Can AI Prepare You for Cross-Examination? The Court of Appeal’s Warning on Witness Coaching
JSH Law | Legal Technology · Evidence · Family Justice
By Jessica Susan Hill | England and Wales | Law and sources checked 20 October 2026
There is a serious difference between using AI to understand court procedure and allowing it to suggest what a witness should say about disputed facts.
In R v FGD [2026] EWCA Crim 918, a criminal trial was halted after a witness had used AI to prepare for cross-examination. The Court of Appeal reversed the stay, but its warning about witness coaching is of real significance to legal practitioners, witnesses and litigants in person.
This article explains the case, the legal distinction between helpful procedural support and improper coaching, and the particular risks for evidence in the Family Court.
1. What actually happened in R v FGD?
The Court of Appeal (Criminal Division) handed down judgment in R v FGD [2026] EWCA Crim 918 on 4 June 2026. The judgment concerns a criminal rape trial. It is not an October decision, even though renewed reporting has drawn attention to it this month. The court took care to anonymise the parties and exclude identifying details.
The complainant, referred to as C, was the sole prosecution witness. During cross-examination, a question arose about where C had been on a particular night. After additional screenshots were supplied, a court-ordered download of C’s phone revealed a summary account and 24 proposed cross-examination questions with suggested answers. C said they had put material into an AI system to help prepare for court.
The defence disputed whether the documents had actually been generated by AI rather than another person. The appeal proceeded on the assumption that they were AI-generated. The precise prompts were unknown, and the judgment does not establish which product produced them.
The trial judge regarded the material as witness coaching and stayed the prosecution on the basis that a fair trial would be impossible. The Crown appealed. The Court of Appeal allowed the appeal and directed a retrial.
2. What the Court of Appeal really decided
The appeal did not succeed because AI coaching was acceptable. It succeeded because a stay of criminal proceedings is a remedy of last resort, and the Recorder had not adequately considered whether the ordinary trial process could manage the possible unfairness.
Three distinctions were central. First, C’s original Achieving Best Evidence (ABE) interview had been recorded years before the AI interaction, leaving the possibility of assessing earlier evidence separately from later answers. Second, the content of the AI-generated documents had been recorded and could be scrutinised. Third, the criminal trial process could potentially address the issue through questioning, appropriate evidential rulings and directions to the jury.
Paragraph 23 contains the central warning: witnesses should be firmly discouraged from using AI to prepare for giving evidence, and coaching by any means is not permitted. The court identified risks to trial fairness, the potential for extensive and intrusive digital enquiries, and the possibility that sensitive personal information might be placed into an open AI system.
It would therefore be inaccurate to report that the Court of Appeal approved AI witness coaching or that it declared all AI-assisted documents inadmissible. The decision is about the conduct, its possible evidential consequences and the proportionality of the court’s response. It did not decide whether the allegation at the centre of the criminal trial was true.
3. Witness familiarisation versus witness coaching
Explaining how a court hearing works is not the same as suggesting what a witness should say about disputed facts. A person may properly learn about where to sit, what cross-examination means, how to ask for clarification or a break, and why it is important to acknowledge genuine uncertainty.
The danger arises when an adviser or AI tool tells the witness how to respond to specific allegations, documents or inconsistencies. The court referred to the principles explained in R v Momodou and Limani [2005] EWCA Crim 177. The concern is that a witness may consciously or unconsciously change their recollection to match what they believe is expected.
| Procedural assistance — generally lower risk | Fact-specific coaching — high risk |
|---|---|
| Explain what examination-in-chief and cross-examination mean | Generate persuasive replies to the actual allegations |
| Provide a blank chronology template | Fill missing memories with plausible details |
| Explain that a witness can say they do not remember | Create a confident answer when the witness is uncertain |
| Proofread for errors without changing meaning | Add factual claims, quotations or emotional descriptions not provided by the witness |
| Discuss available participation measures | Rehearse answers to the precise evidence or exhibits in the case |
The comparison is an editorial guide, not a substitute for examining the facts and applicable procedural rules. Even proofreading requires close checking: AI can alter meaning while apparently improving grammar. A witness statement is not supposed to be a perfectly optimised narrative.
4. What this means for Family Court proceedings
R v FGD was a criminal appeal; it does not itself amend Family Procedure Rules or dictate the appropriate remedy for a family judge. Nevertheless, its underlying concern about independent testimony is highly relevant to disputes over child arrangements, allegations of domestic abuse, social-care records, contested witness statements and financial disclosure.
Imagine a parent alleging a pattern of coercive behaviour. They have original messages, contemporaneous notes and a genuine, if incomplete, recollection. A carefully used tool could help arrange dates or build an evidence index, subject to confidentiality. But asking AI to produce the strongest possible answers to questions about why they remained in contact risks transforming their account into a product of suggested testimony.
The concern applies equally to a parent denying allegations, a supporting family member and a professional witness. AI should not turn an allegation into an established fact, infer someone’s motivation as if the witness observed it, or replace a lack of memory with a convenient explanation.
Where domestic abuse is relevant to child arrangements, Practice Direction 12J provides essential safeguards concerning the identification and treatment of abuse and risk. A fluent AI draft cannot replace the court’s assessment of evidence, fairness or the child’s welfare.
5. Witness statements must remain the witness’s own evidence
The relevant family evidence framework includes Family Procedure Rules Part 22 and Practice Direction 22A. Under paragraph 4.1 of PD22A, a witness statement must, if practicable, be in the maker’s own words and expressed in the first person. Paragraph 4.3 requires identification of matters within personal knowledge and the sources of matters of information and belief.
Those requirements do not forbid every form of administrative or drafting support. They do demand that a statement should accurately reflect what its maker knows and believes, rather than incorporate invented facts, changed levels of certainty or suggested explanations the witness cannot personally adopt.
The Family Procedure Rules Part 17 also require witness statements and other specified documents to be verified by statements of truth. PD22A sets out the witness-statement declaration and refers to potential contempt consequences for false statements made without honest belief.
A witness cannot outsource responsibility for the factual accuracy of their evidence to an AI assistant, a legal adviser or a drafting service. Equally, the simple fact that a computer assisted with formatting does not mean a whole statement must be excluded. The court will consider the nature of the involvement, whether it materially influenced the testimony, the original evidence and the appropriate procedural response.
6. The hidden disclosure and data-protection risk
One reason R v FGD is so instructive is that the digital material was preserved and later surfaced through a court-ordered phone download. The AI-generated suggestions could therefore be compared with the witness’s testimony. However, the Court of Appeal also expressed concern about the breadth of the phone download and referred to the importance of reasonable, proportionate enquiries.
This is not a new entitlement for the other party in Family Court proceedings to demand an entire phone, every AI chat or unrestricted access to cloud records. Relevant disclosure duties and court powers depend on the nature of the proceedings, confidentiality, proportionality, privilege where applicable and the actual directions made.
Nevertheless, if an AI conversation contains suggested answers to disputed facts and may have affected a statement or oral evidence, it can become relevant to an evidential disagreement. A litigant should not fabricate, alter or destroy potentially relevant records. They should also not assume that the answer is to circulate their entire private chat history without professional guidance.
In children proceedings, Practice Direction 12G and applicable court orders govern restrictions on communicating case information. Protected addresses, child health and education records, safeguarding concerns and material about third parties may be highly sensitive. Where data-protection law applies, the ICO’s guidance on AI and personal data explains relevant obligations. The fact that a chatbot accepts an upload is not permission to disclose the material.
7. Trauma-informed hearing preparation without scripting answers
It would be harmful to suggest that a witness who is frightened, disabled, traumatised or unfamiliar with court simply has to cope alone. Procedural familiarisation, interpreters, reasonable adjustments, special measures and support from appropriate services can be crucial to fair participation.
The solution is not to have a chatbot generate polished versions of answers to uncomfortable questions. In disputes involving coercive control, for example, the person may need help explaining the chronology and linking it to genuine source evidence; they should not be encouraged to replace the difficulties of recalling trauma with an AI-generated narrative of certainty.
The other danger is excessive suspicion. Not every AI user is dishonest, and use of AI for a benign procedural question is not proof of contamination. Any inquiry should be evidence-led, focused on what the tool actually did, and proportionate to the issue in dispute.
8. Eight practical steps for litigants in person
- Understand the process, not suggested testimony. Ask about what happens at hearings, not what you should say about your disputed case.
- Preserve the sources. Keep original notes, messages, files and earlier statement drafts without altering them to match later text.
- Separate personal knowledge and other information. Mark what you witnessed, what somebody reported and what is a disputed allegation.
- Check every factual sentence. Do not adopt dates, quotations or certainty introduced by a machine.
- Use templates cautiously. A blank evidence register can help, but an AI-generated witness account may produce major problems.
- Protect confidential case material. Do not upload identifiable children’s records to unsuitable public AI services.
- Preserve potentially relevant AI interactions. If there is a genuine concern about contamination, seek advice before editing, deleting or disclosing those records.
- Seek proper participation assistance. If the hearing is overwhelming, consider the correct route to court directions, safeguards and regulated advice.
Someone who simply asked ChatGPT what a fact-finding hearing means should not assume they have committed witness coaching. The material distinction is whether a tool supplied or shaped answers to the actual disputed facts.
9. What legal teams and support services should change
Lawyers, charities, clinics and non-reserved court-support providers should make the witness-familiarisation boundary explicit. It is sensible to explain how cross-examination operates and ensure a witness understands procedural rights. It is not sensible to generate convincing answers for them to repeat.
- Establish which AI systems, if any, are approved for sensitive client or court documents.
- Keep a source trail that identifies which evidence supports each factual statement.
- Check that proofread edits do not add details, certainty or interpretations the witness did not express.
- Do not run case-specific mock questioning that scripts anticipated answers.
- Identify participation needs and safeguarding concerns early.
- Where AI may have influenced testimony, evaluate the actual interaction and seek regulated advice about any necessary steps.
For JSH Law, responsible non-reserved support includes chronologies, evidence indexes, organisational assistance, document structure and practical court preparation. It does not include coaching someone to deliver predetermined factual testimony.
10. Frequently asked questions
Can I use ChatGPT to proofread a Family Court witness statement?
Proofreading and formatting may be permissible if the meaning remains unchanged, the witness verifies every fact, and the material is handled securely. It is important to check whether the tool subtly added assertions or changed the tone of certainty.
Is AI coaching now permitted because the prosecution appeal succeeded?
No. The appeal was about the disproportionate remedy of staying the prosecution. The court expressly discouraged using AI to prepare factual testimony.
Does the judgment automatically control Family Court witness statements?
No. It is a criminal appellate decision with relevant reasoning, but family judges apply the Family Procedure Rules, the court’s evidential powers and the circumstances of each case.
Can the other parent demand to see all my ChatGPT conversations?
No blanket entitlement follows from this case. Relevance, legal power, proportionality, confidentiality and any court order all matter.
What if an AI tool inserted a false fact into a signed statement?
Do not repeat or rely on something you know is inaccurate. Keep the original records and obtain appropriate legal advice about correcting the matter consistently with any directions.
11. Legal sources and related JSH Law reading
Primary sources checked on 20 October 2026
- R v FGD [2026] EWCA Crim 918, 4 June 2026, especially paragraphs 5–6, 12 and 16–23
- Practice Direction 22A — Written Evidence
- Family Procedure Rules Part 22 — Evidence
- Family Procedure Rules Part 17 — Statements of Truth
- Practice Direction 12J — Domestic Abuse and Harm
- Practice Direction 12G — Communication of Information
- Judiciary Position Statement on AI (9 October 2026)
- ICO: AI and Data Protection
© 2026 JSH Law Ltd. All rights reserved.
© 2026 JSH Law Ltd
