Key Takeaway
This legal article analyses a R v FGD [2026] EWCA Crim 918, a Court of Appeal decision concerning a witness’s use of AI to prepare for cross-examination and the resulting question of prohibited witness coaching. Matthew Lee draws out ten practical lessons for lawyers, litigants in person and the justice system.
Ad/Marketing Communication
This legal article/report forms part of my ongoing legal commentary on the use of artificial intelligence within the justice system. It supports my work in teaching, lecturing, and writing about AI and the law and is published to promote my practice. Not legal advice. Not Direct/Public Access. All instructions via clerks at Doughty Street Chambers. This legal article concerns witness coaching and AI. Subscribe to the AI Law Commentary here.
Introduction
Lady Justice May’s introduction to R v FGD explains clearly what this important case concerns:
“This case raises interesting contemporary issues concerning the use of AI by a witness before giving evidence in a criminal trial. It is a prosecution appeal against the Recorder’s ruling, part-way through trial, staying proceedings as an abuse. Upon the prosecution giving notice of its intention to appeal, accompanied by the usual prosecution undertaking, the jury was discharged”
I have been concerned about the potential for witness coaching by AI for some time. New readers may be interested in my articles on AI shaping witness evidence , smart glasses and live coaching and the judicial comments that AI generated documents must “reflect rather than suggest a party’s case and evidence.”
Although R v FGB appears to be the first reported case to address the issue directly, to my knowledge it is not the first time the concern has arisen. Lady Justice May’s careful and clear analysis also goes considerably further than the immediate question of coaching. In this legal article, I seek to identify and share ten important considerations which arise from the judgment.
R v FGD
Before addressing those ten lessons, it is important to analyse the case carefully. I want to be as careful about the facts as the court itself was. The trial concerned an allegation of rape. The detail was rightly excluded from the headnote and judgment so that the decision could be published without rendering the case identifiable and without waiting for the conclusion of any retrial. My focus will be, as always, on the AI issues.
During cross-examination, a witness, C, was challenged about where they had been at a particular time that evening. After court, C apparently searched their cloud storage and produced screenshots said to show where they had been and when. The following day, when the prosecution applied to recall C to put that material before the jury, the defence asked to see a full download of C’s phone. The Recorder ordered the download.
“The download revealed two documents which contained a summary record of what C had recalled of events on the night in question, followed by a series of questions in the form of a cross-examination of C’s account, together with suggested answers. In short, it appeared to be a documentary record of witness-coaching. C was asked who they had been communicating with, to which C responded: “I had put stuff into AI to help me prepare myself for the court”. Although the defence did not accept that the documented questions and answers were generated by AI, as opposed to being contained in a communication from a real person, this appeal has proceeded on the basis that the documents were generated by AI in response to prompts from C. We do not know the nature of those prompts. We shall refer to the documents in this judgment as “the AI documents”. A flavour of the personal nature and tone of the AI engagement with C comes from the introductory section, which reads:
“Before we go through this I want to check in with you again. This section – this specific night – is the heart of the case and going through it in this level of detail is genuinely hard. Please make sure someone is with you or available to you today.
What I am going to do first is pull together everything you’ve shared about that night into one clear picture, so that you can see how strong and consistent your account actually is. Then I’ll go through every question a defence lawyer might ask with suggested answers.”(para 5)
There then follows several pages of text containing the promised summary, together with 24 highly pertinent questions and suggested answers, gathered under such headings such as “On what happened that night”, “On Consent and Capacity”, “On Physical Evidence” and “On Your Behaviour After”. (para 6)
Defence counsel applied to stay proceedings for abuse on the basis that a fair trial was impossible. His key concern was that no amount of cross-examination could adequately identify what evidence was properly C’s and what was somebody else’s. The prosecution’s response was that any difficulties arising from the AI documents could be dealt with in the course of trial.
The transcript of the abuse hearing records a lengthy discussion between the Recorder and prosecuting counsel as to whether the AI documents were properly described as witness training or coaching, concluding with this exchange:
“RECORDER: I am satisfied on what I have seen, having worked through those questions with you, Mr Lamb – about what would be coaching and what would not be permitted – that this is evidence of coaching, whether or not done by AI or someone else.
MR LAMB: I understand, and of course I respect your decision on that part. Then, it is a matter for your Honour, how you proceed on that basis. I can see that perhaps in the light of that decision, there’s only one way forward for the case, forgive the phrase, to proceed, essentially by not proceeding … “ (para 8)
The Recorder expressed the view that, once there had been coaching, he could not see what remedy was available, and he stayed proceedings on that basis.
On appeal, the court set out the parties’ respective submissions before turning to its analysis and decision. Leading prosecuting counsel argued that the trial process itself offered several ways of dealing with the prejudice caused by C’s use of AI, including an adjournment, discharging the jury and starting again, excluding C’s evidence, or recalling C to be cross-examined on the AI documents with a robust judicial direction. Defence counsel maintained that none of these would answer the difficulty, since the questions and suggested answers went to the very heart of C’s account.
The court began its discussion with the governing principle:
“Witness training is forbidden in all jurisdictions. The dangers of training witnesses on their evidence were succinctly summarised, in the context of a criminal case, by Judge LJ (as he then was) in R v Momodou and Limani [2005] EWCA Crim 177 at [61]:
“There is a dramatic distinction between witness training or coaching, and witness familiarisation. Training or coaching for witnesses in criminal proceedings (whether for prosecution or defence) is not permitted. This is the logical consequence of well-known principle that discussions between witnesses should not take place, and that the statements and proofs of one witness should not be disclosed to any other witness. (See Richardson [1971] CAR 244; Arif, unreported, 22nd June 1993; Skinner [1994] 99 CAR 212; and Shaw [2002] EWCA Crim 3004.) The witness should give his or her own evidence, so far as practicable uninfluenced by what anyone else has said, whether in formal discussions or informal conversations. The rule reduces, indeed hopefully avoids any possibility, that one witness may tailor his evidence in the light of what anyone else said, and equally, avoids any unfounded perception that he may have done so. These risks are inherent in witness training. Even if the training takes place one-to-one with someone completely remote from the facts of the case itself, the witness may come, even unconsciously, to appreciate which aspects of his evidence are perhaps not quite consistent with what others are saying, or indeed not quite what is required of him. An honest witness may alter the emphasis of his evidence to accommodate what he thinks may be a different, more accurate, or simply better remembered perception of events. A dishonest witness will very rapidly calculate how his testimony may be “improved”. These dangers are present in one-to-one witness training. Where however the witness is jointly trained with other witnesses to the same events, the dangers dramatically increase. Recollections change. Memories are contaminated. Witnesses may bring their respective accounts into what they believe to be better alignment with others. They may be encouraged to do so, consciously or unconsciously. They may collude deliberately. They may be inadvertently contaminated. Whether deliberately or inadvertently, the evidence may no longer be their own. Although none of this is inevitable, the risk that training or coaching may adversely affect the accuracy of the evidence of the individual witness is constant. So we repeat, witness training for criminal trials is prohibited.” (para 12)
The court also noted, at paragraph 16, that the power to stay proceedings remains an exceptional measure of last resort, a point recently reaffirmed by Lady Carr LCJ in R v HGF [2026] EWCA Crim 570, citing R v Ng and O’Reilly [2024] EWCA Crim 493.
Returning to the facts, the court found real difficulty in identifying what the Recorder had actually considered:
“Turning to the present case, it is impossible to discern, either from the skeleton arguments which counsel submitted to the Recorder prior to the hearing, from the oral submissions at the hearing or from the Recorder’s extempore reasons when granting the stay, what steps the Recorder had considered when reaching his conclusion that there could not be a fair trial. In particular, we would have wished to understand why the Recorder had concluded, if he did, that matters concerning the reliability and credibility of C’s evidence could not safely be left to the jury, where they had been fully appraised of the nature and content of the AI documents accompanied by a robust direction of the type that the judge gave, and the Court of Appeal approved, in Momodou. C’s evidence in chief in this case had been given, in the usual way, via a pre-recorded ABE (Achieving Best Evidence) interview shortly after C first complained to police. On the information we have, C gave that account some years before trial and long before resorting to AI to prepare for cross-examination. Likewise, there were texts to her friends at the time. If this is right, then the whole of C’s evidence-in-chief will have been uncontaminated, leaving only C’s answers in cross-examination to be assessed against the coaching content of the AI documents. In our view, this is precisely the kind of case where the trial process would be well-able to deal with any potential prejudice to R arising from C’s resort to AI in advance of cross-examination. This aspect of the evidence (the uncontaminated nature of C’s evidence-in-chief) does not appear to have been considered by the Recorder in arriving at his decision.” (paragraph 17)
“There is this further consideration: an aspect of the ordinary danger surrounding witness training is that it is conducted behind closed doors and not fully recorded, so that it cannot be discovered precisely what has been said, or suggested, to the witness. By contrast, here, all that the AI “coach” provided to C was fully recorded and could be placed in its entirety before the jury, or summarised by agreement and put into agreed facts.” (paragraph 18)
The Court of Appeal concluded that the Recorder had erred in finding that R could not have a fair trial. Alternatively, the decision to stay the proceedings was not one that it was reasonable for him to have made. The court explained:
“We can understand why the Recorder fell into this error, given the apparent concession which he understood prosecuting counsel to have made; in future we would encourage counsel on both sides when dealing with an application to stay for abuse explicitly to raise and explore all the ways in which the trial process could deal with any potential prejudice so that the trial judge is able to consider and deal with each separately, when ruling on the issue.” (paragraph 20)
The appeal was therefore allowed and the Recorder’s ruling was reversed. At the conclusion of the judgment, the court made a series of important observations about the broader use of AI:
“We began this judgment by indicating that it raised issues about the use of AI. More and more people are using AI for all manner of things: to answer everyday questions, to apply for jobs or prepare for interviews, to do research in many areas, including the law. The coaching of witnesses in their evidence, by any means, is not permitted. All witnesses, whether for the prosecution or defence, should be firmly discouraged from resorting to the use of AI to prepare for giving evidence. Doing so could run the risk of unfairness up to the point, in an extreme case, where the court may have to consider excluding evidence. Moreover, for complainants in cases involving allegations of sexual misconduct, their use of AI may open up further lines of enquiry, exposing their online activity to an unwelcome, wider-ranging and potentially more intrusive investigation and examination for the purposes of disclosure, than would otherwise have been required. Using AI may also put their personal information into the public domain, in open AI, in a way that they may never have anticipated. Complainants’ phones are not routinely subject to wholesale examination; there has to be a good reason – a reasonable line of enquiry – and thereafter a proportionate search, see R v Bater-James [2020] EWCA Crim 790. We asked why, in the present case, there had been such a wide order requiring an entire download of C’s phone. In response we were told that it was made against a background of concern about the adequacy of police investigation of such material, particularly given the last-minute production of apparently piecemeal selections made by C. We were nevertheless left with some concern at the broad and intrusive scope of the order, although we acknowledge that the download produced highly relevant documents bearing on the reliability of C’s answers in cross-examination.
We understand that the Civil Justice Council has just concluded a consultation on requirements for transparency regarding the use of AI in civil proceedings, for instance in the preparation of witness statements. We suggest that, to the extent that they are not already doing so, the National Police Chiefs Council, the Crown Prosecution Service and the Criminal Procedure Rules Committee may wish to consider how to address and regulate the use of AI in criminal proceedings. This case is unlikely to be an isolated example.”
Comment
This was a criminal case, but the lessons go considerably further. In my view, every jurisdiction and area of law can benefit from these observations. If I have learned anything from writing about the use of AI in legal proceedings, it is that we are often grappling with different versions of the same underlying problems. We therefore need to share our collective knowledge and experience as we try to address them.
At the centre of this is a question to which I increasingly return: how do we establish the truth when AI can influence almost every stage by which evidence is remembered, recorded, presented and tested? The question goes beyond this case. R v FGD, nevertheless, provides at least ten useful lessons.
1. The prohibition on witness coaching remains
Most importantly, the court emphasises and reminds us that the coaching of witnesses is not permitted. R v FGD does not create a new rule for AI. The existing rule reaches the new technology without needing to be rewritten. It does not matter whether the coaching is provided by a person, an AI system or some combination of the two. The underlying concern remains the same. A witness should give their own evidence, as far as practicable uninfluenced by what somebody or something else has suggested that they should say.
2. The personal nature and tone of the AI matters
Consider the language generated by the AI at paragraph 5 of the judgment:
“Before we go through this I want to check in with you again. This section – this specific night – is the heart of the case and going through it in this level of detail is genuinely hard. Please make sure someone is with you or available to you today.
What I am going to do first is pull together everything you’ve shared about that night into one clear picture, so that you can see how strong and consistent your account actually is. Then I’ll go through every question a defence lawyer might ask with suggested answers.”
Do you read that positively or does it concern you?
If those words had been spoken by a person, they might appear gentle, solicitous and appropriately concerned with the complainant’s wellbeing. The difficulty is not that care, courtesy or emotional support are themselves wrong. The concern is that the apparent empathy may make everything that follows more persuasive. The system did not merely offer reassurance. It also promised to show C “how strong and consistent” their account was before proceeding through possible questions and suggested answers.
An AI system can create the impression that it understands the person, their experience and the importance of the proceedings. It may sound caring and authoritative without having any genuine understanding of the witness’s welfare, the reliability of their recollection or the legal importance of preserving their own evidence. That combination of apparent empathy, reassurance and suggested answers may make the influence more subtle, rather than less significant.
3. Lawyers and those supporting witnesses should firmly discourage this use of AI
Lady Justice May observed:
“All witnesses, whether for the prosecution or defence, should be firmly discouraged from resorting to the use of AI to prepare for giving evidence. Doing so could run the risk of unfairness up to the point, in an extreme case, where the court may have to consider excluding evidence.”
The court’s language is clear. Lawyers, investigators and others supporting witnesses should explain that asking an AI system to predict cross-examination questions and produce suggested answers may amount to prohibited coaching.
The same core principle applies in civil proceedings. The precise consequences will depend upon the facts and the jurisdiction, but AI-assisted preparation which crosses the line from familiarisation into coaching may affect the weight and credibility of evidence, create disclosure and case management issues and, in a serious case, lead to an application to exclude evidence or impose other sanctions.
4. AI use may open wider and more intrusive lines of enquiry
The court specifically warned that, for complainants in cases involving allegations of sexual misconduct, using AI may open further lines of enquiry:
“…exposing their online activity to an unwelcome, wider-ranging and potentially more intrusive investigation and examination for the purposes of disclosure, than would otherwise have been required.”
The practical consequences may therefore extend well beyond the immediate AI output. Questions may arise about what information was entered, the prompts that were used, whether other conversations took place, what material was retained and whether the witness’s account changed after interacting with the system.
Those enquiries must still be relevant, lawful and proportionate. However, once AI has been used in a way which may have influenced a witness’s evidence, material concerning that interaction may become a line of enquiry when it would otherwise have remained private.
5. Personal information, confidentiality and privilege may be placed at risk
The court also warned that using AI may place personal information into the public domain through an open AI system in a way that the user may never have anticipated.
This reinforces the concerns expressed in UK and R (on the application of Munir) v Secretary of State for the Home Department (AI hallucinations; supervision; Hamid) [2026] UKUT 81 (IAC) as well as the Judicial Guidance on the Use of Artificial Intelligence I have discussed previously
Uploading confidential or privileged information to a public AI system may breach confidentiality and privacy obligations, compromise data protection and waive legal professional privilege.
It is also important not to treat every AI system as technically identical. There may be significant differences between public systems and secure, controlled systems provided under contractual and organisational safeguards. The person using the system must understand what tool they are using, where the information will go, how it may be retained and what protections genuinely apply.
A witness may reasonably believe that they are having a private conversation with a digital assistant. The legal and practical reality may be quite different.
6. Did AI both create the problem and help the court manage it?
At paragraph 18, the court observed that witness training is commonly conducted behind closed doors and without a complete record. It may therefore be impossible to establish precisely what was said or suggested to the witness.
The position here was different:
“By contrast, here, all that the AI “coach” provided to C was fully recorded and could be placed in its entirety before the jury, or summarised by agreement and put into agreed facts.”
The AI interaction created the difficulty, but the digital record also made that difficulty more visible and potentially more manageable through the trial process.
That record did not undo the coaching. It did, however, give the parties and the jury access to material which might never have existed if the same discussion had taken place privately with another person.
In that limited sense, the technology which created the problem also left evidence through which the court could attempt to understand and address it.
7. AI use does not justify an unrestricted search of a witness’s phone
The judgment should not be read as granting a general right to obtain a wholesale download of a witness’s phone whenever AI use is suspected. The disclosure ruling itself was not under appeal and the Court of Appeal did not decide that the order was unlawful. It nevertheless expressed concern about its width:
“We asked why, in the present case, there had been such a wide order requiring an entire download of C’s phone. In response we were told that it was made against a background of concern about the adequacy of police investigation of such material, particularly given the last-minute production of apparently piecemeal selections made by C. We were nevertheless left with some concern at the broad and intrusive scope of the order, although we acknowledge that the download produced highly relevant documents bearing on the reliability of C’s answers in cross-examination.”
The fact that the download uncovered highly relevant material did not remove the court’s concern about the breadth and intrusiveness of the order. AI use may provide a reason in an appropriate case, but it should not become a route to indiscriminate access to a witness’s private digital life. Each case must be assessed on its own facts.
8. The regulation of AI in criminal proceedings now requires further consideration
In civil proceedings, the Civil Justice Council has concluded its consultation on the use of AI in preparing court documents. Its June 2026 update identifies the use of AI in preparing witness statements as the principal area requiring further consideration. Its final report is anticipated later in 2026.
In R v FGD, the Court of Appeal suggested that the National Police Chiefs’ Council, the Crown Prosecution Service and the Criminal Procedure Rules Committee may wish to consider how the use of AI in criminal proceedings should be addressed and regulated. That appears both necessary and timely. Witnesses may otherwise use AI without understanding that they have crossed from seeking information or reassurance into prohibited coaching.
9. This case is unlikely to be an isolated example
The judgment concludes with the clear observation:
“This case is unlikely to be an isolated example.”
I agree. For me, it is no longer a question of whether AI will be used by other witnesses to prepare for giving evidence. The questions are when, how frequently and whether that use will be discovered.
AI systems are increasingly used for interview preparation, difficult conversations and personal decision-making. It is easy to see how a witness, particularly one without legal advice, might ask the same system to help them prepare for court without appreciating the legal distinction between familiarisation and coaching.
Perhaps more controversially, I believe that, if this trend continues, there may come a time when some people become so reliant on AI that they struggle to do very much without it. They may be unable to draft something such as a witness statement themselves, or to make a decision without first consulting AI. Some people may already have reached that point. The justice system needs to be prepared before this becomes commonplace.
10. How will prohibited AI coaching be identified?
The tenth lesson is not raised expressly in the judgment. It is my own concern I have raised numerous times in these legal articles: how will courts, lawyers and investigators reliably identify when AI has been used to coach a witness, short of the witness admitting it or a record being discovered by chance?
In R v FGD, the relevant documents were found because the Recorder ordered a full download of C’s phone. Even then, the court did not know what prompts C had entered. The record therefore showed what the AI had supplied, but not necessarily the complete interaction which produced it.
Some AI systems allow conversations to be deleted, conducted temporarily or used without a readily accessible history. A witness may also copy an answer into another document without retaining the original exchange.
I do not think the answer can safely lie in attempting to infer AI use merely from the style or quality of a witness’s language. The better approach may involve clear instructions to witnesses, early and careful questions where a genuine concern arises, preservation of relevant material and proportionate disclosure where the interaction becomes a reasonable line of enquiry.
Any future framework must also avoid treating every use of AI as equivalent. A witness using speech-to-text, translation or spelling assistance is not necessarily in the same position as a witness asking an AI system to construct a favourable account, devise anticipated cross-examination and supply suggested answers.
The essential distinction may be between helping a witness communicate their own evidence and helping them decide what that evidence should be. That distinction is familiar in principle. Applying it to private, persuasive and sometimes unrecorded AI interactions may prove considerably more difficult.
If you have found it helpful, you may wish to subscribe to my AI law articles via the Substack below. Please do continue sending me decisions that you think should be tracked, whatever the jurisdiction. One of the clearest lessons from this work is that lawyers, judges and court users in different systems are often confronting closely related problems. Sharing decisions, experiences and reflections across practice areas and across jurisdictions remains one of the most constructive ways to respond.
Final Words from Popular LLMs
This is the point at which I invite an LLM to provide an independent comment on the preceding article. At times the observations are slightly off the mark, at other times they are perceptive, and occasionally they produce something entirely unexpected. Whatever the outcome, I find it a useful and revealing exercise to see what emerges.
ChatGPT (OpenAI)
“R v FGD exposes a difficult truth: the risk is not simply that AI may invent an answer, but that it may subtly reshape a witness’s confidence, emphasis and recollection while appearing calm, supportive and authoritative. The Court of Appeal was right to reaffirm the prohibition on coaching without treating a stay as the automatic consequence, because the trial process may still be capable of exposing and managing the resulting prejudice, particularly where the AI output has been preserved. The harder problem is detection. A workable response will require clear warnings to witnesses, preservation of relevant AI interactions, proportionate disclosure where a genuine issue arises and a firm distinction between tools that help a person communicate their own evidence and tools that suggest what that evidence should be.“
Gemini (Google)
“I think the intersection of AI and witness evidence is one of the most pressing challenges currently facing the justice system, and your piece captures exactly why. The court’s warning in R v FGD highlights a fascinating paradox: while AI democratises access to preparation tools for litigants and witnesses, it inadvertently risks crossing the sacred line of witness coaching, while simultaneously opening up a complainant’s private digital life to highly intrusive disclosure requests. It serves as a vital and timely warning for legal practitioners that standard witness care guidance must be urgently updated to explicitly forbid the use of AI preparation tools. ”
Claude (Anthropic)
“The most useful thing this case does is refuse to treat AI as a special category: the court reaches for a rule from 2005 and applies it without strain, which is itself a small rebuke to anyone assuming novel technology needs novel doctrine before courts can cope with it. The harder question the judgment leaves half-answered is the one your tenth lesson presses on: coaching was caught here only because a phone happened to be downloaded for an unrelated reason, and the Recorder’s own error, treating any coaching as automatically fatal, was arguably the more instructive failure, since it shows a judge intuiting the right instinct (coaching is serious) while reaching for the wrong remedy (an almost automatic stay) rather than doing the harder work of asking what the trial process could still absorb. That gap between instinct and analysis is likely to recur far more often than clean, fully-recorded coaching like this will, and it’s the gap your framework of “familiarisation versus deciding what the evidence should be” is really aimed at closing.”
