Guest post by Reuben Vandercruyssen, a senior associate, and Olivia Drawbell, an associate, at City firm Hogan Lovells
Reuben Vandercruyssen
AI is now part of professional work, from transcripts and summaries to disclosure support, data analysis and first drafts.
Used well, it can make work faster, cheaper and sometimes better. Used badly, it can produce confident mistakes, expose confidential material, or encourage professionals to rely on outputs they have not properly checked.
That creates an obvious professional negligence risk. A professional who uses AI carelessly may fall below the standard of reasonable skill and care. As AI tools become more reliable and more embedded, there may also be circumstances in which a reasonably competent professional would be expected at least to consider using them.
The legal framework
The starting point remains Bolam v Friern Hospital Management Committee [1957] 1 WLR 582: a professional is not negligent if they follow a practice accepted as proper by a responsible body of professional opinion. Bolitho v City & Hackney Health Authority [1998] AC 232 adds an important control: that opinion must withstand logical analysis.
That is also the approach adopted by the UK Jurisdiction Taskforce’s (UKJT) draft legal statement on private law liability for AI harms in England and Wales. It asks whether liability may arise from using AI, or failing to use it, in professional services.
The standard is dynamic. As the UKJT statement recognises, reasonable skill and care can change as technology and professional practice develop. A method once acceptable may become outdated. A new tool may be too risky today, but difficult to ignore once it is reliable, adopted and materially better than the manual alternative.
Negligent use
The immediate risk is negligent use. The UKJT statement identifies the kinds of factors likely to matter: inadequate due diligence; insufficient understanding of the tool; lack of testing; inadequate client transparency; confidentiality, privilege or data-security failures; and inadequate human oversight.
In a claim, the evidence will focus on what the tool was, what it was used for, what checks were done, what the client was told, and what a competent peer would have done.
Verification is the clearest example. The UKJT statement gives the example of a barrister who fails to review incorrect AI-generated written submissions.
Recent case law points the same way. In Elden v HMRC [2026] UKFTT 00041 (TC), the First-tier Tribunal stated that it is “the human who relies on [AI’s] use [who] bears the responsibility for the accuracy”. Ayinde v The London Borough of Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin) carries the same practical message: lawyers remain responsible for material placed before the court, including where AI may have been involved.
That point should not be confined to lawyers. AI can assist professional work, but it cannot take responsibility for it. A polished output is not the same thing as a verified output. The more consequential the output, the more important human checking becomes.
Olivia Drawbell
Negligent non-use of AI
The more novel point is whether failing to use AI may itself become negligent. In principle, there is no reason why not.
The UKJT statement makes the point clearly: the question whether an AI tool should be used is no different in principle from the question whether any other professional tool should be used.
Its examples are useful. A radiologist may be expected to use an effective tumour-detection AI system if it is available at reasonable cost and improves detection. An auditor may be expected to use AI to identify anomalies and fraud across very large transaction sets where purely human review is not realistic. A solicitor may need to consider AI-assisted or technology-assisted review in large-scale disclosure.
The disclosure example is particularly useful because the direction of travel is already visible. PD 57AD and the disclosure review document are built around scoped, co-operative and proportionate disclosure, and technology-assisted review (TAR) is defined broadly as document review undertaken or assisted by technology.
More recently, the Competition Appeal Tribunal’s disclosure ruling in Gormsen v Meta Platforms Inc [2025] CAT 85 gave a further nudge in the same direction.
The tribunal did not mandate any particular tool or AI approach, but it left methodology to the producing party’s supervising solicitor, declined to treat AI-assisted review as necessarily experimental where it supplements human review with quality control and sign-off, and recognised that advances in TAR can affect the baseline of what is reasonably necessary and proportionate.
The practical point is not that AI must be used in disclosure. It is that, in suitable large-scale matters, a purely manual approach may become harder to justify if available technology could materially reduce cost and time without undermining defensibility.
This does not mean non-use will often be negligent today. For now, there may be good reasons in a particular scenario not to use AI: confidentiality, privilege, data security, client instructions, regulatory limits, unreliability, cost or lack of suitability.
But those reasons need to be thought through at the time, not reconstructed later as a generic objection to AI. If a tool becomes reliable and materially better than the manual alternative, blanket refusal may become harder to defend.
Guidance, defensibility and disputes
Professional guidance will matter. The UKJT statement recognises that guidance from professional bodies is likely to feature in the court’s assessment.
But guidance is not a complete safe harbour. Falling below applicable guidance may point towards breach, but compliance will not automatically defeat a claim. That follows from Bolitho: even professional consensus must be logical and defensible.
Future disputes will turn on records and evidence: why the tool was chosen, whether it was suitable, what due diligence was done, how outputs were checked, what the client was told, how confidentiality was protected, and whether the loss was caused by the alleged misuse or non-use of AI.
Causation and scope of duty remain important controls. Even if breach is established, a claimant must still show that the breach caused the loss and that the loss falls within the scope of the duty that was breached.
Manchester Building Society v Grant Thornton UK LLP [2022] AC 783 remains the leading authority on scope of duty, and a reminder that liability does not extend to all consequences of a breach, but only to losses within the scope of the duty assumed.
Conclusion
AI does not require English negligence law to be reinvented. Bolam and Bolitho already allow the court to ask whether the professional acted in accordance with responsible and logically defensible professional practice.
Careless adoption of AI can plainly create the possibility of liability. Over time, unjustified refusal to use effective AI tools may also become harder to defend.
The safest position is defensible AI use: tools selected for a proper reason, used for appropriate tasks, supervised by humans, protected by confidentiality safeguards, and supported by a clear record.
Leave a Comment