AI Hallucinations and Professional Liability: Further lessons, this time from UK v Secretary of State for the Home Department [2026] UKUT 81 (IAC)

Introduction

For the legal profession, Artificial intelligence has moved remarkably swiftly from novelty to utility within the past few years. Firms of solicitors are increasingly using, and indeed relying upon, large language models (‘LLMs’) and other generative AI-powered tools to assist with legal research, drafting and reviewing documents, disclosure, and case management. Used appropriately, these tools offer significant opportunities to improve efficiency and reduce costs. Used carelessly, they have the potential to generate no more than a professional negligence claim and a referral to the SRA or other legal regulatory body.

The dangers of over-reliance on such tools have been known for some time. The decision in R (Ayinde) v London Borough of Haringey [2025] EWHC 1383 (Admin) (‘Ayinde’) was a watershed moment, as the first reported case in the senior courts to confront directly the use by legal representatives of inaccurate generative AI in litigation. The Divisional Court made it abundantly clear that lawyers have a professional duty to check the accuracy of AI-generated research.

The decision in Ayinde prompted the Law Society and Bar Council to each issue guidance on the use of generative AI in legal practice. However, such guidance did not prevent the Defendant firm in Cork v Smith [2026] EWHC 1199 (Ch) from recently attracting the ire of the Insolvency Court for seeking to rely upon a non-existent Insolvency Rule, as recently discussed in an article on this website.  On 18^th May 2026 the Bar Standards Board issued further guidance for the Bar on the use of AI, see; Artificial-Intelligence-Guidance-May-2026.pdf.

Many larger firms have bespoke systems or systems adapted from AI specifically designed to assist legal practice.  Even then, over-reliance (or inadequately supervised reliance) on such systems is dangerous, as Cork v Smith proved.  But many smaller, less well resourced firms have resorted to using free online AI tools, with even greater risks.

The decision of the Upper Tribunal (Immigration and Asylum Chamber) in UK v Secretary of State for the Home Department (AI hallucinations; supervision; Hamid) [2026] UKUT 81 (IAC) (‘UK v SSHD’) is another, perhaps less well-publicised example of this growing phenomenon of legal practitioners relying on hallucinated AI-generated material. The key takeaways from its decision are relevant to all legal professionals. It also provides further valuable insight into the Courts’ likely future approach to professional negligence claims arising from the use of AI.

The Decision

The Upper Tribunal of the Immigration and Asylum Chamber considered, in one judgment, two separate cases in which false authorities had been advanced by legal practitioners acting on behalf of their clients in immigration proceedings. In both cases, concerns arose regarding the provenance of the authorities, the possibility of this being ‘hallucinated’ AI-generated material, and the adequacy of supervision exercised by those responsible for the litigation.

In the first case, an immigration firm had cited a non-existent authority, ‘Horleston v SSHD’, said to be a 2007 decision of the Court of Appeal. Upon this coming to the attention of the Tribunal, the firm was ordered by a show cause notice to identify the person responsible and provide an explanation why the case did not appear to exist. An immigration adviser was identified, who initially denied using AI, but later accepted that he could not rule out the possibility that the fictitious case had been generated through inadvertent use of Google’s AI-assisted search functionality. He also admitted to uploading confidential client correspondence into an online AI tool.

The second case concerned judicial review proceedings in which multiple inaccurate authorities had been cited in the pleadings. The claim form lodged in the Upper Tribunal had been signed by a ‘senior solicitor’, but the Grounds for Judicial Review simply bore the name of the firm. An Upper Tribunal Judge ordered that the Compliance Officer for Legal Practice at the firm should provide a signed statement identifying the author of the grounds. A ‘part-time trainee lawyer’ was identified as the drafter, who was said to have been working at the firm under supervision. The firm was ordered to show cause why its conduct should not be referred to the SRA; a hearing was listed to consider the firm’s response. The supervising solicitor argued that the errors arose from inadequate supervision and reliance upon outdated firm precedents, and claimed there was no mechanism by which staff at his firm could use AI tools.

In the first case, the Upper Tribunal noted that the danger in using AI for legal research was not confined to generative AI models such as ChatGPT; the use of Google AI (within Google search, which feature the immigration adviser and solicitor in question suggested he may have inadvertently used) was equally likely to generate false results which may at an initial glance be thought to be accurate. Ultimately, the Tribunal declined to make a regulatory referral as the adviser had already self-reported (to two bodies). However, it made clear that referrals would ordinarily be appropriate where false authorities are placed before the court as a result of inadequate corroboration procedures. Additionally, putting client letters into an open-source AI tool was to place this information on the internet in the public domain, and thus breach client confidentiality and waive legal privilege.

In the second case, the Upper Tribunal was unimpressed with the conduct and explanations offered by the supervising solicitor. It held that the critical issue was not merely whether AI had been used, but whether appropriate supervision and checking of the work done had taken place. The solicitor, as the supervising professional, was ultimately responsible for ensuring the accuracy of material filed with the Tribunal. A referral to the Solicitors Regulation Authority followed. The Tribunal was further surprised at the solicitor’s naivety regarding the availability of AI tools to his staff.

Three Key Takeaways

The Upper Tribunal drew several broader conclusions, of relevance to litigation before any Court or Tribunal:

1. Legal professionals are obliged to ensure that the legal arguments they present are factually and legally accurate. Those who cite false cases fail to comply with that professional obligation and risk referral to their regulator.
2. A solicitor or other legal professional who delegates their work to another fee-earner remains responsible for the supervision of their work and for ensuring its accuracy. They must ensure that those they supervise are aware of the dangers of using non-specialist AI for legal research and drafting. Failure to do so is likely to result in a referral to the SRA. A supervisor who failed to ensure that the work of a more junior fee-earner did not contain false cases or citations was likely to be more culpable than the junior fee-earner themselves.
3. Uploading confidential documents into an open-source AI tool, such as ChatGPT, is to place this information on the internet in the public domain, and thus to breach client confidentiality and waive legal privilege, which might itself warrant referral to the SRA and should in any event be referred to the Information Commissioner’s Office.

Importance of the Decision

The Upper Tribunal decision in UK v SSHD adds to the series of cases, across several practice areas, in which judges have had to grapple with legal professionals presenting documents containing inaccurate and misleading recitals of statutory provisions, procedural rules, and authorities, arising from the ‘hallucinations’ of generative AI-based LLMs. This is by now, sadly, well-trodden territory.

But perhaps the most important takeaway is that while the Upper Tribunal decision arose thanks to the misuse of AI, fundamentally the cases concerned straightforward considerations of legal professional liability. It mattered little whether the errors in question originated from ChatGPT, Google AI, a junior fee-earner or some other source; the key issue was whether the qualified legal professional responsible for the matter had exercised appropriate supervision and corroboration.

This approach aligns comfortably with orthodox professional negligence principles. There appears to be no need for the courts to re-consider existing duties and standards of care merely because AI was involved. The question remains whether the reasonably competent solicitor or barrister (or other type of lawyer) would have acted in the same way in the circumstances.

The same would hold true for other professionals such as surveyors, accountants and others who may make use of AI tools.  The tools are just that; it is the work product which is judged, though of course methodology remains important.  A lawyer or other professional is only likely to be non-negligent if the product of AI is such that it is not possible to verify its accuracy independently and (to use the Bolam test) it is reasonable for the professional to trust the AI to provide that product.  As far as the authors are aware, there is currently no AI system which could safely be left to generate work on its own and the output of which could not be verified by a human (but that day of ‘black box legal AI’ might yet come!).

As such, many future AI-related negligence claims are likely to focus on familiar issues:

• Inadequate supervision of junior staff;
• Failure to corroborate legal authorities;
• Reliance upon unreliable research methods;
• Deficient document review processes;
• Inadequate internal governance procedures; and
• Failures to protect confidential client information.

The Tribunal’s observations regarding supervision are particularly notable. It stated that a supervisor who permits a junior fee-earner’s work to contain hallucinated authorities may be more culpable than a lawyer who introduces such errors into their own work. This is a crucial reminder of the importance for firms of maintaining adequate systems for supervising delegated work. Reference should also be made to the recent case of Mazur v Charles Russell Speechly [2026] EWCA Civ 369[1], on the question of carrying on the conduct of litigation and supervision, of course.

Confidentiality presents a further area of potential exposure. The Upper Tribunal’s observations regarding open-source AI systems are stark. Uploading confidential client documents to publicly available AI tools risks breaching client confidentiality and waiving privilege. This area appears fertile for future disputes. Questions are likely to arise concerning the adequacy of client consent, firms’ AI governance and information security policies, compliance with data protection legislation, and insurers’ expectations regarding AI use.  That said, it is not currently clear how loss could be caused from such confidentiality breaches, as the data presumably ends up on a Google (or similar) server somewhere in the world but will be of little interest to those with access to it.

Significantly, the Tribunal distinguished between open-source AI products and closed systems, referring specifically to Microsoft Copilot as an example of a tool capable of performing certain functions without creating the same risks identified in relation to public platforms. That distinction is likely to become increasingly important as professional standards evolve.

The Next Wave of Litigation

Judicial criticism following a ‘show cause’ order in respect of AI hallucinations, regulatory referrals and findings of professional misconduct are likely to provide fertile ground for subsequent negligence claims by dissatisfied clients.

But looking further ahead, hallucinated authorities are unlikely to represent the most significant source of AI-related professional liability. The more substantial risks may arise from subtler failures.

What happens when AI-assisted drafting introduces a critical factual inaccuracy into a pleading? What if an AI-generated witness statement contains fabricated details which go unnoticed? What if confidential transaction documents are uploaded to an unsuitable platform without appropriate safeguards and the platform is hacked? What if a firm adopts AI systems without implementing appropriate governance, training and auditing procedures?

Each scenario gives rise to conventional negligence questions framed by this developing new technology.

The additional risk of…not using AI?

However, one of the most interesting questions is whether professional standards will ultimately move in the opposite direction. The current debate is focused on the risks posed by the use of AI. Yet the Tribunal itself recognised that properly designed and appropriately deployed legal AI tools may be “a step forward in legal practice” and of “enormous benefit”.  The senior judiciary, and the Master of the Rolls Sir Geoffrey Vos in particular, continue to see the responsible and proper use of AI as the way forward (see his recent further speech in Scotland in April 2026, the transcript of which is here: Speech by the Master of the Rolls: Artificial Intelligence and the Judiciary – Courts and Tribunals Judiciary).

As AI systems become more reliable, it is conceivable that the standard of care will likewise evolve. If validated AI tools become demonstrably better at identifying authorities, detecting inconsistencies or reviewing disclosure than human lawyers acting alone, future claimants may argue that a reasonably competent solicitor should have used them.

The professional negligence claim of the future may therefore be less likely to involve the allegation that the solicitor or counsel negligently used AI, but rather that they negligently failed to use it.

Conclusion

The decision in UK v SSHD provides yet another indication as to the approach of the court system to this burgeoning area of professional liability involving AI-assisted legal work. It reinforces the principle that legal representatives remain personally responsible for everything placed before the court or tribunal, regardless of whether the immediate source of the error is a trainee, a precedent bank or an AI system. AI-related errors are not really a novel category of wrongdoing. Existing principles of competence, supervision, verification and confidentiality remain the governing framework.

But the decision also hints at a future in which the professional standard of care is shaped by technological development. As AI tools mature, the question may become whether it is negligent not to use AI.

[1] In which one of the authors, James Hall, was junior counsel for an intervener, the Law Centres Network, led by Gatehouse’s PJ Kirby KC.

Article by James Hall and Mark Erridge