August 3rd, 2026
Artificial intelligence in legal work: what it is doing well, where it is going wrong, and where the rules are heading
Tatiana Svetlova · Founder and Principal Solicitor, Svetlova LLP · July 2026
In R v FGD [2026] EWCA Crim 918 a criminal trial was stayed part-way through after a download of the complainant’s phone revealed that she had prepared for cross-examination using an AI chatbot. It had produced a tidied narrative of her account, together with two dozen questions a defence advocate might ask and suggested answers to each. The Court of Appeal reversed the stay: the coaching was not in dispute, but the trial process had ample means of dealing with it — an adjournment, exclusion, recalling the witness on the documents, or a firm direction of the kind approved in Momodou. Two features mattered. Her evidence-in-chief was a pre-recorded ABE interview given years earlier and could not have been contaminated; and, unusually, everything the AI had told her existed in writing and could be put before the jury.
Coaching by a human being happens behind a closed door and leaves no trace. This left a transcript — which is why it was caught, and why most of it will not be.
The court added an observation that will outlast the case: this is unlikely to be an isolated example. It invited the National Police Chiefs’ Council, the Crown Prosecution Service and the Criminal Procedure Rule Committee to consider how to address and regulate the use of AI in criminal proceedings. That is the question this article is really about: is a regulatory check now due, and of what kind?
What AI is quietly doing well
It is worth saying plainly, because the reported cases are all disasters and give a distorted picture: in day-to-day practice these tools are doing a great deal of good, and much of it is invisible.
The most important effect is on access. People who would not previously have had proper — or affordable — access to legal help can now enter this territory with far more confidence. They arrive understanding roughly what kind of problem they have, what the process is likely to involve, and what questions are worth asking. That is a significant shift for anyone who has historically had to choose between paying for advice they could not afford and going without it altogether.
The effect is visible on the client side every week. Clients — particularly younger buyers in residential transactions — now arrive with genuinely well-formulated briefs. They have thought about what they want, identified the points that worry them, and can articulate an issue in terms a solicitor can act on immediately. A buyer who flags potential issues is not a threat to professional work – it is a better instruction, and better instructions make better outcomes.
The ordinary gains are real too: first-pass review of long documents, translation triage in cross-border matters, chronologies, structuring a set of enquiries, and stress-testing an argument by having it attacked.
Where it goes wrong
Fabrication. The most familiar failure is the invented authority. In Ayinde v Haringey LBC and Al-Haroun v QNB [2025] EWHC 1383 (Admin), the Divisional Court considered its Hamid jurisdiction in respect of non-existent citations placed before the court. The Crown Prosecution Service has apologised for hallucinated material in court documents; at least one substantial firm has referred itself to the SRA. The problem is never that the tool was used. It is that nothing was checked.
Contamination of evidence. This is the FGD problem, and it is more insidious than fabrication because nothing produced is false. An AI asked to help a witness prepare will do exactly what it is asked: it will smooth the account, reconcile the inconsistencies, supply the confident answer, and — as the documents in FGD did — reassure the witness that their account is strong and consistent. That is precisely the danger identified in R v Momodou and Limani [2005] EWCA Crim 177 at [61]: an honest witness shifts emphasis without realising it, and a dishonest one calculates rapidly how the testimony might be improved. These tools are agreeable by design. Agreeableness is a defect in a witness preparation aid.
Confidentiality and privilege. Putting client material into a general-purpose consumer system is a disclosure of that material to a third party. Whether it is a breach depends on the terms, the deployment and the safeguards, but the analysis has to be done rather than assumed — and it engages both the duty of confidentiality and data protection obligations. Sanctions-sensitive and cross-border matters raise this in sharper form.
Disclosure exposure — the trap nobody anticipates. The Court of Appeal in FGD made a point that deserves wider attention. A complainant’s use of AI can itself create a reasonable line of enquiry, opening their phone and their online activity to a far broader and more intrusive examination than would otherwise have been justified.
Fluent wrongness. The output reads like competent work whether or not it is. That is a particular hazard where the user cannot yet evaluate what they are reading — the litigant in person, the client who believes they have received advice, and the practitioner working at the edge of their knowledge. The tool that helps you learn faster is the same tool that will let you sound expert before you are.
So is AI due a regulatory check?
Yes — but not the kind usually imagined. The answer is not a prohibition on AI in legal work. That is both unenforceable and undesirable, and it would fall hardest on the users who benefit most: small firms and unrepresented litigants.
The civil side is already some distance ahead. The Civil Justice Council, through a working group chaired by Birss LJ, consulted between February and April 2026 on whether rules are needed to govern the use of AI by legal representatives in preparing court documents. Its interim position is instructive: existing professional frameworks are treated as sufficient for pleadings, skeletons and advocacy documents, because the lawyers who sign them are already regulated and already owe duties to the court.
The proposals bite where regulation runs out — a declaration that AI has not been used to generate the content of a trial witness statement, including by altering, embellishing, strengthening, diluting or rephrasing it; a transparency requirement for experts; and continuing work on witness statements and on litigants in person, where the issues are hardest. A final report is expected later this year.
The criminal side has no equivalent and now has a Court of Appeal invitation to build one.
Four things would matter more than a rule change alone:
Warning at the point of contact. Witnesses learn what is expected of them from the officer in the case, from witness care, and from the ABE process. That is where the message belongs — before the conduct, not in a rule they will never read.
A record, not a ban. FGD shows the value of an auditable trail. Rules that push AI use into the open produce evidence a jury can weigh; rules that drive it underground produce nothing.
Guidance on the remedy. The Court of Appeal’s practical direction to counsel — to raise and explore each available trial-process remedy explicitly, so the judge can rule on them separately — should be standard practice in any abuse application of this kind.
Restraint on disclosure. If discovering AI use routinely justifies a full device download, complainants will pay a disproportionate price for a mistake most of them do not know they are making. Bater-James proportionality has to hold.
What this means for firms now
For solicitors, the regulatory position is already clear enough to act on. The SRA has not restricted the use of AI; it has confirmed that responsibility for the work is undiminished by the fact that a machine helped produce it.
A closing thought
The lesson is not that these tools are dangerous. It is that they are indiscriminate. They will help a solicitor understand an unfamiliar jurisdiction, help a buyer arrive with an intelligent brief, and help a witness rehearse evidence they were never permitted to rehearse — with the same fluency and the same air of authority in every case. The regulatory task is not to decide whether AI belongs in legal work. It is already there. The task is to draw the line between preparation and contamination, in terms that ordinary people encountering the justice system for the first time can actually understand.
Svetlova LLP, 2 Allen Street, London W8 6BH
This article is published for general information and does not constitute legal advice. Specific advice should be taken on any particular matter.
