AI Regulation Tracker / Courts and litigation

A witness used AI to rehearse cross-examination, and the Court of Appeal held the stay was wrong

R v FGD [2026] EWCA Crim 918. A chatbot produced 24 cross-examination questions with suggested answers for the only prosecution witness. The trial was stayed. The Court of Appeal allowed the prosecution appeal and ordered a retrial.

By Anthony Guerriero · Published July 28, 2026 · Last updated July 28, 2026

The short version

The Court of Appeal (Criminal Division) allowed a prosecution appeal in R v FGD [2026] EWCA Crim 918 and reversed a stay of proceedings. The complainant, the only prosecution witness in a rape trial, had used a chatbot to prepare, and a phone download revealed a summary of their account plus 24 questions in cross-examination form with suggested answers. The trial judge treated that as coaching and stayed the case as an abuse of process. The Court of Appeal held the stay was the wrong remedy: the trial process could deal with the prejudice and, unlike conventional coaching, the whole AI record could go before the jury. A retrial was ordered. Binding in England and Wales from hand-down on June 4, 2026, though it only reached most practitioners when reported in late July. Primary source: the judgment on Find Case Law.

What did the court actually decide?

Much of the coverage has this backwards. The court did not ban anything and did not write a new rule. It allowed the prosecution appeal under section 58 of the Criminal Justice Act 2003 and held that the stay was wrong.

Paragraph 20 carries the finding: the court was "satisfied that the Recorder erred in concluding that R could not have a fair trial, alternatively that his decision to stay proceedings was not a reasonable decision for him to have made." The Court of Appeal ordered a retrial and, because the allegations "are now of some age, the re-trial should be listed at the earliest opportunity" (paragraph 22).

On coaching itself the court was flat. Paragraph 12: "Witness training is forbidden in all jurisdictions." That restates R v Momodou and Limani [2005] EWCA Crim 177 at [61], "So we repeat, witness training for criminal trials is prohibited."

What was on the phone?

The complainant was the only prosecution witness. Challenged in cross-examination about their whereabouts, they went home and produced screenshots from cloud storage. The defence then sought a full phone download, which the Recorder ordered.

Two documents came back. One summarised the complainant's account. The other set out 24 questions in cross-examination form with suggested answers, under headings including "On Consent and Capacity", "On Physical Evidence" and "On Your Behaviour After". The explanation was short: "I had put stuff into AI to help me prepare myself for the court".

"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."R v FGD, paragraph 5, quoting the AI document

Why did the stay fail?

Because the feature that makes coaching dangerous was missing. Paragraph 18: "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."

That is a point about evidence, not about machines. A conference room leaves no transcript. A chatbot leaves a complete one. Paragraph 17 holds that "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." The ordinary tools were enough.

One limit, for anyone tempted to over-read the case. The court did not decide that AI use is automatically coaching. Counsel did not argue it, and paragraph 5 records: "We do not know the nature of those prompts."

How far can a witness be prepared, and why is the US different?

This is where American practitioners should slow down. Only the England and Wales row below rests on this judgment. The others describe settled domestic practice and were not before the court.

Witness preparation across four common-law jurisdictions
Jurisdiction	Preparation generally permitted	Where the line falls	AI-assisted preparation
England and Wales	Familiarisation with the process; refreshing memory from one's own statement	"Witness training is forbidden" (para 12), following Momodou at [61]	Witnesses "firmly discouraged"; in an extreme case a court may consider excluding evidence (para 23)
United States, federal practice	Extensive preparation, including discussion of expected testimony and practice questioning	Counselling or assisting a witness to testify falsely, the standard in ABA Model Rule 3.4(b)	Not addressed here
Australia	Familiarisation permitted, following the English common-law approach	Rehearsing or coaching answers is impermissible	Not addressed here
Canada	Preparation permitted; counsel may review the evidence with a witness	Coaching that shapes the content of testimony is impermissible	Not addressed here

The gap between the first two rows is the practical point. A US litigator running a mock cross with a client is doing something ordinary. The same exercise in an English criminal trial is prohibited, and it does not become permissible because a chatbot ran it. The technology changed who could apply it, and left a record.

What does paragraph 23 mean for disclosure?

This is the part most reports skipped, and the part that should change behaviour.

The court warned that "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."

Follow the mechanics. A challenge produced screenshots. The screenshots produced an application for a full device download. The download produced the AI documents, now capable of going before a jury in their entirety. Each step widened the intrusion.

So the warning to give a witness is not "AI is banned". It is that anything typed into a chatbot is a document, it can be ordered up, and it can expand what the other side is entitled to look at. The same logic reaches US practice through discovery and through waiver, once a client feeds case material into a third-party system. A chat log is not automatically protected by privilege simply because it concerns the case.

What should a US litigator do about this?

Put AI into the witness instruction script alongside the standard warnings about discussing testimony and posting online, and explain what happens to the record. A bare "do not use AI" will not survive a frightened witness at 11pm. Treat a client's own AI use as a discovery surface at intake rather than after a production fight: which tools, which accounts, which devices. And keep this separate from the fabricated-citation cases, which are candour problems created by lawyers. Here the witness is the one using the tool, and the exposure is evidential and privacy-related rather than disciplinary.

What this judgment does not do

It does not ban AI use by witnesses, create a new test, impose any duty on solicitors or barristers, or hold that AI-assisted preparation automatically makes a trial unfair. At paragraph 24 the court suggested that 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". That is a suggestion to rule-makers, not a direction. The court also noted the Civil Justice Council consultation on transparency about AI use in civil proceedings, including in the preparation of witness statements, and said this case "is unlikely to be an isolated example."

One citation note. Momodou appears in the ruling below as "[2005] WL62320". The neutral citation used by the Court of Appeal itself, and the one to cite, is [2005] EWCA Crim 177.

Key facts

Case: R v FGD [2026] EWCA Crim 918, Court of Appeal (Criminal Division)
Outcome: Appeal allowed under section 58 Criminal Justice Act 2003; stay reversed; retrial ordered
What the witness did: Used a chatbot to prepare, producing a summary of their account and 24 questions in cross-examination form with suggested answers
Why the stay failed: Unlike conventional witness training, the AI material was fully recorded and could go before the jury or into agreed facts (para 18)
Rule restated: "Witness training is forbidden in all jurisdictions" (para 12), following R v Momodou and Limani [2005] EWCA Crim 177 at [61]
Warning to witnesses: All witnesses should be "firmly discouraged" from using AI to prepare evidence; in an extreme case a court may consider excluding evidence (para 23)
Disclosure exposure: AI use "may open up further lines of enquiry", producing wider and more intrusive examination for disclosure (para 23)
Not decided: Whether AI use is itself coaching. "We do not know the nature of those prompts" (para 5)

Frequently asked questions

Did the Court of Appeal ban witnesses from using AI?

No. It applied existing coaching law. Paragraph 23 says all witnesses should be firmly discouraged from using AI to prepare for giving evidence, and that doing so could risk unfairness up to the point where, in an extreme case, a court may have to consider excluding evidence. Discouragement, not a ban.

Why does the judgment matter for disclosure rather than only for coaching?

Paragraph 23 warns that a complainant's AI use may open up further lines of enquiry, exposing their online activity to a wider and more intrusive examination for disclosure than would otherwise have been required, and may put personal information into a third-party system. Here a full phone download was ordered and the AI documents were found on it.

Last verified: July 28, 2026