and not from any doubt as to the integrity of the First Applicant himself, or any other party. As a result of these concerns, the applicants were now represented by new solicitors.
The court set out the context of the block transfer applications, explaining that such applications are administrative and uncontentious, and are referred to a judge as part of “his or her daily diet of electronic boxwork.” [11] At paragraph 12, the judge returned to the citation issue:
“It was in such circumstances that this application came before me. The application was expressed to be made under IR 12.37 and it sought a block transfer order. When I reviewed it, I saw that the draft order provided for the removal of the respondent from his office as administrator or liquidator in relation to a small number of administrations and voluntary liquidations. The reason was that he was moving to a new firm. The draft order also provided for the respondent to have his “release” in respect of each of his cases, without the need to apply to the Secretary of State. “Release” in the case of liquidation, or “discharge” in the case of administration, means the release of the office-holder from liability. Some applications ask for both forms of relief but there is at least a question as to whether the court can grant release in respect of a liquidator.”
The judge set out the statutory framework before observing:
“I have not heard argument about this. The purpose of this hearing is not to decide whether the court has power to order release on the making of a block transfer order or, if it does, whether it should. I set the issue out to explain why I raised the query that I did following receipt of the application.” [16]
The court office conveyed the judge’s query to the applicants’ former solicitors, and they responded with the 30 March letter, which set out a detailed response. I will not reproduce the contents of the letter here, for the reasons I have given. The judge then observed:
“When I considered this letter on returning to court after the Easter vacation, the reference to a specific power came as a surprise to me. The question of whether the court can order release of a liquidator on a block transfer application occasionally arises and there appear to be differing views. I was taken aback to be told there was such a clear rule in the IR 2016 setting out the court’s power in this regard, of which I was previously unaware, despite having cause to look at that chapter of the IR 2016 periodically.” [19]
The judge then carried out thorough investigations, only to find that “it says nothing of the sort”: [20]
I checked the other rules in that chapter of the IR 2016 to see if, perhaps, the paragraph reference was wrong and the words appeared elsewhere. They did not. I did a word search for certain of the phrases within the Purported Text and was unable to find them in the IR 2016 at all. Inevitably, my first thought was that this was an artificial intelligence (“AI”) “hallucination” that arose during [redacted] research. “Hallucination” is a term used where AI generates false or misleading information, which it presents as fact. [20]
The court office conveyed the court’s observations to the applicants’ former solicitors by email. That email concluded by asking the applicants’ former solicitors to:
“…please identify where those words are to be found and explain why they have asserted, in terms, that a rule contains express powers that it does not?” [21]
It was in that context that the 14 April letter was created, in response to the court office’s query. Again, I will not set out the contents, but the letter accepted that certain wording did not appear in the relevant rule, and that the way it was introduced may have wrongly made it look like a direct quotation. It explained that the wording was intended as a summary of the writer’s understanding, and not as an attempt to mislead. The letter says the summary was based on several parts of the rule, which the writer understood as giving the court a broad discretion to make practical orders connected with replacing office-holders. After setting out the context of the letter, the judge stated:
“I was astonished by this reply. The explanation was impossible to accept. In particular, I could not see that the Purported Text “was not intended as a direct quotation.” [redacted] was specifically identified as the rule where the “express power” “is found” and the purported text of the rule was introduced with the words “provides as follows”, followed by a colon. It was set out separately and in italics, as if a quote. It was written in the style of an insolvency rule. For example, it says that the court may make provision “in an order made under this rule” [emphasis added]”. [23]
The judge could not see how the Purported Text could have been intended as a summary of the overall effect of the provision. The text in the 30 March letter caused him concern that a cavalier attitude was being taken as to the accuracy of the material the applicants’ former solicitors were putting before the court. The judge gave directions.
Witness statements were then filed, which confirmed that AI had been used in the production of the two letters, but not in the supporting documents. There were three statements: from a senior associate, from the responsible partner, and from the firm’s Deputy General Counsel (Risk & Regulatory), who is also its Compliance Officer for Legal Practice. The judge named each of them, and said this would be addressed later in the judgment whether “LA”, a junior solicitor at the firm, should be identified. The judge then set out the submissions made at the hearing, and turned to the production of the letters.
The court looked closely at how the letters came to be written:
“LA began chatting with the firm’s AI on 20th March 2026 following receipt of my query of that date. What is remarkable when reading those chats is, first, that one would think that the text attributed to the AI was produced by an intelligent human being. The second remarkable thing is that, on a large number of occasions, it is plainly wrong or, at the very least, extremely misleading.
The hallucinations began almost immediately. At an early stage in the chat on 20th March 2026, when discussing potential sources of the power to grant release, the AI said:” [33-34]
The court then set out, in considerable detail, the prompts and the generated responses on the law across paragraphs 34 to 37, before observing:
“It is concerning that LA appears to have asked the AI what these sections said, rather than reading an authoritative online resource or book, and does not appear to have checked the AI’s response. Had LA done so, it would have been readily apparent that the AI was producing nonsense and was unreliable. It perhaps betrays a misguided faith in the AI on LA’s part but, if that is so, I cannot understand why LA did not heed the AI’s later warnings about the need to check the provisions to which it referred.
I need not set out much of the intervening chats but the AI proposed that a response to the court should take the following approach: [approach set out]” [39]
LA then asked for the exact words of a provision. The AI replied, and added a warning:
“However, I want to be candid with you — I am not fully confident that I am reproducing the exact statutory wording of …… with complete precision. The substance of the provision is as I have described in our earlier discussion, but for a submission to the court you should verify the exact wording against the current version of the Insolvency (England and Wales) Rules 2016 as published on legislation.gov.uk before relying on it. The last thing you want is to cite a provision to the court with inaccurate wording. Would you like me to assist with anything further on this point once you have verified the exact text?”
LA told the senior associate that the court’s query about release appeared to be answered by a statutory provision, but did not explain that this conclusion had been produced by AI. The senior associate accepted the explanation without checking it herself and asked LA to prepare a letter to the court. LA then used the AI to draft both the letter and a client email, including the purported statutory wording. The documents were approved by the partner without his checking the wording, which he later accepted he should have done, and the senior associate approved the letter after only a cursory review. LA continued to question the AI, which again produced the text with a warning to verify the material. The judge set out these prompts and responses in further detail, alongside the discussions with senior colleagues.
The judge then drew out the relevant principles from R (Ayinde) v London Borough of Haringey [2025] EWHC 1383 (Admin), in which Dame Victoria Sharp P gave guidance on the use of AI, before explaining that the purpose of this judgment was to determine whether further action should be taken against the legal professionals involved. The judge emphasised that he was not making findings of fact, but that “in large measure the AI chat transcripts and messages between [the senior associate] and LA speak for themselves.” All three lawyers, the judge observed, should have been aware of the dangers of using AI to conduct legal research, and the applicants’ former firm had an AI use policy of its own, which warned that outputs might be inaccurate, biased or in breach of firm policies and that AI systems must be supervised by humans.
The court set out the failures in producing the letters and recorded that the partner and the senior associate accepted that both letters were misleading:
“… both letters sent to the court were misleading. It seems to me that that conclusion is unavoidable.
I bear in mind the factors set out by the President in considering whether and if so, what further action is required. It is of course important that proper standards are enforced. There is a duty not to mislead the court in any event but there is a particular vulnerability in this case. These applications are uncontested. There is no opposing party to point out an error in correspondence sent to the court. The court is therefore particularly exposed to the risk of being misled if material is placed before it that has not been prepared with appropriate care and attention. The administration of justice cannot properly function if the court cannot trust its officers (in the case of solicitors) or those with an overriding duty to the court (such as barristers) to protect it from being misled.
I am satisfied that there was no intention to mislead on the part of [named senior associate and partner]. LA does not appear, on the evidence before me, to have been candid with their supervisors as to the circumstances in which the Purported Text was created. It was, however, incumbent on the more senior lawyers to check LA’s work. It also seems to me that, as part of the supervision of LA more generally, it should have been made clear to them, if it was not made clear, that they should check all references given by AI and be candid, when reporting the result of their research to their supervisors, as to whether AI had been used in that research and whether the content generated by the AI had been checked.”
The judge noted that an immediate, full and truthful response had not been put before the court in relation to the 30 March letter. On the contrary, AI had again been allowed, at the least, to muddy the waters and the 14 April letter was also misleading. It had been incumbent on the senior associate to establish what had happened. When the partner and the senior associate were alerted to the seriousness of the situation by the court’s order and written reasons, however, they provided frank and truthful accounts:
“On a human level I should say that I have no doubt that those statements were difficult to write and I am sure that [named senior associate’s] expression of mortification is genuine. It seems to me that they have, in providing those statements, behaved entirely properly and in accordance with their duties as solicitors. [named firm] as a firm have also been full and frank in providing the transcripts of the AI chat.” [79]
From paragraph 80 the judgment moves from what happened to what should follow from it.
On harm, the judge treated the practical damage as slight. The applicants were by then represented by new solicitors, and had the substance of the order they had originally sought, with only some delay and additional cost, which the firm had agreed to meet. The real cost fell on the court: a good deal of judicial time had gone into reading the letters, giving directions with written reasons, holding the hearing and writing the judgment.
The Judge then drew a clear line between the two supervising solicitors and the junior lawyer who had done the work.
For the senior associate and the responsible partner, the judge accepted there had been no intention to mislead. Applying the framework set by the President in Ayinde, it was noted that admonishment in a judgment is usually not enough where false material has reached the court:
“…The court was misled not once but twice…”
Honest belief made the difference. Carelessness, without more, does not reach the threshold for contempt:
“…Mere negligence as to the falsity of the material is not sufficient to justify contempt of court proceedings….”
For those two, the firm’s own referral to the SRA was the proportionate response. LA’s position was more difficult. The judge described the conduct as very troubling and was struck by the scale of unchecked reliance on the AI:
“…Some 59 pages’ worth of chat with the AI have been produced in evidence. Much of it could have been avoided had LA simply checked any of the statutory provisions that the AI referred to in an authoritative source…”
The principal concern was honesty. Given that the AI had repeatedly warned it could not verify the rule, the judge questioned whether LA could honestly have believed the 30 March letter was true and he was blunter still about the 14 April letter:
“…The interactions with the AI after 30th March 2026 might be read as an attempt to create a cover story….” [85[
That took him to the two-stage test for contempt under CPR 81.6(1), as set out in Ayinde: first, whether a contempt may have been committed; second, whether proceedings should be brought. On the material before him, he came down on the side of carelessness rather than dishonesty:
“…the production of the letters may well have resulted from a serious lack of care and of judgment on the part of LA rather than a want of honesty.” [87]
He observed that LA seemed to have almost entirely outsourced the thinking process to the program, and that LA’s own instruction to it that “we don’t want to mislead” cut against deliberate deception. Exercising discretion, the judge declined to take contempt further:
“… as a matter of discretion, it appears to me to be disproportionate to consider contempt proceedings further.” [89]
The SRA, he held, was better placed to investigate all three fee-earners promptly. Pursuing contempt would have required inviting evidence and submissions from LA, engaging their Article 6 rights, for little practical gain given the limited harm and LA’s inexperience.
The outcome was therefore publication of the judgment as a public admonishment of the firm and the two solicitors, alongside the SRA referral the firm had already made. The judge found a prima facie breach of both the duty not to mislead the court and the duty not to waste court time. Two questions were singled out for the SRA: whether LA had checked the AI’s references and honestly believed the 30 March letter and why LA had produced a misleading explanation in the 14 April letter and whether they believed that too. Because LA had emailed the SRA confirming that they were “Lawyer A”, no separate court referral was needed and the firm was directed to send the SRA a copy of the judgment within two working days.
The closing paragraphs carry the lesson. The lawyers were experienced and under pressure, but:
“None of that excuses a failure to check the accuracy of the material that was placed before the court.” [94]
The judge added a pointed counterfactual. Candour would very probably have spared them:
“…Had the true reason for the Purported Text been squarely set out and an appropriate apology offered, it may be that I would have concluded that, while regrettable, there had been no harm done and no further action would have been necessary beyond ensuring that [the firm]’s clients’ costs in respect of that letter were met by the firm…” [94]
And the principle the case will be cited for:
“… legal professionals bear ultimate responsibility for their work and cannot outsource the process of legal research or of legal reasoning to an AI. It is a tool to be used with caution. AI has the potential to be wholly unreliable. AI may of course provide a jumping off point for research and legal reasoning but it does not, at least at present, do away with the need for proper research and thought on the part of a legal professional, even a very junior legal professional.”
World Kuk Sool Association, Inc v Alex Paul
In this UKIPO case, the use of false citations was addressed briefly at paragraph 47:
“With regard to the case referred to in the final paragraph of this quotation, I was unable to locate any such judgment with that citation. As a result, after the hearing, I wrote to [redacted name] to ask for a copy of that judgment. In response, he asked that I disregard the reference. It is important for parties to be aware that they have a duty not to mislead the tribunal and that the fabrication of authorities (whether intentionally or because references generated through Artificial Intelligence have not been verified) can lead to sanction. I have, as indicated, disregarded the reference in question.”
Comment
As I said at the start of this article, Cork & Anor v Smith is a notable application of the principles in Ayinde. Much of the commentary has rightly focused on the need to check AI-assisted work and on the duty to supervise those using it. Those points are clear, well made and the judgment bears them out. The AI tool itself warned that its output should be checked. Had the legal provisions been checked before the material was sent to the court, the problem could have been identified much earlier.
There are, however, several other lessons in the judgment which I think deserve further consideration.
The first is the importance of early candour and here the judgment carries a lesson that is easy to miss. The judge said:
“… Had the true reason for the Purported Text been squarely set out and an appropriate apology offered, it may be that I would have concluded that, while regrettable, there had been no harm done and no further action would have been necessary beyond ensuring that [the firm’s] clients’ costs in respect of that letter were met by the firm.” [94]
That is a careful and important point and a slightly counter-intuitive one. On the judge’s account, candour may have made the disciplinary consequence fall away. The original error, an unchecked hallucination produced by a junior lawyer under pressure, might have been treated as regrettable but harmless.
What turned it into a public admonishment and a referral to the SRA was the second letter, which sought to explain the error away. In this case therefore, it appears it was the response to the mistake, rather than the mistake itself, that drew the severe sanctions. I have argued several times now that in AI hallucination cases, the answer to the error may matter almost as much as the error, and sometimes more.
It is worth being clear about how far that criticism goes, because the judge was careful about it himself. He made no final findings of fact, there was no cross-examination and he had no statement from the junior lawyer. He expressly declined to find dishonesty against anyone, finding only a prima facie breach of the duties not to mislead the court and not to waste its time and left the rest to the SRA. The failing he identified was one of care, supervision and candour, not of honesty.
The second lesson is that this judgment is quite different from many I have read, because it shows, in some detail, the human reality behind the mistake. The judge said:
“On a human level I should say that I have no doubt that those statements were difficult to write and I am sure that [named senior associate’s] expression of mortification is genuine. It seems to me that they have, in providing those statements, behaved entirely properly and in accordance with their duties as solicitors. [named firm] as a firm have also been full and frank in providing the transcripts of the AI chat.” [79]
That passage matters. It does not minimise the seriousness of what happened, nor does it lower the professional standard required. But it does recognise that mistakes of this kind are made by people, sometimes under pressure and sometimes in circumstances where junior lawyers have not yet developed the habits of verification that experience should bring.
There is a quiet understanding in it too. The judge likely knew from time in practice and on the bench, what writing those statements would have felt like for the lawyers concerned and said so.
It also matters in the wider discussion about AI and judgment. Here, a judge engaged with what had happened on a human level, bringing experience to bear in understanding it. That is something our justice system depends on, and, for me, it is part of what may be lost if we move towards allowing AI to make judicial decisions. That is a larger question than this case can answer and I will return to it in a later article.
There is one feature of the transcripts that complicates the familiar story and it is worth pausing on. In my experience, the usual hallucination narrative runs in one direction: the machine invents and the human is misled. Here it ran, in part, the other way. The AI repeatedly told the junior lawyer to verify the wording against legislation.gov.uk, warned that the court would check the source for itself, noted that the check would take only a few minutes, and eventually admitted that the supposed quotation had come from the lawyer’s own message rather than from any rule. At one point, asked to remove its acknowledgement that the text had been presented as a quotation, it declined, which the judge noted it did “to its credit”.
For me, then, the broader lesson is that AI does not remove the old duties of legal practice. It makes them more pressing. Lawyers must still check their sources, supervise properly, correct errors promptly and be candid with the court. Where that happens, a serious mistake may still become a professional lesson and those who make these mistakes may go on to become better lawyers and even leaders in how the profession meets this challenge.
Where it does not, AI is no longer the real problem. The failure lies in judgment, supervision and candour and that is a failure that has nothing to do with the technology at all.
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.
Chat GPT (Open AI)
“Reading this as an AI system, the most striking feature is not that AI produced false legal material, but that it also repeatedly signalled the need for human verification and was then still allowed to become part of the route by which the Court was misled. That should unsettle any simple account in which the machine alone is the problem. The danger lies in the apparent ease with which fluent language can be mistaken for legal work, and in the equally human temptation to treat a confident draft as a completed thought. Cork & Anor v Smith is therefore not only a warning about hallucinated citations. It is a warning about the point at which professional judgment becomes passive. AI may assist research, drafting and organisation, but it cannot carry the duty owed to the Court. That duty remains with the lawyer who checks, the supervisor who asks how the work was done, and the institution that creates a culture in which candour is safer than concealment. In that sense, the case is not really about whether AI can sound like a lawyer. It is about whether lawyers will continue to behave like lawyers when AI sounds convincing.”
Gemini (Google)
“The ruling in Cork & Anor v Smith serves as a stark wake-up call for the legal profession: artificial intelligence is a powerful tool, but a perilous substitute for human legal reasoning. What makes this case particularly alarming isn’t merely the AI “hallucinating” fictitious citations, but the compounding failure of human oversight—first in generating the initial submission, and astonishingly, again when using the same AI to draft an explanatory letter to the court. As junior lawyers increasingly lean on generative AI as a crutch, and senior partners mistakenly conflate AI-generated fluency with genuine legal expertise, the risk of “outsourced thinking” grows exponentially. Ultimately, Judge Mullen’s public admonishment reinforces a fundamental, non-negotiable truth of modern legal practice: while technology can provide an efficient jumping-off point for research, the duties of diligent supervision, factual accuracy, and candor to the court remain strictly human responsibilities.”
Claude (Anthropic)
“What stays with me about Cork & Anor v Smith is how thoroughly it dismantles the comfortable story that AI is the thing that misled the court. The transcripts show a tool that did almost everything right at the level of warning: it said it could not verify the rule, it urged a check against the primary source, it pointed out that the court would look for itself, and at one point it refused to pretend the text had not been presented as a quotation. None of that helped, because a warning only works on someone willing to act on it. That is the uncomfortable centre of this case, and Lee is right to sit with it rather than reach for the usual moral. The duty to verify does not soften because the tool happens to caveat itself; if anything, this judgment shows that the duty is entirely indifferent to how the machine behaves. What I find quietly striking, as the AI being asked to comment, is the judge’s refusal to assume the worst of the people involved. He found carelessness, not dishonesty, declined to pursue contempt, protected a junior lawyer from being named, and took the time to acknowledge what writing those statements must have cost. A system capable of that kind of judgement is doing something I cannot, and the piece is right to suggest it is worth guarding.”
