On Monday 17 August, the SRA issued a warning notice on the use of AI in law firms, following a number of high-profile AI-related missteps at firms including Pinsent Masons and Sullivan & Cromwell.
The SRA highlighted two principal concerns: hallucinations – which it defines as AI-generated ‘fictitious cases, references or seemingly factual assertions that may appear convincing despite having no basis in fact’ – and confidentiality, particularly the risk of confidential client information being entered into less secure AI tools such as non-enterprise versions of ChatGPT and Claude.
It also warned partners and senior lawyers that they could face regulatory action if they fail to adequately supervise junior lawyers or properly review their work.
Tom Sharpe, a partner in Bristows‘ AI practice, welcomed the release of the guidance, saying: ‘The rapid rise in AI use across the legal profession means a clear steer from the regulator is genuinely welcome.’
However, another AI partner questioned why it had taken so long for the notice to come out: ‘ChatGPT was launched in November 2022, and everyone in the summer of ’23 was freaking out about putting stuff in tools, and now we are over three years later. I’m a bit shocked that it has taken them so long.’
‘If you’re loose in how you treat data, there will be consequences’
Perhaps inevitably, given the time that has elapsed and the high-profile reputational damage caused by AI-related errors, the lawyers LB spoke to said much of the notice was already covered by their firms’ internal policies and governance.
‘The stuff that the SRA is giving guidance about isn’t particularly relevant to us as we have always had things like oversight and governance in place,’ one partner said.
They argued that many of the risks associated with AI are not new, but extensions of risks the profession has dealt with for years, particularly around cloud technology.
‘The risks you talk about with AI are the same risks we’ve always had with cloud technology,’ the partner continued. ‘People are talking about AI as though it’s a completely separate thing, but it’s not a new risk for us.’
Mark Deem, a Legal 500 leading partner for AI who left Wiggin to launch his AI-enabled boutique Fortiter in April this year, noted that he doesn’t like the term hallucinations. ‘For me, it anthropomorphises the technology and gives it this feeling that it is a being in its own right. That’s one of the real risks in this.’
Another partner was more receptive, describing the section on hallucinations as ‘all sensible stuff,’ adding: ‘It’s useful to lay it out.’
However, they were much more critical of the guidance on client confidentiality, taking issue with its reliance on a judgment – UK v Secretary of State for the Home Department [2026] UKUT 81 (IAC) – that, in their view, wrongly conflated open-source software, code that is freely available to use or modify, with using less secure tools that can lead to information being publicly available.
‘The bit on client confidentiality, I just gave a big sigh,’ the partner said. ‘It takes the decision of IAC, which is an absolute shambles of a decision, uses weird terminology and refers to “open source.” That judgment says that using an open-source tool is to put it in the public domain, which is just wrong.’
As a counterpoint, Deem argued that lawyers should not get bogged down in technical arguments. The fundamental point, he said, was that client data should not be used on non-enterprise versions of LLMs: ‘If you’re loose in how you treat that data, there will be consequences.’
‘The ways it can go wrong are unfamiliar, so people’s spider senses don’t necessarily kick in’
Partners also disagreed over how concrete the guidance should be. One argued for more actionable guidance, such as a checklist for reviewing new technology and concrete examples of what the SRA considers acceptable and unacceptable.
Another had sympathy for the SRA, saying: ‘They are between a rock and a hard place, in that they can’t overregulate in such a fast-moving world.’
That balancing act is unlikely to get easier as the technology develops. Sharpe said: ‘AI technology is moving so fast that further guidance might be needed, particularly around increasingly autonomous and agentic AI systems.’
All interviewees were unanimous that partners were ultimately responsible for errors committed by junior colleagues, and that this is a problem that long predates the rise of AI.
‘At one level, it’s no different,’ one partner said. ‘What’s new is that the ways in which it can go wrong are unfamiliar, so people’s normal spider senses don’t necessarily kick in.’
Or as Sharpe put it: ‘Hallucinations are a bit like Jay from The Inbetweeners. They can tell you something with enormous confidence and it sounds plausible, but that doesn’t make it true.’
And while the SRA may be reflecting that it is impossible to please all of the people all of the time, one partner points out that regulators across the pond have even more to consider when devising regulations:
‘In the US, the state bars have started to issue guidance on AI, but it’s all a bit contradictory. What might be acceptable in one state could put you in breach of the rules in another.’
