AI witness coaching derails UK rape trial: Lessons for Australian practitioners 

Part-way through a rape trial at Lewes Crown Court a complainant was ordered to produce her phone contents.¹ The download produced two documents: a summary of the complainant’s account, and 24 questions in the form of a cross-examination, each with a suggested answer. 

The complainant said she had put material into AI to prepare for court. 

The chatbot worked with the complainant to determine “how strong and consistent your account actually is”, then supplied 24 questions a defence lawyer might ask with accompanying suggested answers drawn from her statement. 

The trial judge found this was coaching and stayed the proceedings. The jury was discharged.² 

The Court of Appeal (Criminal Division) reversed the stay, ³ finding that on the facts of the case the fault could be remedied.⁴ The core point remains, however. Witnesses using AI to write statements or to prepare for cross-examination risk contamination of their evidence.  

Nothing about the complainant’s use of AI should be a surprise. A stressed lay person, facing a barrister’s questions, asked a free tool for help the night before. AI use by litigants (even represented ones) is now close to ubiquitous. 

As I have written previously in Proctor, professional competence now requires us to anticipate AI use and to understand of what AI does badly, whether or not you use it yourself.⁵ This understanding is important to protect our client’s interests and also to protect the administration of justice. 

Three features make AI tools a particular threat to evidence:  

• The first is sycophancy. These tools are tuned to validate, not argue. A witness who asks whether their account stands up will be told that it does and potentially given a script to sound more convincing. Unless prompted correctly, the AI will not seek to establish the truth it will look for ways to present a seamless narrative. 

• The second is creation or reinforcement of false memory. In a 2024 MIT study, participants shown a crime video who then discussed their observations with a generative chatbot formed more than three times as many immediate false memories as a control group, and a week later those memories persisted with confidence intact.⁶ 

• The third is our old friend, hallucination. If asked to prepare a claim or witness statement the tool may draw from a generic library of past examples, presenting other people’s facts blended with your own. It will probably be an excellent statement – but it won’t necessarily be true. 

Is robo-coaching really such a problem? 

Bell CJ of the NSW Supreme Court has put the principle plainly: improper reformulation of a witness’s evidence is unacceptable when a lawyer does it, and should not be permitted with AI’s assistance either.⁷ 

Rule 24.1.2 of the Australian Solicitors Conduct Rules prohibits coaching a witness by advising what answers to give. Rule 25.1 prohibits conferring with more than one lay witness at a time on contentious issues,⁸ the purpose being to stop witnesses being encouraged, deliberately or through inadvertence, to speak with one voice.⁹ 

An AI tool handed several witnesses’ accounts and asked to produce statements will possibly harmonise them, because coherence is what it has been optimised for. “Should I be doing this” is not a question it is likely to be asking itself unprompted. 

An AI tool with access to one statement may well use it to prepare the text of another in the same matter. Tools such as Copilot often have extensive data rights within a firm environment, usually being able to “see” everything the user has access to. A practitioner who has sent witness A their proofed statement via email may find shadows of that document when drafting an affidavit for witness B. 

Although the rule was not drawn with AI in mind that is the mischief rule 25 seeks to prohibit, occurring without a conference and without anyone deciding to do it.  

The critical importance of evidence integrity

No Australian court has yet dealt specifically with a witness who rehearsed with a chatbot.  

On the adjacent question of written evidence, all jurisdictions are unified that evidence integrity is critical. New South Wales prohibits generative AI in generating the content of affidavits, witness statements and character references, or in “altering, embellishing, strengthening, diluting or rephrasing” a witness’s evidence, and requires a statement of non-use¹⁰; The Federal Court compels disclosure where AI has summarised or analysed material a witness relies on; Victoria and the Federal Circuit and Family Court require the document to reflect the person’s own knowledge and words.¹¹ 

Queensland’s Practice Direction 5 of 2025 goes only to the accuracy of references in submissions,¹² but Guidance Statement #37 (at paragraphs 4.3 and 4.4 stress the importance of evidence integrity and being in a position to explain how AI was used and accuracy controlled. 

Practical solutions 

Telling clients not to use AI is unlikely to solve the problem. They do need to understand the risks, though. QLS has draft materials suggesting how to address other points such as privilege and hallucination. Add the problem that evidence may be disallowed or less persuasive if AI use becomes apparent.  

For an example, see the UK case of Godwin v Godwin¹³ in which the judge discounted a litigant’s evidence because he was unconvinced it had not influenced their testimony. 

Critical points to raise:  

• Your evidence must be your own recollection in your own words;  

• Do not ask a chatbot what to say or what you will be asked;  

• Do not put material from the file into an AI tool.  

And do not stop at text: photographs, recordings and documents may have passed through editing tools, filters and enhancement features. In all likelihood that did not alter the substance, but if called upon to address such issues it will be important for you to be in a position to do so. 

Consider your own use within the firm. Running a proof through a chatbot to tidy the expression risks a document that is no longer the deponent’s account, so keep the first draft in case it is necessary to demonstrate that no fabrication was injected during the review process. 

Footnotes
¹ R v FGD [2026] EWCA Crim 918, [3]-[5].
² Ibid [8]-[9].
³ Ibid [12], quoting R v Momodou [2005] EWCA Crim 177, [61]. That passage was applied in Australia in Majinski v Western Australia (2013) 226 A Crim R 552; [2013] WASCA 10, [29] (Martin CJ).
⁴ R v FGD [2026] EWCA Crim 918, [16]-[18].
⁵ David Bowles, ‘Your obligations when AI is used by others’, Proctor (online, 18 December 2025); Queensland Law Society, Artificial Intelligence in Legal Practice (Guidance Statement No 37, 31 May 2024) [4.1].
⁶ Samantha Chan et al, ‘Conversational AI Powered by Large Language Models Amplifies False Memories in Witness Interviews’ (arXiv Preprint No 2408.04681, 8 August 2024).
⁷ Chief Justice A S Bell, ‘Change at the Bar and the Great Challenge of Gen AI‘ (Speech, Australian Bar Association Conference, Sydney, 29 August 2025) [46]-[47].
⁸ Australian Solicitors Conduct Rules (Qld) rr 24.1.2, 25.1.
⁹ Day v Perisher Blue Pty Ltd (2005) 62 NSWLR 731, 746; Day v Perisher Blue Pty Ltd [No 2] [2005] NSWCA 125; Legal Practitioners Complaints Committee v Caine [2010] WASAT 178, [142]-[143]; Legal Services Commissioner v Hansen [2008] LPT 7, [26] {Note, the practitioner was exonerated. Counsel had spoken to one witness in proximity to another but there was no evidence that they could hear each other. The matter had also been delayed by over three years, a punishment in itself}.
¹⁰ Supreme Court of New South Wales, Use of Generative Artificial Intelligence (Gen AI) (Practice Note SC Gen 23, 28 January 2025) [10]-[13]; the same requirements now sit in the NSW Uniform Civil Procedure Rules rr 31.4(3A)-(3C), 35.3B.
¹¹ Federal Court of Australia, Use of Generative Artificial Intelligence (Practice Note GPN-AI, 16 April 2026) [3.1], [4.7]-[4.12]; Supreme Court of Victoria, The Use of Artificial Intelligence by Court Users (Practice Note SC Gen 25, 14 May 2026) [6.10]-[6.11]; Federal Circuit and Family Court of Australia, Use of Artificial Intelligence (Practice Direction, 29 May 2026) [4.9].
¹² Supreme Court of Queensland, Accuracy of References in Submissions (Practice Direction No 5 of 2025, 24 September 2025); Queensland Courts, The Use of Generative Artificial Intelligence (AI): Guidelines for Responsible Use by Non-Lawyers (15 September 2025).
¹³ Godwin v Godwin [2026] EWHC 923 (Ch), [44]-[48].
¹⁴ See, for example, Wisconsin v Kyle Rittenhouse where a “zoomed in” Ipad image was excluded from trial evidence and the technical discussion here: | Science and Technology Law Review