# Is it witness coaching if a witness uses AI to prepare for cross-examination?

> It can be. In R v FGD [2026] EWCA Crim 918 the Court of Appeal of England and Wales said that “the coaching of witnesses in their evidence, by any means, is not permitted”, and that all witnesses should be firmly discouraged from using AI to prepare for giving evidence. A complainant had put their account into an AI tool and kept cross-examination questions with suggested answers. The court did not stop the trial: jury directions and cross-examination could address the prejudice. The harder problem is detection. AI preparation is found only if it leaves recoverable traces, on a device or in an online account, and an examination that finds nothing does not show that it did not happen. Hong Kong courts apply their own rules; take advice.

Markdown copy of https://d3forensics.com/insights/ai-witness-coaching-forensic-evidence. By Alan Jeffries, Principal Consultant, D3Forensics Limited (Hong Kong BR No. 53603590). Published 2026-11-16. First published on LinkedIn on 2026-08-20. Covers the week of 17 August 2026.

Primary question this page answers: Is it witness coaching if a witness uses AI to prepare for cross-examination?

---

On 4 June 2026 the Court of Appeal (Criminal Division) of England and Wales decided R v FGD [2026] EWCA Crim 918, which reached the wider profession only in late July. It has been read almost entirely as a rule-making case: a witness who uses a generative AI tool to rehearse their own evidence may have been coached, and coaching “by any means” is not permitted. That reading is accurate and incomplete. A prohibition matters only in proportion to the chance of detection, and detection in FGD was a matter of luck.

## What happened

The following is drawn from [the judgment itself](https://caselaw.nationalarchives.gov.uk/ewca/crim/2026/918), published on the National Archives Find Case Law service. The court comprised Lady Justice May, Mr Justice Griffiths and His Honour Judge Picton.

The appeal arose from a rape trial in which the complainant, anonymised as C, was cross-examined about where they had been at a particular time. After court, C went through their cloud storage and produced a series of screenshots showing their movements that night. The prosecution sought to recall C. The defence applied for a download of the device.

The judgment records that “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”. The court described “24 highly pertinent questions and suggested answers”, gathered under headings such as ‘On what happened that night’, ‘On Consent and Capacity’, ‘On Physical Evidence’ and ‘On Your Behaviour After’. Asked who they had been communicating with, C answered: “I had put stuff into AI to help me prepare myself for the court.”

The Recorder stayed the proceedings as an abuse of process. The Court of Appeal allowed the prosecution’s appeal and lifted the stay, holding that a stay is an exceptional remedy and that the ordinary trial process, including jury directions and cross-examination on the recovered documents, was capable of addressing the prejudice. In doing so it restated the principle: “The coaching of witnesses in their evidence, by any means, is not permitted.” It added that “all witnesses, whether for the prosecution or defence, should be firmly discouraged from resorting to the use of AI to prepare for giving evidence”, and that 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 invited the National Police Chiefs’ Council, the Crown Prosecution Service and the Criminal Procedure Rules Committee to consider how to address and regulate AI use in criminal proceedings.

No finding of dishonesty was made against the witness, and none should be inferred. The point of professional interest lies elsewhere.

## The issue behind the headline

Reconstruct the sequence and the fragility of it becomes obvious. C volunteered screenshots. Because C volunteered screenshots, the defence sought a download. Because there was a download, two documents came to light. Remove any link and the AI-generated cross-examination script stays where it was, and the trial proceeds without anyone knowing it existed.

This is not a criticism of anyone involved. It is an observation about visibility. The Court of Appeal has restated a norm whose enforcement depends entirely on forensic recovery, at the moment when the artefacts required for that recovery are becoming less durable and less locally held. Three trends matter:

- Assistants are moving into the operating system. Functionality is migrating from discrete applications, which leave containers and conversation databases, into keyboards, browsers and system-level assistants, which leave far less.
- History is held in the account. Conversation history is increasingly held server side in the user’s account rather than on the handset, so a device extraction alone may recover nothing while the account holds everything.
- Ephemeral modes. Temporary chat modes, marketed as a privacy feature, are designed not to persist history at all.

## Practitioner analysis

If AI preparation is now a line of enquiry, it needs to be treated as one, with a defined scope rather than a hope that something turns up.

Where the material actually sits. In descending order of reliability: exported or saved documents, which is what FGD turned on, and which are the easiest artefacts to recover and the most probative; application containers and local conversation caches; browser history, autofill and service worker caches for web-based assistants; and cloud-side conversation history in the user’s account, which is frequently the only complete record. An examiner who images a handset and reports “no AI artefacts found” without addressing the account has answered a narrower question than the one asked.

Second-order artefacts. Even where content is gone, use may still be evidenced. [Recent published research on iOS notification handling](https://blog.digital-forensics.it/2026/07/after-signal-fix-exploring-apples.html) indicates that one notification database retains application identifiers, contact identifiers and delivery timestamps for roughly sixty days, materially longer than other notification streams, without retaining message content. That can be enough to indicate that an application was in use, and when, after the application itself has been removed. Artefacts of this class are how a pattern of use is established when the substance has been deleted, and they decay on a clock.

Proportionality. England and Wales spent several years building a regime to restrain indiscriminate examination of complainants’ devices, now found in [Chapter 3 of Part 2 of the Police, Crime, Sentencing and Courts Act 2022](https://www.legislation.gov.uk/ukpga/2022/32/part/2/chapter/3) and the statutory code of practice on extraction of information from electronic devices. Nothing in FGD unpicks that, and it should not be read as a licence for wider downloads. The correct response is the opposite of a broader trawl: a narrow, articulable line of enquiry directed at a defined artefact set, justified on the facts of the case and recorded. A targeted request that names what is sought and why is more defensible than a general download.

The negative finding. This is where reports most often fail. An examination that finds nothing does not establish that no AI preparation occurred. It establishes that, within the scope examined, using the tools and versions applied, at the date of acquisition, nothing of that class was recovered. Where temporary chat modes, cloud-only history or intervening device resets are in play, the honest report says so and identifies what could not be examined. An expert who allows “we found nothing” to be read as “it did not happen” has overstated the evidence.

Attribution. Recovering a document containing model-generated text does not identify who created it, on which account, or when. Timestamps, account identifiers, synchronisation records and corresponding application activity are what turn an interesting file into an evidential proposition. In FGD the witness answered the question directly. That will not usually happen.

## What organisations should learn

For litigators and in-house teams, the immediate step is preventative and costs nothing: a short written instruction in the witness pack, given before preparation begins rather than after, telling witnesses in plain terms not to use AI tools to rehearse their evidence, and asking them to disclose it if they already have. The same applies to witness statement preparation in civil proceedings and in arbitration, where the prohibition on coaching is equally clear and the supervisory framework is thinner. Hong Kong courts and tribunals apply their own rules on witness preparation; take advice.

For those instructing experts, the discovery consequences are already visible elsewhere. In Conservation Law Foundation, Inc. v Shell Oil Co. (D. Conn., 18 May 2026), [as reported by Greenberg Traurig](https://www.gtlaw.com/en/insights/2026/7/the-new-litigation-metadata-experts-ai-prompts-may-not-stay-behind-the-scenes), a US Magistrate Judge held that AI prompts used by an expert to reduce a document set formed part of the expert’s methodology and were therefore discoverable. The practical implication is that prompts and outputs are working papers. They should be preserved, versioned and capable of production, and the engagement letter should say which tools are approved, for which tasks, operated by whom.

For evidence custodians and investigators, the scope question is now a standing one. In any matter that turns on a witness’s account, generative AI artefacts belong in the collection scope from the outset, and the preservation notice needs to reach the account as well as the handset.

## Conclusion

The Court of Appeal has drawn a line that most practitioners will accept without argument. The harder question is what happens as the evidence that a line has been crossed stops being written down anywhere.

There is a corollary. An allegation of AI coaching is now available to any party, costs nothing to make, and is difficult to disprove. Answering it requires the same acquisition, the same scope discipline and the same carefully bounded reporting as establishing it. On current trends, the first honest and competent expert report on this subject will be one that says clearly what could not be recovered, and why that silence is not evidence that nothing happened.

See [mobile phone forensics](/services/mobile-phone-forensics) and [expert witness reports](/services/expert-witness-reports).

## Sources

- [R v FGD [2026] EWCA Crim 918, Court of Appeal (Criminal Division), 4 June 2026 (National Archives Find Case Law)](https://caselaw.nationalarchives.gov.uk/ewca/crim/2026/918)
- [ICLR case record: R v FGD](https://www.iclr.co.uk/ic/2026006263)
- [Herbert Smith Freehills Kramer, 24 Jul 2026: A witness's own use of AI to prepare for cross-examination may amount to prohibited witness coaching](https://www.hsfkramer.com/notes/litigation/2026-07/a-witness-own-use-of-ai-to-prepare-for-cross-examination-may-amount-to-prohibited-witness-coaching)
- [Slaughter and May, The Lens, 27 Jul 2026: AI as witness coach? Court of Appeal draws the line](https://thelens.slaughterandmay.com/post/102need/ai-as-witness-coach-court-of-appeal-draws-the-line)
- [Bond Solon: Court of Appeal warns that AI use by witnesses may amount to prohibited witness coaching](https://www.bondsolon.com/news-and-insights/court-appeal-warns-ai-use-witnesses-may-amount-prohibited-witness-coaching/)
- [Society for Computers and Law: Don't prompt the witness: R v FGD](https://www.scl.org/dont-prompt-the-witness-r-v-fgd/)
- [Police, Crime, Sentencing and Courts Act 2022, Part 2 Chapter 3](https://www.legislation.gov.uk/ukpga/2022/32/part/2/chapter/3)
- [Home Office: Extraction of information from electronic devices, code of practice](https://assets.publishing.service.gov.uk/media/634810a38fa8f53463dcb957/E02802691_Electronic_Devices_Code_of_Practice_WEB.pdf)
- [Crown Prosecution Service: Disclosure Manual Chapter 30, Digital Material](https://www.cps.gov.uk/prosecution-guidance/disclosure-manual-chapter-30-digital-material)
- [Greenberg Traurig, 30 Jul 2026: The New Litigation Metadata: Experts' AI Prompts May Not Stay Behind the Scenes](https://www.gtlaw.com/en/insights/2026/7/the-new-litigation-metadata-experts-ai-prompts-may-not-stay-behind-the-scenes)
- [M. Epifani, 28 Jul 2026: After the Signal fix: exploring Apple's notification database](https://blog.digital-forensics.it/2026/07/after-signal-fix-exploring-apples.html)

General commentary on public reports, not legal advice and not a finding about any organisation or person named in those reports.

## Frequently asked questions

### Is it witness coaching if a witness uses AI to prepare for cross-examination?

It can be. In R v FGD the Court of Appeal of England and Wales said coaching by any means is not permitted and that witnesses should be firmly discouraged from using AI to prepare for giving evidence. The harder problem is detecting it, because it is found only if it leaves recoverable traces. Hong Kong courts apply their own rules; take advice.

### Did the Court of Appeal stop the trial in R v FGD?

No. The Recorder had stayed the proceedings, but the Court of Appeal lifted the stay, holding that jury directions and cross-examination on the recovered documents could address the prejudice. It made no finding of dishonesty against the witness.

### Where would evidence of AI preparation be found?

In saved or exported documents, application data and local caches, browser artefacts, and conversation history held in the user's online account, which is often the only complete record. Some system records may show that an application was used, and when, even after it has been removed.

### If an examination finds no AI artefacts, does that mean none were used?

No. It means nothing of that class was recovered within the scope examined, with the tools used, at the date of acquisition. Temporary chat modes, cloud-only history and device resets can all leave nothing to find, and a careful report says what could not be examined.

### What should lawyers tell witnesses about AI?

Give a short written instruction before preparation begins, telling witnesses not to use AI tools to rehearse their evidence and to say if they already have. Experts' own AI prompts and outputs should be treated as working papers and preserved.

## Contact

Telephone +852 5808 1071; email info@d3forensics.com; case review form https://d3forensics.com/#contact
