Beware the siren call of open-source AI
In UK and R (on the application of Munir) v Secretary of State for the Home Department [2026] UKUT 81 (IAC) the Upper Tribunal has given important guidance on the use and misuse of AI tools in legal practice.
The facts
Two sets of proceedings were before the Tribunal. The second matter did not deal with AI misuse issues. The purpose of the first hearing was to decide whether to refer Mr Mohammed (a level 3 accredited adviser) of TMF Immigration Lawyers to the Immigration Advice Authority (“IAA”), previously the OISC, for investigation.
The Upper Tribunal (“UT”) made a ‘show cause’ order when considering the application for permission to appeal to the Upper Tribunal from the decision of the First-tier Tribunal on 18th September 2025. In the show cause notice, TMF Immigration Lawyers were required to provide a signed written notice within seven days identifying the person responsible for the case and the grounds, and to: “provide an explanation for the fact that the case Horleston v SSHD [2007] EWCA Civ 654 cited in the grounds of appeal to the Upper Tribunal is not available on BAILII, and explain why the citation is in fact for South Tyneside Metropolitan Borough Council v Anderson & Others (a case about equal pay for female support staff and of no possible relevance to sufficiency of protection in a protection law context).” TMF Immigration Lawyers were to state explicitly whether an AI large language model such as ChatGPT was used to draft the grounds or not.
Mr Mohmmed responded to this direction on the same day setting out that he was the person responsible for the case and drafting of the grounds of 17th March 2025 in this matter. He stated unequivocally that no AI large language model (such as ChatGPT) was used in drafting the grounds of appeal. Mr Mohamed subsequently filed a witness statement, dated 10th October 2025, in which he confirmed again that he drafted the grounds. He reiterated that he was certain that he had not used ChatGPT to create the grounds ,but accepted that after consideration of his internet browsing history he could not offer a cogent explanation based on his searches as to how this false case with another case’s citation was included in the grounds. In these circumstances he states: “In absence of an explanation and with how AI operates, I cannot dismiss the fact that the case was an AI creation as there is no other explanation.” He expressed his disappointment in himself and that he was aware of his professional duties as outlined in the IAA’s Code of Standards including the duty to conduct himself with honesty and integrity. He explained that he had difficult matters in his personal life which had placed him under a lot of distress and anxiety.
On 14th October 2025 Mr Mohammed reported himself to the IAA and to the SRA for having unknowingly inserted a non-existent case in grounds of appeal by email, enclosing a copy of his statement of the 10th October 2025 prepared for this hearing.
The UT said:
“Mr Mohammed has admitted including one fake case in grounds of appeal and concluded himself that it was probable that this had happened because of the inadvertent use of an AI large language model. As far as the Panel could establish there had never been a reported case called Horleston but Google AI nevertheless can, depending on the question posed, produce information such as panels of Court of Appeal judges who sat on this non-existent case and suggest that it had to do with the fairness and asylum applications. By asking Google AI the same question in slightly different ways, we were able to elicit various different compositions of the bench which decided that fictitious case. Plausibly, each of the judges suggested by Google was sitting in the Court of Appeal at that time but not one of them could have sat on a case of that name because there is no such case. The danger in using Artificial Intelligence for legal research is not confined to generative AI models such as ChatGPT, therefore; the use of Google AI for legal research is equally likely to generate results which are false but which might initially be thought to be accurate.
We do not suggest for a moment that the use of legal AI programmes by properly trained professionals is anything other than a step forward in legal practice. The software which is currently available is of enormous benefit in properly focused legal research, as it is in other contexts such as large disclosure exercises. But any practitioner who uses non-specialist AI to undertake research or drafting is obliged to undertake rigorous checks to ensure that any information gleaned from those sources is true and accurate. Anyone with responsibility for legal practice at a firm of solicitors or regulated legal advisers must be aware of those pitfalls and of the need to warn staff about the dangers of using non-specialist AI.”
The Tribunal concluded:
“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, and any such conduct might itself warrant referral to the regulatory body and should, in any event, be referred to the Information Commissioner’s Office.”
Conclusion
The UT was careful to make no criticism of the use of AI in the daily work of a lawyer; rather the judgement concerns itself with ‘open-source AI’. The decision is a warning bell for over-enthusiastic use of AI by hard-pressed legal practitioners and makes clear that the Tribunals and Courts are likely to come down hard on the misuse of AI tools by practitioners in future.