If a trainee at your firm cites a case that does not exist, who faces the greater regulatory exposure — the trainee or the partner who signed off the work?
In November 2025, the Upper Tribunal answered that question directly. And the answer should concern every supervising solicitor in England and Wales.
In Part 1 of this series, we looked at client confidentiality, legal privilege and the three principles firms should apply when selecting an AI provider. This second part addresses a different but equally significant risk: AI hallucinations, the Tribunal's response to them, and the new obligations that now sit with anyone who supervises the work of others.
What Is an AI Hallucination?
AI hallucination is the term used to describe a situation where a large language model — the technology that powers tools like ChatGPT, Microsoft Copilot, Google Gemini and Claude — generates content that is factually incorrect but presented with apparent confidence and authority. In legal practice, this most commonly manifests as fabricated case citations — references to judgments that do not exist, attributed to courts that never decided them, complete with neutral citations that look entirely plausible.
This is not a bug that will be fixed in the next software update. These tools generate text by predicting the most probable next word in a sequence. They are optimised for fluency, not accuracy. If you have ever asked ChatGPT or Copilot to find a case and received an authoritative-sounding citation that turned out to be entirely invented, you have experienced this first hand. For a solicitor, the danger should be obvious: a fabricated citation misleads the court, wastes judicial time and exposes the firm to regulatory referral.
The Case Law: How We Got Here
The English courts and tribunals have addressed AI-generated fabrications with increasing severity over the past twelve months. Three decisions in particular trace the trajectory.
Ayinde v London Borough of Haringey [2025] EWHC 1383 (Admin)
On 6 June 2025, the Divisional Court handed down judgment in two joined cases referred to it under the court's Hamid jurisdiction.
In the first case, a barrister had drafted grounds for judicial review citing five authorities which did not exist. When the opposing solicitors raised this, the claimant's team dismissed the concerns as “cosmetic errors”. The barrister was referred to the Bar Standards Board; the supervising solicitor was referred to the SRA. In the second case, Al-Haroun v Qatar National Bank, a lay client had conducted legal research using AI. Of the 45 cases cited, 18 did not exist. The solicitor who filed the material without verification was referred to the SRA.
The court emphasised that a solicitor's professional responsibility for the accuracy of material filed with the court cannot be delegated — not to a junior colleague, not to a client and not to an AI tool.
MS v SSHD (Professional Conduct: AI Generated Documents) Bangladesh [2025] UKUT 305 (IAC)
In August 2025, the Upper Tribunal confronted the issue in the immigration context. A barrister had used ChatGPT to draft grounds of appeal and to verify a citation — which turned out to be fabricated. When challenged, he initially attempted to explain the errors away before eventually admitting to AI use.
The Upper Tribunal established that AI large language models can produce misinformation including fabricated judgments complete with false citations, that the consequence is likely to be referral to a professional regulator, and that taking unprofessional shortcuts which will very likely mislead the Tribunal is never excusable. The barrister was referred to his regulatory body. His lack of understanding of ChatGPT's limitations was not accepted as an excuse.
R (Munir) v SSHD (AI hallucinations; supervision; Hamid) [2026] UKUT 81 (IAC)
The Munir judgment, handed down by a three-judge Upper Tribunal panel in November 2025, is the most significant of the three — and the one with the broadest implications for law firms.
City Law Practice Solicitors and Advocates filed grounds for judicial review on behalf of Ms Munir. The grounds contained multiple bogus authorities: cases attributed to the wrong court, given wrong citations, and in one instance a citation that did not correspond to any existing case at all. The senior solicitor who signed the claim form — the firm's Compliance Officer for Legal Practice — stated that the grounds had been drafted by a junior caseworker working under his supervision. The COLP had signed the statement of truth on the UTIAC1 claim form. He was responsible for the accuracy of those grounds. He had not checked them.
Supervision, Not AI, Was the Issue
Judge Lindsley's analysis at paragraph 37 reframes the entire debate:
“It would be easy to think that this is a case about the naïve use of generative AI, but it is not merely about that; it is principally about supervision and the obligation to ensure that the tribunal is not misled. It matters not how such citation errors come about. Whether they are inserted by a hapless trainee or by ChatGPT is really neither here nor there; the point is that the qualified legal professional with conduct of the matter is expected to ensure that such documents are checked, that errors are identified, and that only accurate documents are sent to the tribunal.”
Why the Supervisor Carries the Greater Regulatory Exposure
The Tribunal went further at paragraph 38:
“In our judgement, a supervisor who fails to ensure that the work of a more junior fee-earner does not contain false cases or citations is likely to be more culpable than a lawyer who fails to ensure that his own work is free from such ‘hallucinations’. An individual in the latter camp fails the tribunal, the public and his lay client, whereas an individual in the former camp fails, in addition, to aid the development of more junior lawyers.”
Supervisory failure compounds the original error. A solicitor who personally relies on an AI-generated hallucination fails in one direction. A supervisor who allows that error to pass unchecked fails in two: they fail the court and client, and they fail the junior lawyer whose development depends on proper oversight.
The New Statement of Truth
In direct response, the Upper Tribunal has amended the UTIAC1 claim form to include a new statement of truth requirement. A legal representative must now confirm that any authority cited within the form or in any documents appended to it:
(a) exists,
(b) may be located using the citation provided, and
(c) supports the proposition of law for which it is cited.
A legal representative who signs such a statement in a case where false authorities are cited should ordinarily expect to be referred to their regulatory body. Other forms and directions are to be similarly amended.
While this particular amendment applies to the Upper Tribunal's immigration jurisdiction, the direction of travel is clear. Solicitors already owe duties to the court as officers of the court, and the SRA Code of Conduct requires that they do not mislead or allow the court to be misled. These new requirements simply make the consequences of breach explicit and procedurally automatic.
Free Download: AI Provider Evaluation Checklist
A practical checklist for UK law firms evaluating AI providers — covering model training, data retention, abuse monitoring and geographic data residency. Work through it with your prospective provider before signing any agreement.
The Practical Implications for Firms
The combined effect of Ayinde, MS (Bangladesh) and Munir creates a clear set of expectations.
Verification is non-negotiable. Every case citation in any document filed with a court or tribunal must be verified against a reputable source — BAILII, the National Archives' Find Case Law service, Westlaw, LexisNexis or an equivalent. The excuses ventured in these cases — illness, time pressure, lack of access to subscription databases — were all rejected.
AI policies must address supervision, not just usage. It is not sufficient to tell staff not to use ChatGPT. The problem is unchecked work. Firms need supervision workflows that catch errors at the point of review, irrespective of how those errors were introduced.
Supervisors must understand the risk. In Munir, the COLP described himself as not a “tech savvy person”. The Tribunal did not accept this as any explanation for his failure to put appropriate checks in place. A basic understanding of AI capabilities and limitations is now part of what it means to be a competent supervisor.
Record-keeping matters. The Tribunal was troubled by the firm's inability to demonstrate who had worked on which files. Poor record-keeping made it impossible to identify whether the same errors had occurred in other matters — and that uncertainty became an aggravating factor.
What This Means for AI Adoption
None of this is a reason to avoid AI. These cases are cautionary tales about the absence of supervision, not about the presence of technology. The firms that adopt AI well will be those that embed verification as a default — selecting tools purpose-built for regulated legal work rather than repurposing consumer AI products and hoping for the best.
In Part 1 we set out three principles for evaluating AI providers: no training on your data, no retention of messages and clear data residency commitments. To those three principles, Part 2 adds a fourth:
Principle Four — Select tools that are designed to support verification, not replace it. The AI tools you adopt should make it easier, not harder, for supervising solicitors to check the accuracy of output. That means tools with traceable sources, audit trails and the ability for a human reviewer to confirm, amend or reject any AI-generated content before it is acted upon.
How Chancery AI Reduces These Risks
Chancery AI has been designed with the lessons of Ayinde, MS (Bangladesh) and Munir built into the product — not bolted on after the fact.
Human-in-the-loop review. Chancery is built around the principle that AI assists the solicitor; it does not replace their judgment. A Supervising Solicitor must be assigned to every matter — enforced at the data model level, not left to process or goodwill — and the platform provides structured approval workflows that allow supervisors to review, amend or reject AI-generated work before it is acted upon.
Source and citation referencing. Where Chancery AI references legal authorities, it draws from verified legal sources rather than generating citations from prediction alone. When the platform cites a case or a statutory provision, the supervising solicitor can trace that reference back to its source, check it and confirm it before the output leaves the platform. The goal is to make verification straightforward rather than something that requires a separate manual research exercise.
Bespoke hallucination guards. Chancery includes purpose-built safeguards that assess AI-generated responses for accuracy before they are presented to the user, providing an additional layer of protection beyond what any general-purpose AI tool offers. This does not eliminate the need for human review — nothing does — but it materially reduces the risk that a fabricated citation or an inaccurate legal proposition reaches the solicitor's screen in the first place.
Each of these is a direct response to the obligations that firms already have. The platform is designed to make compliance easier, not to introduce new risks that require managing.
Summary
The Upper Tribunal has sent an unambiguous signal. AI hallucinations are not a technology problem — they are a supervision problem. For supervisors, the position is now clear: you carry greater regulatory exposure than the fee earner who produced the error.
Choosing the right AI tools — tools designed for regulated legal work, with verification workflows, source tracing and hallucination safeguards built in — is no longer just a technology decision. It is a regulatory one. Firms that invest in proper verification workflows and supervisory processes now will be far better placed than those that wait for the consequences to arrive at their door.
This blog post is part of our two-part mini-series, Using AI Safely in Legal Practice. Part 1 addresses client confidentiality, legal privilege and the three principles for evaluating AI providers.