Thirty one UK Cases of Hallucinated Citations, AI and Otherwise: “It is not ‘co-operating with the Tribunal’ to neither confirm nor deny the use of AI”

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. Subscribe here.

Update

Thank you to everyone who has shared this post. Readers have sent me further hallucination cases, taking the total beyond 31. These have been added to the tracker below and the update to this post can be read here.

Introduction

This post records the thirty-first case in the UK I have added to my ongoing list of decisions involving AI hallucination cases, whether generated by AI or arising through other forms of fabrication or error. If you would like to read my analysis of the previous thirty cases, you can click on the case that interests you from the table below, which will take you directly to my analysis.

Elden v HMRC

[2026] UKFTT 41 (TC)

The case arose from an application by HMRC to strike out an appeal against closure notices. The factual and procedural background is set out fully in the decision, but will focus on the issues relevant to AI and inaccurate citations from paragraph 59 onwards:

“On 18 September HMRC emailed the Tribunal, the Appellant and his Representative raising concerns about the use of AI in the Appellant’s skeleton argument, and the inaccurate citing of authorities within it. The Tribunal responded that these matters should be brought to the attention of the Tribunal at the hearing itself.”

The Appellant explained that he had not contributed to the preparation of the skeleton argument. HMRC believed AI had been heavily used. When this was put to the representatives, the Appellant explained that:

“73…he did not contribute to the preparation of the Skeleton Argument. HMRC believe AI was heavily used to produce the document. This was put to the Appellant’s Representatives by the Appellant, and they neither confirmed nor denied the use of AI, but he stated they stood by everything in the document save where the Appellant himself contradicted it.”

The Tribunal then set out the Appellant’s submissions and the cases relied on. From paragraph 89, the Tribunal addressed summaries of five cases included in the skeleton argument. These summaries contained no direct quotations. HMRC contended that the summaries were inaccurate and produced with the assistance of AI. The Appellant maintained that he had not drafted the document. Through the Appellant at the hearing:

“92 …the representative neither confirmed nor denied the use of AI, and said they stood by the summaries and it didn’t matter whether AI was used or not.”

In further submissions, the representative argued that:

‘”93…The suggestion that citing a published authority amounts to providing false material is misconceived. A court decision is a matter of public record. Whether a case applies is a matter of legal argument and opinion, not misrepresentation. It is entirely proper for parties to put forward different interpretations for the Tribunal to consider. To characterise this as “false material” is both unfounded and inappropriate.’ It is not clear who the representative is quoting as saying false material was used. The wording used by HMRC was ‘inaccurate use of AI/inaccurate authorities”

The Tribunal responded:

“94…To some extent, the Tribunal agrees that whether or not AI was used is not directly relevant. AI is a powerful tool that can be used to great effect, but the human who relies on its use bears the responsibility for the accuracy. At the same time, because AI is known to ‘hallucinate’, that is, to generate false or inaccurate information and present it as if it were factual, if AI has been used to produce a document and flaws are found in that document, particularly if the flaws, once pointed out, are not corrected, this leads to the rest of the document being treated with great caution. This then has a knock on effect on the time taken to consider and check all relevant points.“

The Tribunal noted that the use of AI in legal proceedings is a rapidly developing area of case law and referred to earlier decisions, including Harber, Zzaman and Ayinde v London Borough of Haringey (link to these in table below). The Judge explained that it was therefore appropriate to test the accuracy of the summaries in order to determine whether the duty to verify AI output had been met.

Having analysed each cited case, the Judge overserved: “I am confident that no human being competent to summarise case law could summarise this case…” The Tribunal found that the summaries had been produced using AI and had not been verified with the degree of care required for tribunal submissions. This was said to raise serious implications for the administration of justice and public confidence. Tribunal time, as a public resource, had been significantly wasted by the need to analyse authorities that were either irrelevant or inaccurately summarised.

Following circulation of a draft decision, the Tribunal entered into correspondence with the Appellant’s representatives. The Tribunal, upon reviewing the draft judgment, asked the advisers to explain why the summaries in the skeleton were inaccurate. Only at that point was the use of AI acknowledged (it helped with conciseness) and it was accepted that the case references needed clarification.

A subsequent reply sought to justify the relevance of two of the cases but did not directly address why the summaries were inaccurate, beyond referring to time pressures. The Tribunal noted that this did not meaningfully engage with the concerns identified. A further explanation was later provided, attributing the errors to late instruction, urgent refocusing of the skeleton argument, human editing and document control mistakes. The Tribunal set out these explanations in detail but remained concerned.

The Tribunal also considered the Appellant’s conduct of the proceedings more broadly. While noting that he had largely left matters to his advisers, the Tribunal observed that he had been copied into most correspondence and could not have been unaware of missed deadlines. He appeared not to have appreciated the central importance of a witness statement in advance of the hearing.

In its findings of fact, the Tribunal’s view was:

“139. The case summaries were produced using AI and they have not been verified for accuracy with sufficient care as should be used when producing submissions for a Tribunal hearing. This lack of sufficient care amounts to professional incompetence on the part of any regulated individual or firm involved in the production of the skeleton argument.”

The Tribunal also identified:

“ (3) Obfuscation in relation to answering the question about the use of AI in the production of the skeleton argument. It is not ‘co-operating with the Tribunal’ to neither confirm nor deny the use of AI. A straightforward answer detailing the use of anything that could amount to the use of AI would not be difficult.”

Despite these findings, the Tribunal declined to strike out the appeal at that stage. It took into account the absence of a hearing date, the Appellant’s attendance at the hearing, and his stated intention to comply with future directions. The Tribunal instead imposed a series of Unless Orders.

These included requirements for a witness statement within a specified timeframe, confirmation of receipt of the hearing bundle, and strict conditions governing any future skeleton argument. Those conditions required the provision of full judgments, direct quotations with precise references, explanations of relevance, and statements of truth identifying who had checked each factual assertion or case summary and on what basis.

The strike-out application was therefore refused, although the Tribunal made clear that non-compliance with the directions would result in automatic strike out without further reference to the parties. For completeness, the Tribunal noted that by the time the decision was published, the witness statement had already been produced.

Comment

This case demonstrates with particular clarity the importance of being open with the court about how documents are prepared. The Tribunal was explicit that AI itself is not the issue. The difficulty arises where its use is obscured, its output is not properly verified, and inaccuracies are defended rather than corrected. Here, the Tribunal stopped short of striking out the appeal, but only after making serious findings and imposing robust safeguards.

As regular readers will notice, I have again refrained from setting out a detailed analysis of the false citations and summaries contained in the judgment. That decision reflects an ongoing concern of mine about inadvertently worsening the very problem these AI hallucination cases expose. As I have written elsewhere, well intentioned judges sometimes set out hallucinated cases and their false legal principles in full in order to demonstrate the scale of the difficulty. The risk is that doing so may unintentionally compound the problem, because AI generated inaccuracies can then become indirectly integrated into the wider legal record.

I have discussed this a few times now with colleagues both on this blog and in webinars. Some interesting observations were also made on LinkedIn on this issue. Not everyone agrees on the correct approach. Much of the discussion centred on the tension between transparency and protection. On one view, there is a strong instinct to spell out fabricated authorities and their purported principles in full, both to explain precisely what has gone wrong and to ensure accountability. On the other, there is an equally pressing concern about safeguarding the integrity of the body of law itself.

For now, I remain concerned. I am aware that large language models access and draw from my legal writing, and for that reason I continue to keep my full AI hallucinations database private. I do so because I am concerned that, if it were publicly searchable, search engines and AI tools would begin to republish false information and confer unintended legitimacy through repetition. If any readers wish to see this research, they are welcome to contact my clerks.

I have included below a section from my global hallucination tracker (AI and non‑AI hallucinated citations) that focuses solely on the UK position. I would be very grateful if readers who are aware of further cases not yet listed could share them with me, so that I can keep the tracker as accurate and up to date as possible. There is more analysis to follow shortly. If you find these articles helpful, please consider subscribing below.

Final Words from Chat GPT 5.2 Pro and Gemini 3 Pro

This is where I invite premium AI models to comment on or critique the preceding discussion. The below are the interesting aspects of each response.

ChatGPT:

“What stands out is how the Tribunal ties together candour, verification, and confidence in the process: AI can be a useful drafting tool, but the person putting the document forward carries responsibility for accuracy, and “neither confirm nor deny” responses about AI use only deepen suspicion when inaccuracies appear. Your further concern about whether judgments should reproduce fabricated cases captures the core dilemma in AI hallucination cases: transparency and accountability push towards spelling errors out in full, but doing so risks laundering false “law” into the searchable ecosystem that both humans and models draw from”

Gemini:

“The ruling in Elden v HMRC underscores a critical juncture in legal practice where the convenience of generative tools collides with the absolute imperative of professional candour. The Tribunal’s sharp criticism of the representatives’ initial refusal to confirm the origin of their submissions reveals that the core issue is not just the technology itself, but the attempt to evade responsibility for verifying its output. As AI hallucination cases continue to accumulate, they expose a systemic vulnerability: when legal professionals fail to act as a rigorous filter for automated content, they do not merely present weak arguments—they actively waste judicial resources and erode the foundational trust of the adversarial process.“

Tracker

The full international AI and non‑AI fabricated/false citations tracker can be here, the below is just a summary of the UK AI and non‑AI fabricated/false citations. The section below provides only a summary of the UK cases involving AI and non AI fabricated or false citations. You can click on each case name to read my detailed discussion of it on this blog, and from there you can use the hyperlinks to obtain the full judgment. The table will update automatically as the main tracker is expanded, even if the surrounding blog text remains unchanged over time.