AI and Professional Negligence
Use of AI and professional negligence has been sought to be clarified by the UK Jurisdiction Taskforce (UKJT) in their Statement on Liability for AI Harms. See the report by LegalIT insider which states:
The UK Jurisdiction Taskforce (UKJT) – an industry-led initiative created by LawtechUK – has this week published a Legal Statement on Liability for AI Harms, which finds that professionals may be liable for negligent use of AI and also for failing to use it when a competent member of their profession would have done so. The Taskforce also finds that English law is already capable of resolving most AI liability disputes without the need for AI-specific legislation; and that businesses cannot avoid legal responsibility simply because a chatbot produced false information – the developers of general-purpose AI models will not usually be liable for unforeseeable downstream harms.
The Telegraph, reporting on this, said:
Any lawyers who were found to have acted negligently by failing to use AI could face regulatory punishment, the guidance said. Such punishment could include being banned from practising as a solicitor or barrister.
The UKJT said lawyers would be expected to use AI in the same way that any professional would if doing so would significantly improve the outcomes of their work.
The actual words used in the Legal Statement by the UKJT are:
A professional could be liable for failing to use AI in circumstances where a competent member of their profession would have done so.
The Statement later elaborates as follows:
It is important to be aware of the possibility that a professional could also be liable for failing to use AI for a task when a professional exercising reasonable care and skill would have done so. Whether there is a breach of duty in failing to use AI will, of course, be judged according to whether a reasonable professional of a comparable rank/specialism should have used AI in that context. That will, again, turn upon professional bodies’ regulations and/or guidance (if any) and on the expert evidence as to the actions of competent professionals in the field. The possibility of breaches of duty arising from a failure to use AI reflects the fact that AI, in a professional’s hands, is a tool. The question of whether such a tool should be used, and, if so, how, is no different from that which arises in respect of any other tool available to a professional.
Sir Geoffrey Vos, Master of the Rolls and Head of Civil Justice in England and Wales, said:
Generative AI is now used by a vast number of individuals, businesses and professional services providers. Since the advent of generative AI in 2022, lawyers and other users of AI have expressed concern about when and whether Large Language Model developers and others involved in the AI supply chain might be liable for harms caused by AI. This Legal Statement attempts to provide much needed legal certainty and legal predictability in an area where few cases have yet reached the courts. I believe the Legal Statement will be of inestimable value to the legal community, the technological community, and users in a time of rapid development in both the capability and usage of generative AI.
It probably does the opposite of what Vos hopes it does.
Alex Heshmaty, commenting on LinkedIn, said:
At first glance, this appears to place lawyers between a rock and hard place: if they use AI and it produces hallucinations which they miss, the lawyer is negligent, and if they don’t use it and make a mistake they are also negligent. Either way, the very existence of AI appears to be opening up lawyers to more claims of negligent liability, slowing down legal work and increasing costs, closing off more avenues for access to justice. Would be interested to know your thoughts Brian Inkster.
I replied:
👀 I may have to blog about this!
Now I am.
Parker
I recall Lord Keen of Elie, speaking in his capacity as member of the LawTech Delivery Panel (an initiative sponsored by the Ministry of Justice and the Law Society of England & Wales) in 2018, referencing Parker as good legal tech. It was awful.
I blogged at the time:
I imagine he [Lord Keen] has never used it. A very simplistic chatbot that asks a few questions to determine whether you need to comply with GDPR and when you inevitably do simply states “Would you like to speak to my team members about what you need to do?” Hardly cutting edge or very useful legal tech.
The problem is that the Lords, Knights and British Empire medal holders on these panels have no real idea about the technology that they are pontificating about. They don’t work in private legal practice. They don’t know what works or doesn’t and why.
Barristers
To be fair, the Statement by the UKJT was prepared by a group of English barristers. But again they were looking at it from a strict liability perspective and, I wager, without much knowledge of the actual technology at play or how it might or might not be employed in legal practice (the work of a barrister being very different to that of a solicitor).
AI and Professional Negligence
The question of AI and professional negligence is not new. I recall at a Lexpo Conference in Amsterdam a speaker suggesting that it would not be long before it would be considered professionally negligent to not use AI in legal practice. That was before the advent of GenAI. I think it may have been 2017! I have heard this repeated on a number of occasions since.
Now we have a UK Jurisdiction Taskforce suggesting that it could now be. The UKJT states that it “is an industry-led initiative, tasked with promoting the use of English law and UK’s jurisdictions for technology and digital innovation”. So although it has UK in its name it only promotes English law. I therefore sincerely hope its musings bear no weight up here in Scotland!
Flawed conclusions?
However, where do you start and stop with this suggestion?
It is clear that, in most cases, the dangers of using AI (especially GenAI) in legal practice is greater than not doing so. Until the balance shifts (if it ever does) should that not be enough to dismiss any suggestion that not using AI could result in professional negligence?
Also when and how can you demonstrate use of AI that “a competent member of their profession would have done”?
Does “a competent member of their profession” have to:
- understand how AI works as good as any other member of their profession?
- have prompting skills as good as any other member of their profession?
- understand when and when not to use AI?
- understand the verification burden of using AI and the choice of not using it if that burden is too high?
- have access to the same technology as another competent member?
- be able to access and use every AI platform available in case one is better than another? For example if your law firm uses Harvey but Legora is better at performing a certain task than Harvey (or vice versa) are you negligent for not using the better platform for that particular task? What if Mike OSS is better than either of those? 😉
In the absence of considering any of these points the Statement by the UKJT is inherently flawed and I trust not one the courts will rely on when deciding professional negligence cases against lawyers.
Contradictions in AI and Professional Negligence
The Statement by the UKJT contradicts itself when it earlier sates that:
It seems likely that if Anita [a name used in the Statement to represent the lawyer] chooses to use AI, she should have a sufficient understanding of it. She should also be able to explain (at a minimum, in broad terms) how the AI she is intending to use works to her client. If Anita used an AI tool without having at least a sufficient understanding of the AI tool and what it was doing, it will in most cases follow that she did not use reasonable care and skill.
So is Anita then, in turn, negligent for not using an AI tool that she had no understanding of and could not explain to her client how it works?
More work for Lawyers
Again, as I have said before, the winners will be lawyers who specialise in professional negligence cases representing clients and/or lawyers and/or Professional Indemnity Insurers in the mess caused by AI use. We will need more of those lawyers to deal with AI and professional negligence.
By the way, Damien Charlotin has now (as at 12 July 2026) tracked 1,751 reported decisions globally in cases where generative AI produced hallucinated content. That is up from 1,394 on 4 May 2026. 357 more in 70 days. On average 5 new ones per day. Or 7 new ones per business day. Those statistics do not feature in the UKJT Statement. And they contradict the assertion by Sir Geoffrey Vos of “an area where few cases have yet reached the courts”. He may, of course, be referring to professional negligence cases arising from those decisions. However, those decisions lay the problem bare.
Reactions to AI and Professional Negligence
On LinkedIn the following comments have been made:-
Huw Evans (Company Director at The Advisers Toolbox LTD):
I agree with this. AI is a dangerous experimental technology and should be treated as such.
Not saying it shouldn’t be used, but it should only be used with caution and not relied upon.
Me:
Indeed. As Cory Doctorow put it: “AI is the asbestos we are shoveling into the walls of our society and our descendants will be digging it out for generations”: https://pluralistic.net/2025/09/27/econopocalypse/#subprime-intelligence
-+-+-+-+-+-
Deborah do Carmo (Lawyer-Linguist. Raising the bar in legal translation since 2002. Dutch & Portuguese to English. CIOL-certified. Former insolvency practitioner (SA), CPPI-qualified (Scotland). Triple national: 🇬🇧 🇿🇦 🇵🇹):
Excellent blog post, Brian.
So, basically, professional negligence would be judged under a Pied Piper test: once enough professionals follow the tune, everyone else risks being called negligent for not joining the procession.
The trouble is, we all know where that procession led.
Me:
Good analogy!
-+-+-+-+-+-
Ian Dodd (Consultant):
Until Legal AI in the UK is covered by legislation and regulation the work the Task Force does and your very cogent arguments are, sadly, going nowhere as there is no framework for any governance.
It’s all putting the cart before the horse.
Me:
Indeed. Whilst the Task Force considers that English law is already capable of resolving most AI liability disputes without the need for AI-specific legislation!
Ian Dodd:
The ICO and the SRA believe(I have it in writing) that GDPR is more than enough to manage all the problems Legal AI in the UK will generate.
I’ve asked them to tell me exactly how GDPR will do this.
I’m still waiting.
-+-+-+-+-+-
Graeme Johnston (mainly law-related):
I think “could” is carrying a lot of weight in the sentence from the summary of the report.
Later on, the authors’ fuller thought on the topic is as in this image – they envisage usual principles of prof neg being applied and the only solicitor example given is of someone who fails even to advise their client to use some form of AI-assisted tool for disclosure in the E&W B&PCs.
And to grasp what they in mind there it’s necessary to look up the endnote (n96), which suggests that they mean TAR in the sense already understood in edisclosure, and they support the proposition by reference to the existing obligation in the applicable practice direction to consider TAR, and to explain why it’s been rejected, if it is, in the disclosure review doc.
Given all this, and bearing in mind that TAR has now been around since the 2000s and has well-established benefits when applied with proper process, I think it’s a reasonable example but also by its narrowness illustrates that they’re not envisaging anything other than a situation where any competent solicitor would have used, or considered using, demonstrably helpful tech.
Me:
Indeed. I quoted that paragraph 67 in my blog post. The example given is indeed limited and portrays a lack of proper consideration of, as you say, the rather weighted “could”. However, that summary has been latched upon by the press (no doubt as part of a limited press release by the UKJT that quoted it – with the Telegraph embellishing upon it) and is phraseology that could have been put better. The lack of fuller consideration of the points I highlighted demonstrates, I believe, a lack of understanding on the part of the authors that undermines their findings. You give them more benefit of the doubt than I can!
Graeme Johnston:
Thanks – agreed! And yeah, the Telegraph thing about people being struck off for not using it was quite a leap, even bearing in mind the duty of competence!
-+-+-+-+-+-
Matthew Letts (Strategic Tech Advisor, Founder, Entrepreneur, Solicitor, General Counsel and Legal Quant – Helping lawyers avoid burnout on busy work and law firms make more profit – Building big things in 2026 | Legal 500 – Key lawyer):
A completely bald assertion: “It is clear that, in most cases, the dangers of using Al (especially GenAl) in legal practice is greater than not doing so.”
You don’t actually have the evidence to back this and, as demonstrated by the Q&A at a recent event where you spent 15 minutes saying AI isn’t fit for use in the profession, professionals are already using AI responsibly and are seeing the benefits of doing so.
A handful of examples of people being ignorant to how the technology works and submitting hallucinated cases doesn’t make the technology bad. It just makes training and responsible use important.
Me:
Here we go again! As stated to you before, the evidence (unlike your brief generalised statements) is set out in 29 very detailed blog posts that precede this latest one: https://thetimeblawg.com/chatbots/ – I suggest you read and learn. Oh, and 1,751 reported decisions globally (increasing on average by 7 per business day) in cases where generative AI produced hallucinated content is not “a handful”! Also, I never said at the GBLO that “AI isn’t fit for use in the profession”. I actually gave examples in the Q&A where it could be of use. You clearly were not listening.
Matthew Letts:
1,751 is a handful compared to the number of incidents of negligence in the same time period.
Using AI without understanding how it works, and not checking outputs of AI, is negligent.
I’m not going to waste my time reading all your blog posts when the basic premise upon which you speak is flawed. It’s just ideology on your part.
Me:
A handful is defined as a “small number or amount”.
The negligence cases are being built up from those 1,751 reported decisions and all the unreported ones (which likely will be countless more) taking place daily that will only come out of the woodwork as time goes on.
I agree that “using AI without understanding how it works, and not checking outputs of AI, is negligent”. That is what I have written about extensively! However, you would not know that if you choose not to read what I write.
As I have said to you before, and will say again, I note that you and Simon P MARSHALL provide AI training to lawyers. If you and/or he are not highlighting the problems and risks that I do, and consider those to be ideological and not real, then you will be exacerbating the problem that currently exists with AI use within legal practice. On your heads be it.
Matthew Letts – editing a previous comment made by him to include:
You mad [sic] a big point about radiology at GBLO which is wrong on the facts. You then said you’ve used AI but it’s never gotten you to an answer quicker as you’ve spent more time checking than you would have if you’d done it yourself without AI.
Me:
I see you are now editing your previous comments to include new points! Usually better to add those as new ones after the original points have been answered.
I made one point in one slide out of 27 about radiologists. That is correct on the facts. Radiologists in regular cars are still driving to work in 2026 despite Geoffrey Hinton predicting, in 2016, that they would be obsolete in 5 years and Elon Musk predicting in 2015 that we would have self-driving cars by 2017. Are you suggesting we no longer have any radiologists?
I did not say I’ve “used AI but it’s never gotten” me “to an answer quicker as” I’ve “spent more time checking than” I “would have if” I’d “done it” myself “without AI”. In fact in the last Q&A at GBLO, I gave an example of when that was not the case and GenAI was very useful to me.
You are now just making things up!
Matthew Letts:
the reason I’m critical of your fear mongering is because I understand the technology, risks, and opportunities.
Using this technology intelligently can transform the way lawyers work resulting in increased profitability, efficiency, and work-life balance.
A handful: “A very small number of things or people compared to a larger group.”
The larger group here is the total number of negligence cases.
People who listen to your talks will steer clear of AI and pay the price for that. People who listen to me will get to learn about the pros and cons of AI and can make an informed decision about usage and implementation.
No response has been edited after you have responded.
People will vote with their feet. They can vote for fear and doubt (your approach) or learning and careful deployment (my approach).
Me:
It is not “fear mongering”. I am explaining the risks against the opportunities. You minimise/ignore those risks. Lawyers do so at their peril.
My approach is not “fear and doubt”. It is full and detailed explanations (backed up with evidence) to give real understanding. More often than not, I am told how refreshing this is when it comes to discussions around GenAI.
I hope your approach is in fact the same as you, unfortunately, display none of that on here.
Matthew Letts:
I don’t ignore or minimise those risks at all. You haven’t had the pleasure of listening to my 15 minute talk. Nor have you seen any of my training materials.
I speak extensively about tech bros who overhype the capabilities of AI but there are so many examples of use cases where the tech is already consistent and good enough to be used successfully.
Me:
If that is the case I am unsure why you have chosen to do so on here by calling reference to those risks “flawed” and “ideological”!
Matthew Letts:
Brian Inkster this is just not borne out by responsible use of the right tools “It is clear that, in most cases, the dangers of using Al (especially GenAI) in legal practice is greater than not doing so. Until the balance shifts (if it ever does) should that not be enough to dismiss any suggestion that not using Al could result in professional negligence?”
That is the issue I take.
The professional negligence point is easily answered by the legal test for negligence. What is the standard of a reasonably competent professional working in that area? If the conduct falls below that then a claim in professional negligence will succeed.
We are already there in terms of negligence use of AI. I believe that a day will come, within the next five years, where there will be a negligence claim for failure to use AI that will succeed.
Me:
Someone other, in the comments, has suggested it won’t happen within 15 years. Your 5 years is the standard futurist prediction time frame. I have put it in my diary and we will circle back then (in 5 years time) to see if you or they were correct. I rather think that it will be like the Self Driving Cars/Radiologists time line from my ‘GenAI ain’t magic’ talk.
Matthew Letts:
self-driving cars are now here though. Tesla is rolling out robo taxis. FSD (full self driving) is coming rapidly to more and more markets.
I’ll put it in my diary too and let’s see who the history books are kinder to.
Me:
You haven’t been paying attention. Elon Musk’s time line was 2 years from 2015. Geoffrey Hinton’s re. Radiologists was 5 year’s from 2016. My comparison is based on those 2 and 5 year predictions. As we now agree, time will tell. But beware of the seven deadly sins of LegalTech predictions: https://thetimeblawg.com/2018/10/01/travels-through-the-blawgosphere-2-artificial-intelligence-and-law-robots-replacing-lawyers/
Matthew Letts:
Do you think the seven deadly sins still hold?
Me:
Yes. See them committed regularly.
-+-+-+-+-+-
Bobby Murray (Partner in Dispute Resolution at Harper Macleod LLP. I help people resolve construction disputes, pursue negligence claims against professionals, and navigate other difficult challenges across business and life.):
Don’t for a moment believe that any court anywhere would agree with that at any time within the next fifteen years at least.
Me:
I agree and certainly hope they don’t too.
-+-+-+-+-+-
Stephen Canning C.A. BSc Eng (Extensive experience governance & op risk at Sovereign Wealth Fund | Invest only in People | Family Office | “Kindness, that’s it” | Blockchain is coming | Share your luck and confidence | Only ever assume it’s possible):
Enjoyed your exchange Brian Inkster and Matthew Letts would love to connect with you both and follow the AI legal journey.
I would love to get your thoughts on blockchain and its potential impact on the legal profession.
Matthew Letts:
Stephen Canning – Very happy to connect. I think blockchain is an interesting area which helps with the idea of zero trust contracts.
Quentin Solt takes this one step further, away from blockchain and into the living contract. It’s a concept I’m a big fan of and think it has real potential.
Stephen Canning:
Certainly an interesting time to be a commercial lawyer.
The problem I foresee for the legal profession is when trust across large communities is possible without contracts, using technology.
I never give anyone advice however knowing what I know, if I was a commercial lawyer right now I would be looking to use my undoubted talents in a different direction. AI will give plenty of advice on that, please take my comment in the kind way they are intended.
Have a memorable week ahead.
-+-+-+-+-+-
Dr Annette Davison (The Water Risk Doctor, helping protect environments, communities and organisations from exposure to contaminated water):
Megan Dyson
Megan Dyson (Law & Policy):
Thanks, Annette. Such important questions and a real challenge for our profession. We’ve been told during Law Soc CPD that we may be failing our clients – and significantly, not appropriately charging professional fees – if we aren’t using AI for tasks that it ‘clearly’ should be able to do. For example, AI should be able to be used for junior lawyer tasks such as document management – indexing, searching, possibly doing initial parts of the discovery exercise. So in answer to the question should a modern lawyer be proficient in the use of appropriate AI prompts and have systems in place for appropriate checking of products – absolutely. It should be part of the requirements for being admitted to practice and probably CPD too.
Me:
Is your Law Society taking into account the verification burden depending on the task in question? If it is a task that it ‘clearly’ should be able to do does that mean one where the verification burden is very low? If it is high the cost to the client may be similar or greater to not using AI for the task at all? Is your Law Society’s view based solely on cost i.e. if it is cheaper to use AI for a task than not using it you should use it and charge the client accordingly?
-+-+-+-+-+-
Ben Snape (Claim.co.uk & Litigated.com | £100M+ recovered – legal Technology, Advanced Marketing solutions, AI development. 20+ years serving the legal industry):
I don’t think the standard should be “did you use AI?”, it should be “did you exercise reasonable skill and care?”
As AI matures, there may come a point where failing to use it for certain routine tasks could be difficult to justify. Equally, blindly relying on AI without understanding or verifying its output could be just as negligent.
The future isn’t AI versus humans. It’s competent professionals using AI appropriately, with accountability remaining firmly with the professional.
Brett Walker (Project Manager, Interaction Insights & Performance:
Ben Snape ⭐️ I think that the more complete question is: Did you exercise reasonable judgement in determining how to perform the task, and did you exercise appropriate professional expertise in carrying out the task using the chosen tools and methods?
Me:
Thanks Ben Snape ⭐️ and Brett Walker. I am sure the standard will be developed through time by the courts when they tackle the thorny question of a professional negligence claim where a lawyer did not use AI and the claimant thinks they should have! I’d like to think a professional negligence claim would be unlikely to arise for failing to use AI for certain routine tasks that you performed equally as well without it.
-+-+-+-+-+-
Nikunj Mistry (Helping CIOs & Digital Leaders Accelerate AI, Data & Enterprise Modernization | UK & North America):
This raises an important question for every profession adopting AI Brian. The standard shouldn’t be whether AI is used or avoided, but whether professionals can apply it responsibly, understand its limitations and exercise sound judgment. Ultimately, accountability still rests with people, not the technology.
Me:
Indeed and agreed.
-+-+-+-+-+-
Rob Westmacott (Empowering Enterprise to deploy AI securely and without friction, powered by AliasPath™ I Agentic systems | Building better DSAR processing outcomes for DPO’s and HR | 3x Founder | 2x Exits | Golfer when possible):
The concern is understandable, but the two propositions are not necessarily contradictory. The UKJT is not saying lawyers are under a general duty to use AI. Its proposition is conditional: non-use could be negligent where a reasonably competent professional would have used a particular tool for a particular task. That requires more than showing another firm had a better platform. Was the tool established, suitable, reasonably available and demonstrably capable of reducing the relevant risk? Would using it probably have prevented the client’s loss? Competence does not require lawyers to master every platform or match the profession’s best prompters. The standard is reasonable care and skill in context – not technological perfection or hindsight optimisation. The point about Anita is still reconcilable. If she uses AI, she must understand it sufficiently to supervise it safely. If an AI-assisted method eventually becomes necessary for competent practice, she may have to acquire that understanding, obtain assistance or decline work she cannot perform competently. Lack of knowledge would not indefinitely excuse failure to use an essential tool. For most GenAI applications, however, we are some distance from that position.
Me:
This presumably would depend on a stage being reached where the vast majority of lawyers were so reasonably competent? That, as you say, is some distance away!
-+-+-+-+-+-
There follows a repost on LinkedIn with comments and reactions to that repost:
Deborah do Carmo (Lawyer-Linguist. Raising the bar in legal translation since 2002. Dutch & Portuguese to English. CIOL-certified. Former insolvency practitioner (SA), CPPI-qualified (Scotland). Triple national: 🇬🇧 🇿🇦 🇵🇹):
Brian Inkster has written an insightful post about whether lawyers could be held professionally negligent for failing to use AI (especially GenAI) and how any such duty could ever be defined and applied in practice.
The exchange in the comments between Brian and Matthew Letts is also worth reading.
To put it mildly, they do not see eye to eye.
Brian refers to Damien Charlotin’s AI Hallucination Cases Database. It tracks legal decisions worldwide in which a court or tribunal has found or inferred that someone relied on AI-generated false material, most often fabricated or misrepresented legal authorities.
Matthew describes the current tally of 1,751 decisions as ‘a handful’ when compared with the total number of negligence cases over the same period, although he provides no figure for that wider category.
That is rather like downplaying the significance of runway incursions in aviation because they account for only a small proportion of all safety-related incidents. The comparison buries an identifiable source of harm within a much larger category, making it appear less significant.
The same applies here. The question is whether AI use poses a particular risk in legal practice, not how the number of known AI-related cases compares with the total number of professional negligence cases. Respectfully, that is the wrong denominator.
Of course, humans make mistakes. The real question is whether AI reduces the risk of human error or introduces new risks of its own, and whether the verification burden – the time and expertise needed to check its output – outweighs any claimed benefit.
There is also the obvious iceberg problem: Damien’s database captures only cases in which AI-generated false material entered legal proceedings, was detected and became important enough for a court or tribunal to mention it in a decision.
It does not capture errors caught before filing, documents corrected or withdrawn, matters settled privately, complaints dealt with outside court, unreported decisions, harm that never led to litigation or mistakes that were never discovered.
That does not tell us exactly how large the unseen part of the iceberg is. But it does mean that the visible figure cannot sensibly be treated as the full extent of the problem, still less dismissed as insignificant by comparing it with a much wider and largely unrelated category.
Whatever our views on AI, its particular risks should be examined on their own terms – neither downplayed nor exaggerated by flawed comparisons.
Matthew Letts (Strategic Tech Advisor, Founder, Entrepreneur, Solicitor, General Counsel and Legal Quant – Helping lawyers avoid burnout on busy work and law firms make more profit – Building big things in 2026 | Legal 500 – Key lawyer):
Deborah do Carmo thank you for the shout out. Stats incoming on the number of filed professional negligence cases in the same period.
Have you considered the use of AI to reduce the risk of negligence? Citation verification, chronology checking, source material interrogation? These are all things which are being looked at but are completely overlooked by those who see AI as simply too risky to use.
Deborah do Carmo:
Thank you, Matthew. I’ll be interested to see the figures, but the number of filed professional negligence cases would still be the wrong denominator. Unless the data show the number of comparable legal matters in which generative AI was used, they cannot tell us whether 1,751 is a ‘handful’ or what the relevant failure rate is.
Potentially useful applications should be examined seriously. I experiment with and am in favour of tech that works in my field. Sadly, much of it doesn’t. In any event, that is a separate question. My post addresses the validity of the comparison you made.
Matthew Letts:
Stats now provided. Would be very interested to hear your thoughts.
Deborah do Carmo:
Will take a look, Matthew, thanks.
-+-+-+-+-+-
Stephen Allen FRSA (Gets S*** Done | Former Lawyer | AI Realist | Fellow, Royal Society of Arts | Fellow, College of Legal Practice Management | Inaugural Financial Times Legal Intrapreneur of the Year | 3 x FT Awards | #COYW):
Matthew Letts – quis custodiet ipsos custodes.
[Translation: “Who will guard the guards themselves?”—a Latin phrase from Juvenal’s Satires questioning how those in power can be held accountable.]
Matthew Letts:
A good example of why clients use AI. They want to speak the same language as their lawyers. We should all be watching the AI companies and seeking to ensure their feet are held to the fire.
Stephen Allen:
But also the line isn’t a legal maxim. It’s about ultimate accountability. It merits reference.
Matthew Letts:
I know what it means. I just think it’s funny that you use Latin which is one of the classic lawyer stereotypes. Oversight is important.
Stephen Allen:
Maybe I am not doing it as a lawyer (if not being a legal maxim and I haven’t practised since 2003) but as someone who reads about the past and the future. Your assumption is – honestly – a little rude.
-+-+-+-+-+-
Matthew Letts (with stats he referred to earlier now produced):
The first of the stats. Of the 1,751 cases, 1,027 are from litigants in person. That’s 58.7%. 690 (including the paralegal cases, the prosecutor cases, and federal defender case) are from lawyers. That’s 39.4%.
Deborah do Carmo:
Thank you, Matthew.
As my post makes clear, the issue is your chosen denominator. The breakdown you have provided here tells us only who was involved in the 1,751 reported decisions. It breaks down the numerator but does not supply the denominator needed to assess the rate of AI-related failure.
Scott Simmons (Turning Lawyers Into Rainmakers | Business Development Coach & Trainer for Lawyers I Value Pricing & Subscriptions for Law Firms | Creator of The BD Breakthrough Blueprint® | Co-Host of Legal Off The Leash podcast):
Deborah do Carmo – The same is true of human errors. How many don’t end up in front of the SDT?
Mine never did. And I know lawyers who made mistakes and their errors never made it in front of judges or disciplinary panels.
Brian’s argument is ‘don’t use AI until you don’t have to verify its output.’
That’s not how we would work with human lawyers and we shouldn’t treat AI differently.
The fact is, while we’re arguing about hallucinations that are down to failure of verification, there are entrants into the legal market that are starting to eat away at market share because they’re focused on getting infrastructure, process, and customer care right.
That’s where our conversation should have moved to by now. The rest of the business world moved on from hallucinations a long time ago.
Deborah do Carmo:
Good to hear from you, Scott. I agree that many human errors never reach a court or disciplinary panel. That reinforces the iceberg point I made: the available figures are incomplete on both sides. It does not, however, supply the denominator needed to assess the rate of AI-related failure or support Matthew’s description, in his exchange with Brian, of the 1,751 recorded decisions as ‘a handful’.
I also do not read Brian’s argument as ‘don’t use AI until you don’t have to verify its output’. The verification burden is part of the assessment. If a tool saves time and reduces risk after proportionate checking, that is relevant. If checking it takes longer than doing the work directly, or still leaves material uncertainty, that is relevant too.
I agree that infrastructure, process and customer care deserve serious attention, particularly where new entrants are competing on those grounds. But commercial traction does not make reliability and verification yesterday’s conversation. Those questions remain part of getting the infrastructure and process right in the first place.
Scott Simmons:
Thanks, Deborah. And I hear everything you are saying.
The problem with Brian’s assertions on verification is that there is no data to suggest verification takes twice as long as simply doing it yourself.
I’ve got clients who, having built proper structures around their AI tools, say they are saving 75% of time on first drafts, and accuracy has increased compared to human lawyers.
And yet Brian has proliferated a ‘law’ as if it is based in truth. It’s not, and it’s dangerous to feed this level of uncertainty and fear into the profession when the same solutions for verification and reliability that we put in place for humans are available to us for AI.
Deborah do Carmo:
Hi, Scott, have you had a chance to read Brian’s original post on ‘Inkster’s Law’? He didn’t coin the term. The post recounts a distinctly light-hearted LinkedIn exchange, and Brian expressly says that the ratio may change as lawyers begin using and testing the technology more extensively. I don’t read it as an assertion that every AI-assisted task has been empirically proved to take exactly twice as long to verify, still less as an attempt to instil fear and uncertainty into the profession.
I’m also not sure the evidential standard is being applied evenly. Brian’s formulation is rejected because there’s no data establishing a universal 2:1 ratio, while reports from particular clients of 75% savings on first drafts and increased accuracy are offered in response. Those reports may accurately describe their experience, but they don’t establish a general proposition.
More importantly, producing the first draft isn’t the central issue. A 75% saving at that stage doesn’t establish a 75% saving in producing final, reliable legal work.
The relevant assessment covers the entire workflow: checking authorities and sources, identifying omissions, testing the reasoning against the facts, correcting plausible but subtle errors and reaching the point at which a lawyer can safely assume professional responsibility for the result.
Joshua Yuvaraj’s paper on the ‘verification-value paradox’ addresses precisely this problem: claimed efficiency gains may be overstated when the cost and professional necessity of verification aren’t properly accounted for. It doesn’t prove a fixed 2:1 ratio, but it does support the concern underlying Brian’s formulation.
One of the central lessons from my Master’s-level course in The Psychology of Human–AI Interaction at the University of Aberdeen was that verification isn’t simply an existing human control applied to a new tool. It’s shaped by factors such as what users believe the system can and cannot do, interface design, task allocation, cognitive load, automation bias, overreliance, cognitive offloading and the persuasive fluency of AI output.
Those factors can affect what people notice, what they question and how readily they accept an answer. Existing controls may still help, but we can’t assume they’ll operate in the same way when someone is reviewing AI-generated work rather than another human’s work.
To my mind, asking valid questions about verification doesn’t create uncertainty; it exposes uncertainty that already exists. Examining it isn’t fearmongering but a necessary part of responsible adoption. If a memorable device such as ‘Inkster’s Law’ helps keep the verification burden in view, I can’t see the harm in that.
It’s simple: if verification doesn’t take twice as long in a particular workflow, don’t apply the ‘law’. We all know laws are made to be broken. 😉
Links:
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5621550
Scott Simmons:
Deborah do Carmo – Again, I appreciate everything you’re saying. And it’s clear you’ve done a lot more work on this than most—including Brian.
I’m putting forward anecdotal evidence in response to Brian’s ‘law’ which has zero basis in research—it’s completely ideological.
And your point about verification being influenced by how we view where the work has come from is clearly true.
That said, two points:
1. If we choose to put more weight on AI-generated work and therefore don’t verify it as we would a human’s work, that’s a human problem, not an AI problem.
2. How much time, effort, and energy do we put into training human lawyers? I wonder if the same amount of time goes into training AI and whether the outcomes are better or worse over time? From what I’m seeing, it’s less time with better outcomes.
Deborah do Carmo:
Thanks, Scott. I’m going to push back on two things before coming to your two points.
1. ‘Inkster’s Law’
If you read the blog post itself (link above), I don’t think you can reasonably treat the 2:1 formulation as a literal empirical claim. Again, Brian didn’t coin ‘Inkster’s Law’; the post is light-hearted in tone, and he expressly says that the ratio may change with greater use and testing.
Nor is it correct to say that the underlying concern has ‘zero basis in research’ or is ‘completely ideological’. Joshua Yuvaraj’s paper on the ‘verification-value paradox’ (link also above) is directly on point and well worth reading. It argues that apparent gains in drafting efficiency may be offset by the professional imperative and cost of manually verifying GenAI output in legal practice. It doesn’t prove a fixed 2:1 ratio, but it squarely addresses the problem Brian’s shorthand is intended to capture. Brian has his own good blog post on the verification burden
In short: the exact ratio is a memorable rule of thumb; the verification burden is a serious subject of legal scholarship, not ideology.
2. Human–AI interaction
My point wasn’t that we check work differently depending on where we think it came from. I’m saying that we need to check it differently because it comes from an AI system, with different characteristics and failure modes, and because the interaction itself affects how people scrutinise the output.
The factors I mentioned above can affect what people notice, what they question and what they accept. Verification therefore isn’t simply the same human control applied to a different producer of text. The source changes the nature of the checking task.
On your two points:
1. The psychology of human–AI interaction is a fascinating field and impossible to do justice to in a LinkedIn comment, but it explains why this isn’t simply a ‘human problem’. The way people trust, interpret and scrutinise AI output is shaped by the interaction between the person, the system and the workflow. The lawyer remains responsible, but that doesn’t make the technology’s characteristics irrelevant to how the failure arises.
2. Of course, we invest heavily in training lawyers. We have to: the buck stops with them. Their training develops substantive legal knowledge, judgement, ethical duties and professional accountability. Training or configuring an AI system is fundamentally different. It may improve performance within a defined workflow, but it doesn’t give the system judgement, duties or responsibility for the result.
That is why the relevant comparison isn’t simply how much time is spent ‘training’ each. It is whether the complete human–AI workflow produces better verified outcomes, once setup, maintenance, checking and professional sign-off are all taken into account. The jury is still out on how far that holds true across legal practice.
That said, we’ve moved some distance from what my original post was about, so I’ll leave it there. Good to exchange views with you, Scott – but I really do need to get back to some paid work!
-+-+-+-+-+-
Matthew Letts (continuing his stats from earlier):
From mid-2023 to mid-2026, professional negligence claims issued in courts against lawyers worldwide plausibly total around 125,000, or roughly 40,000 a year.
Error bars: 45,000 low, 300,000 high. I’d put 80% odds on the true figure sitting inside that range, and no better than even odds on the central 100,000–150,000 band. No jurisdiction publishes an authoritative count of issued claims, and data outside the Anglophone world is thin. It’s also important to remember that most negligence claims never get filed, with insurers resolving them at the pre-action stage.
Against that, Damien Charlotin’s database records 1,751 hallucination decisions over the same window, with 690 involving lawyers. That is roughly 0.5–1% of the ordinary professional negligence caseload, or 1 in every 150–200 claims. It is growing fast since spring 2025, but is still a sliver (or a handful) of the base rate.
Deborah do Carmo:
Matthew, absent links, sources or methodology, the 125,000 figure cannot be checked. But even if it were accepted, it would still not answer the point I am making.
I am not arguing that 1,751 establishes a failure rate. My point is that reported AI-related decisions and estimated professional negligence claims are different kinds of data and cannot support the rate you claim.
Damien’s 690 lawyer-related decisions are reported decisions in which AI-generated false material was identified, not professional negligence claims. Your 125,000 is an unsubstantiated estimate of all issued professional negligence claims arising from every cause.
You have not shown that the 690 form part of that 125,000.
The datasets also measure different procedural stages: reported decisions on the one hand and issued claims on the other.
Dividing one figure by the other therefore cannot establish that AI-related failures represent ‘0.5–1% of the ordinary professional negligence caseload’ or ‘1 in every 150–200 claims’. It simply turns the same invalid comparison into a percentage.
That is the point my post addresses.
Matthew Letts:
Deborah do Carmo – every AI hallucination case is a professional negligence case. That’s my point and it keeps, seemingly, being missed.
Deborah do Carmo:
Not missing it, Matthew. In fact, I have consistently argued that courts and regulators have been far too lenient towards the lawyers involved and that the conduct may amount to professional negligence, professional misconduct or both.
When you cite an authority, you are representing to the court and opposing counsel that it exists and supports the proposition on which you rely.
But whether every lawyer-related decision involves conduct amounting to professional negligence is not the issue my post addresses.
The problem is the comparison you are drawing. Damien’s database counts reported decisions involving AI-generated false material. The figures you describe as ‘stats’ appear to be estimates unsupported, so far, by sources or methodology. They concern issued professional negligence claims arising from every cause. These are different things measured at different stages.
Even if all 690 lawyer-related decisions involved professional negligence, dividing them by the total number of professional negligence claims would still not tell us how often the use of generative AI leads to error or harm. That would require data on comparable legal matters in which generative AI was used.
That’s the denominator issue my post addresses.
-+-+-+-+-+-
Me:
Thanks Deborah for highlighting my post and the importance of the number of reported decisions globally in cases where generative AI produced hallucinated content.
I agree with you that the wrong denominator has been used by Matthew Letts when considering the question of hallucinated decisions and professional negligence claims.
Those reported decisions are really only the tip of the iceberg. There will be many such mistakes being built up in legal firms throughout the world daily by lawyers using AI badly that will only become apparent and result in professional negligence claims some way down the line. There is a big time lag involved.
It is interesting to look at the graphs produced by Damien Charlotin – https://www.damiencharlotin.com/hallucinations/?graphs=1 – and, in particular, the one on ‘Timing: Offending Document v Decision’. This, naturally, shows that offending documents are produced some time before the decisions on them are reached. The same will be true (and the time lag likely to be greater) in professional negligence claims being raised.
Thank you also for alleviating me of the verification burden of correcting the hallucinations of Scott Simmons 😉
I can confirm that my argument certainly is not ‘don’t use AI until you don’t have to verify its output.’ I have never said that and I don’t think that. I highlight the verification burden as an important factor in considering AI use and it is one that should not be downplayed. In some cases that verification burden will be low and in other cases it will be high. That is an important barometer in considering whether or not to use AI depending on the task in hand. It will fit some tasks and not others.
Scott thinks we have to move the conversation away from hallucinations. With 7 new such decisions reported on average in courts each business day I don’t see how we can do that. This is especially so in the UK, where Damien’s graphs show a recent spike in hallucination cases involving lawyers compared to party litigants.
Scott also hallucinates on ‘Inkster’s Law’. He copies Matthew in his language of “fear” and “ideology”. Like Matthew, I assume Scott has not read the extensive blog posts written by me on the topic backed up by research that I refer to such as that by Joshua Yuvaraj, to which you refer. If Scott had actually read about ‘Inkster’s Law’ he would have learned what you pointed out to him, namely “that the ratio may change with greater use and testing”. As mentioned earlier it will also vary depending on the task in hand and the suitability or otherwise of using AI for that task.
The ‘GenAI Verification Burden in Legal Practice’ is covered by me in detail here: https://thetimeblawg.com/2025/11/15/the-genai-verification-burden-in-legal-practice/
Matthew Letts:
Brian Inkster – the number of worldwide issued professional negligence cases are the tip of the iceberg too.
It’s amazing how Deborah do Carmo asked for stats, I gave stats, and then the goalposts were moved.
It’s also interesting that you don’t talk about the fact that well under half of the hallucination cases actually involve the lawyer using AI inappropriately.
I’d prefer to argue the facts, rather than the feelings.
Me:
I haven’t noticed any goal posts moving. Deborah’s point on you using the wrong denominator stands strong.
Of course hallucination cases will involve the lawyer using AI inappropriately. That should go without saying. You can read more on that here (although, I appreciate you won’t): https://thetimeblawg.com/2024/01/14/lawyers-or-machines-who-do-you-blame-for-genai-hallucinations/
As I have said before, all the facts are in my 30 blog posts on the topic. Your refusal to read them suggests you are the one labouring under feelings rather than facts.
Matthew Letts:
Brian Inkster – does it show a recent spike? If you actually look at the UK data the number of cases has dropped for the last two consecutive quarters. You’re just making it up as you go along at this point.
Me:
I’m afraid you are the one making it up and misleading again by not looking at the correct graphs relevant to the point actually being made!
Matthew Letts:
Answer the stats point. What you’ve said about a peak in cases is demonstrably wrong. Case numbers are dropping in the UK.
Me:
I stated that in the UK there is “a recent spike in hallucination cases involving lawyers compared to party litigants.” i.e. more have recently been committed by lawyers compared to party litigants than was previously the case. This fact is taken from Damien Charlotin’s graph (which I have posted above) clearly showing that spike. It is not case numbers but the increase in lawyer generated hallucinations compared to party litigant ones that I was clearly referring to. You, again, use statistics that are not relevant to the point in hand. To do so is demonstrably wrong.
Matthew Letts:
Do you know how to read the graph you’re relying on? How many cases does that “spike” relate to?
Me:
The number of cases is not the relevant point. The relevant point is the fact that recently more hallucinated cases have been caused by lawyers rather that party litigants. Simple as that. That should be a cause for concern. If it is not for you, then so be it.
Matthew Letts:
You are stating that the number of cases is not the relevant point. Pretty convenient given that statistic doesn’t support your argument. The reality is that cases involving lawyers have dropped for two consecutive quarters. The data is reflecting adoption, understanding, and responsible use. The 100% spike you are relying on is, quite literally, a single case. Nobody who knows the first thing about data would attach any statistical significance to it.
Me:
The graph reflects quarters. There have been more than one single case in a quarter! It shows a pattern of comparison between party litigants and lawyers. I am sorry if you cannot understand that.
Matthew Letts:
How many cases are there in Q3 2026? Happy to accept your apology in person or in writing.
Clue: it’s one so far
Me:
One indeed in the UK in Q3 2026 (which has just begun). My apologies. And that one is from a lawyer and not a party litigant. That fact alone when we are so far down the line from the publicity surrounding the first hallucinated cases (3 years ago in 2023) should still be cause for concern. Looking at worldwide cases since 1 July 2026 we have 13 from party litigants and 13 from lawyers (which includes one from a Judge). Thus, even.
So lawyers are still making mistakes in using AI as much as party litigants are. Whilst you earlier gave statistics about party litigants outnumbering lawyers (which is not surprising) there has been no real dip in comparison figures since Q2 2024 (see graph below).
Whilst one would like to think that any drop in numbers would reflect “adoption, understanding, and responsible use” that is an assumption. One would have thought that the publicity around hallucinated cases should and would reduce it to zero amongst lawyers, not 13 in as many days in mid 2026.
The fact is that hallucinations remains a real live issue and whatever stats you use it is difficult to suggest otherwise.
-+-+-+-+-+-
Deborah do Carmo:
Thanks for your comprehensive response, Brian. Correct, the goalposts never moved. And the only stats provided (a) referred to the numerator, not the denominator and (b) were not stats but unsourced estimates.
-+-+-+-+-+-
Scott Simmons:
Brian Inkster – Hey, Matthew—if the number of complaints to the Legal Ombudsman goes up year-on-year, is that a cause for greater concern?
What if the number of complaints was less year-on-year, but you added the previous year’s complaints to this year’s? Should we be happy that the complaints have gone down year-on-year, or concerned that the total number keeps going up?
I’m now confused. Once upon a time I’d have been more focused on the year-on-year numbers, but now I’m thinking I should be looking at the total number across all time and losing my #%^* that the number keeps going up.
-+-+-+-+-+-
Alex Smith (Global Search & AI Product Lead (Senior Director) at iManage | Godfather and IA Frontier Founder of #IAbeforeAI):
And the iceberg is bigger, think of all the contracts with AI issues waiting to surface in a few years!!
Matthew Letts:
A legitimate concern and there will, over the coming years, inevitably be fallout from early stage, irresponsible, AI use.
Deborah do Carmo:
Agree, Alex, they will only surface later (we see this in legal translation, too); that’s the time lag Brian is referring to.
Alex Smith:
Yep, David Howorth expressed what I’d been thinking for ages on this https://www.linkedin.com/posts/david-howorth-8256a892_the-sullivan-cromwell-ai-filing-errors-activity-7452976875708190720-pGrD?utm_source=share&utm_medium=member_desktop&rcm=ACoAAADSlX8BRstQqE0yFcioIvi85MJgf67NjYA
Deborah do Carmo:
Thanks for this, Alex. David does sum it up nicely.
-+-+-+-+-+-
Rafie Faruq (CEO of Genie AI: Agentic Lawyer):
I think the “competent member of the profession” test is doing a lot of heavy lifting here. Adoption is so uneven right now that there’s no stable baseline to measure anyone against, and by the time the courts settle one, the tools will have moved on twice over. My bet is insurers and clients set the real standard of care long before the case law does: once enough firms use AI for first-pass work, billing clients for the fully manual version becomes the harder thing to defend.
-+-+-+-+-+-
Another LinkedIn post on AI and Professional Negligence
See also comments by Philip Armstrong and others on his LinkedIn post.
