What did R v FGD decide?

It decided a question about remedy, not about artificial intelligence. A trial had been stopped as an abuse of process after it emerged that the complainant had used an AI tool to prepare for cross-examination. The Court of Appeal (Criminal Division) held that the judge was wrong to stay the trial, allowed the prosecution's appeal, reversed the ruling and directed that the case be relisted for retrial at the earliest opportunity.

Did the Court of Appeal hold that using AI to prepare is witness coaching?

Not in terms. The finding that the material amounted to coaching was made by the trial judge, and the prosecution did not contest it on appeal, so the Court of Appeal proceeded on that basis without deciding it. What the court did say, at paragraph 23, is that coaching by any means is not permitted, that all witnesses should be firmly discouraged from using AI to prepare to give evidence, and that in an extreme case a court might have to consider excluding the evidence. Discouragement and prohibition are not the same thing, and several published summaries have merged them.

Does R v FGD apply to expert witnesses?

The judgment does not mention expert witnesses at all. It concerns a witness of fact in a criminal trial, and its guidance is addressed to witnesses generally, for prosecution and defence. An expert's position is governed by a different framework: in civil work, Part 35 of the Civil Procedure Rules and the duty to the court. Anyone citing this case as authority about experts is extending it beyond what it says.

Does the case apply to civil proceedings?

The judgment says one thing about civil work, at paragraph 24: it notes that the Civil Justice Council has concluded a consultation on transparency about AI use in civil proceedings. That is an observation about someone else's work, not a ruling. The prohibition on witness training in civil cases rests on its own authority and on Bar Council guidance, which existed before this case. Commentary that presents R v FGD as changing the civil position is inferring rather than reporting.

Why did the Court of Appeal think the trial could have continued fairly?

For two reasons given in the judgment. First, the complainant's evidence in chief was a recorded interview given years before the trial and long before any AI was involved, so it was untouched by the later material. Second, at paragraph 18, the court observed that unlike conventional coaching behind closed doors, everything the AI had provided was fully recorded and could be put before the jury in its entirety or summarised by agreement. The record made the problem manageable.

R v FGD: what the Court of Appeal really decided on AI and witnesses

A rape trial at the Crown Court was stopped part-way through when it emerged that the complainant had used an artificial intelligence tool to prepare for cross-examination. The Court of Appeal reversed that decision on 4 June 2026, allowed the prosecution's appeal and ordered the case relisted for retrial. The judgment is R v FGD [2026] EWCA Crim 918, and it is worth reading rather than reading about, because a good deal of the commentary describes a decision the court did not make.

What happened

The complainant was the only prosecution witness. During cross-examination they were challenged about where they had been at a particular time, went home, searched their cloud storage and produced screenshots of their movements. When the Crown applied to recall them to put that material before the jury, the defence asked for a full download of the complainant's phone. The judge ordered it and the trial was adjourned.

The download produced two documents. One summarised the complainant's recollection of the night. The other set out twenty-four questions in the form of a cross-examination, arranged under headings such as consent and capacity and physical evidence, each with a suggested answer. Asked who they had been communicating with, the complainant said they had put material into AI to help prepare themselves for court.

One evidential caveat is worth carrying through the rest of this. The defence did not accept that the documents were AI-generated rather than written by a person, and the appeal proceeded on the assumed basis that they were. The court also recorded, at paragraph 5, that it did not know what prompts had produced them.

What the Court of Appeal actually decided

The appeal was brought by the prosecution and it was about remedy. The question was not whether AI preparation is coaching. It was whether stopping the trial was the right response to it.

The court held that it was not. The judge had erred in concluding that a fair trial was impossible, or alternatively that the stay was not a reasonable decision to make. Leave was given, the appeal was allowed, the ruling was reversed and the case was directed to be listed before the Resident Judge within two weeks, with the retrial to follow at the earliest opportunity given the age of the allegations.

The court was notably gentle about how the judge had gone wrong, saying it understood how he fell into error given the concession he understood prosecuting counsel to have made. That concession is the heart of the case. Prosecuting counsel had not argued that the documents were anything other than some form of forbidden coaching, and the trial judge took the point as conceded.

What it did not decide

The characterisation of the documents as coaching came from the trial judge, who said he was satisfied this was evidence of coaching whether or not it was done by AI or by a person. On appeal, the Crown did not challenge it. So the proposition everyone is quoting was assumed rather than decided, and the Court of Appeal left it undisturbed without examining it.

The guidance the court did give, at paragraph 23, uses two different registers, and the distinction is being lost. Coaching of witnesses by any means is not permitted: that is a prohibition, and it is technology-neutral. Witnesses using AI to prepare to give evidence is something from which they should be "firmly discouraged", with the risk of unfairness running, in an extreme case, to the exclusion of their evidence. Discouragement with a reserved sanction in an extreme case is a materially weaker statement than a rule that AI preparation is coaching, and the judgment does not contain that rule.

The coaching of witnesses in their evidence, by any means, is not permitted. Judgment of the court, paragraph 23.

Nor does the judgment draw the line that practitioners actually want drawn. The word familiarisation appears once in the whole document, inside a quotation from R v Momodou. There is no test offered for where legitimate preparation ends and coaching begins when a chatbot is involved, and no description of what acceptable use might look like. The court did not attempt it, and reading a test into the case is inventing one.

The point most coverage skips

The most interesting paragraph in the judgment is 18, and it cuts the other way from the headlines. Conventional coaching happens behind closed doors and is not recorded. Here, everything the AI had provided was fully recorded and could be placed before the jury in its entirety or summarised into agreed facts. The court treated that as a reason the trial could proceed.

That is worth sitting with. The feature of AI use that made this discoverable at all, that it leaves a complete transcript, is the same feature that made the problem addressable. An unrecorded conversation with a well-meaning friend would have done comparable damage and left nothing for the jury to weigh.

The court's use of Momodou follows the same grain. That case is cited for the prohibition, but it is leaned on hardest for the other half of its reasoning: that difficulties created for the defence, even great ones, are almost always capable of being addressed by the trial process itself. Momodou was a case in which coaching occurred and the trial went on.

Does any of this touch expert evidence?

Directly, no. The word expert does not appear in the judgment. The witness was a complainant in a criminal trial and the guidance is addressed to witnesses of fact, for prosecution and defence. There is no discussion of expert duties, expert reports or AI-assisted methodology, and the case is not authority about any of them.

The reason it is still worth an expert's attention is the principle underneath, which is older than the technology and which experts already work under. Evidence has to be the witness's own. An expert's obligation is stricter rather than looser, because the opinion is the evidence and the duty to the court under Part 35 overrides the interest of whoever is paying. A tool that reads a bundle, extracts dated facts and shows where each one came from is doing work an expert can check line by line. A tool that suggests what the opinion should be is doing something else entirely, and no amount of transparency about it makes it acceptable.

The court's other observations are addressed to a different audience but land in the same place. It suggested that the National Police Chiefs Council, the Crown Prosecution Service and the Criminal Procedure Rules Committee consider how to regulate AI use in criminal proceedings, adding that this case is unlikely to be an isolated example. It also warned complainants that using AI may expose their online activity to a wider and more intrusive investigation than would otherwise have been required, and may put personal information into a system in a way they never anticipated.

What to take from it

If you are handling evidence, three things follow, none of which requires the case to say more than it does.

Warn witnesses early and plainly. The court did not impose a duty to warn or prescribe a form of words, but a witness who has rehearsed with a chatbot creates a disclosure problem, a credibility problem and, on these facts, a phone download. The warning costs nothing.
Keep the record if a tool is used at all. The auditability of the material is what allowed this trial to be salvaged. Work that cannot be reconstructed afterwards is the harder problem, not the easier one.
Do not cite the case for more than it holds. It reversed a stay. It discouraged witnesses from using AI. It did not rule that AI preparation is coaching, it said nothing about experts, and it changed nothing in civil procedure.

That last point is the practical one for anyone writing about this. The judgment is short, public and readable, and the gap between what it says and what it is being said to say is wide enough that a careful reader will notice.

ALLDOQ is built around the same distinction the case turns on. Tooling that reads the record, orders it and attaches a source page to every fact leaves the expert's judgement exactly where it was, and leaves a trail that can be checked. That is the useful side of the line.