On 6 June 2025, the Divisional Court gave judgment in R (Ayinde) v. London Borough of Haringay [2025] EWHC 1383 (Admin). This was the first case that Court has considered concerning the misuse of artificial intelligence by lawyers. The President of the King’s Bench Division issued a wake-up call to the profession.
At paragraph 9 of the judgment she referred to the “serious implications for the administration of justice and public confidence in the justice system” if AI is misused. She said that “practical and effective measures must now be taken” by those with leadership responsibilities and by regulators. She made clear that: “Those measures must ensure that every individual currently providing legal services within this jurisdiction…understands and complies with their professional and ethical obligations and their duties to the court if using artificial intelligence.” The Appendix to the judgment contains a selection (it stresses that there are “many more examples”) of cases here and around the world which have considered the misuse of AI in the provision of legal services or to assist in the conduct of litigation.
Paragraph 6 of the judgment sets out the known risks of using Large Language Models (such as ChatGPT). These risks are easily discoverable by an ordinary internet search. In addition, at paragraph 87 of the Appendix, the Court criticises the advice given by the First Tier Tribunal in the recent case of Zzamen v. HMRC [2025] UKFTT 00539 (TC). The Divisional Court said: “We do not, however, consider that the risks are materially reduced by “asking the tool not to provide an answer if it is not sure and asking the tool for information on the shortcomings of the case being advanced.””. Again, this criticism is based on shortcomings of LLMs that are well documented and easily discoverable: LLMs have a tendency to “double down” on false information. All these features of LLMs are set out in the Bar Council’s Guidance on AI of early 2024.
The examples set out in the judgment suggest that a proportion of the profession is treating AI technology as being in a sort of special category that does not require the same approach to its adoption and use as other machines or electronic tools. The reasons for that approach are many and extend beyond the scope of this blogpost, but the approach is no longer appropriate. The ordinary approach to the adoption of a new machine, electronic tool or computer software is to ask (at least compendiously) some combination of the following questions: what is it designed to do? What does it actually do? How good is it at doing what it does? And, how safe is it?
There is no reason why an AI based tool should be different in this respect and good reasons why it should be treated just the same. Three reasons will suffice.
First, because most of our experience with using computers to solve problems is with binary systems that always do precisely what they are asked to do and do it accurately (for example, a calculator or a ‘key word’ search) there is a strong human bias to think that all outputs from computers are both more accurate and better than a human output. That is simply not the case when it comes to the output of the different technologies in many AI systems.
Secondly, almost all our experience with using words to convey a question and receiving words in response, is in the context of human-to-human interaction. In nearly all human-to-human interactions, words produced in response to a question are the product of human thought or emotion. Again, there is a strong human bias to think that when a question is put into a computer and a response in words that apparently make sense comes back, a human like process has produced that response. While we don’t fully understand how AI tools such as LLMs get from their inputs to their outputs, that is not what is occurring at least at this stage in the development of the technology. All the AI is doing is producing a series of output words that it associates with the input words through a form of probabilistic computer analysis of its training data (generally an enormous swathe of the internet). There is no human-like regard for truth or accuracy in the computerised process. The Bar Council’s Guidance explains this by reference to the affidavit sworn by the US lawyer who cited fictitious cases in the US case of Avianca (which is also in the Divisional Court’s Appendix).
Third, even those who design and market (for example) LLM tools cannot say precisely how they get from the input text to the output text. In fact, there is a long article on the Anthropic AI website which seeks to try and discover how its model works in this respect by a sort of reverse engineering. This last feature makes the use of AI tools particularly problematic in the administration of justice. An essential feature of the justice system is transparent and interrogable reasoning which the public can assess and have confidence in and individual litigants can challenge or appeal. Most currently available and potentially relevant AI tools cannot provide this.
AI is here to stay. However, if its potential advantages are to be harnessed, each tool that uses AI must be properly assessed on its merits before being deployed in the provision of legal services. It may be time to establish an appropriately qualified (and properly independent of technology firms) assessment mechanism to explain each AI tool that may be used by the legal profession and to assess the capacities and risks of each such tool rigorously. This will assist both the profession and the judiciary to decide which tools should be used and for what purposes and with which guardrails.