AI and witness evidence require a different discussion from AI and legal research. The danger is not simply that a model may invent a case or misstate a proposition. The deeper concern is that a generative system may reshape, embellish or homogenise what is supposed to be the witness’s own evidence.
That is why the most useful dividing line is between organising material for the lawyer’s investigation and generating the witness’s account.
1. PD57AC begins with personal knowledge and the witness’s own words
Practice Direction 57AC applies to trial witness statements in the Business and Property Courts within its scope. It requires the statement to address facts within the witness’s personal knowledge that are relevant to the case, to identify documents the witness has referred to or been referred to for preparing the evidence, and to comply with the requirement that the witness statement be in the witness’s own words where practicable.
The accompanying Statement of Best Practice also recognises that memory is fluid and malleable, and that the process of preparing evidence should avoid practices that may alter or influence recollection.
2. AI can be useful before the witness account is written
There are legitimate organisational tasks around witness preparation. A legal team may need to identify contemporaneous emails involving the witness, build a timeline of meetings, locate versions of a document, find where the witness appears in the disclosed record, or identify inconsistencies that require careful questioning.
Those are evidence-management tasks. They can help the lawyer prepare a focused, non-leading interview and understand the documentary context.
3. Evidence organisation is not evidence generation
The line becomes more sensitive when a model is asked to draft or improve the factual account itself. A language model is designed to produce coherent language. Coherence can be exactly the wrong objective where the witness’s true recollection is partial, uncertain, awkwardly expressed or inconsistent with another source.
A professionally useful system should make those uncertainties visible rather than silently repairing them.
4. The CJC is treating witness statements as a distinct problem
The Civil Justice Council’s 2026 work on AI in court documents has distinguished ordinary professional legal drafting from witness evidence. Its June 2026 update reported broad consensus that additional AI-specific formal requirements were not currently necessary for professional pleadings, advocacy documents and similar drafting, while witness statements remained an area for further consideration because of authenticity, integrity and personal recollection.
As at 22 September 2026, that work remains in progress. The CJC material should therefore be described as an emerging direction and consultation process, not as a new binding rule.
5. Keep three evidential layers separate
A matter workspace should preserve the distinction between:
- what the contemporaneous document records;
- what the witness now recalls;
- what the system or lawyer infers from the combination.
Those layers may support the same conclusion. They may also diverge. A witness may honestly remember a meeting differently from the written record; an email may show that the witness received a message without proving they read it; metadata may establish a timestamp without establishing the witness’s state of mind.
6. A safer AI-assisted witness workflow
- Map the relevant issues before the interview.
- Identify contemporaneous documents without assuming they define the witness’s recollection.
- Use open, non-leading questions on important contentious matters.
- Record the witness’s own account and uncertainty faithfully.
- Keep documentary propositions separate from recollection.
- Where AI assists with organisation or drafting mechanics, ensure the final content does not go beyond the evidence actually obtained from the witness.
- Record the preparation process where the applicable rules require it.
7. Do not convert documentary narrative into personal recollection
PD57AC itself warns against using witness statements to take the court through the documents or to provide argumentative commentary on other evidence. That is an important discipline for AI-assisted work. A model can quickly construct a polished documentary narrative; that does not make the narrative the witness’s memory.
8. Document exposure can itself affect recollection
PD57AC’s Statement of Best Practice is careful about the documents shown to a witness because memory can be influenced by later material. AI introduces an additional layer: a system may summarise several documents into a narrative before the witness is interviewed. Showing the witness that synthetic narrative may influence recollection just as surely as showing a curated bundle of source documents.
The safer approach is to preserve the distinction between the lawyer’s documentary analysis and the witness’s own recollection. Where documents are used to refresh memory, the applicable professional and procedural rules should govern how that occurs.
9. Interview support should avoid synthetic answers
AI may help prepare a list of topics, identify chronological gaps or organise the documents that need to be considered. It should be used carefully if it begins proposing the factual answer to a question before the witness has given their own account.
For important contentious matters, open questions and accurate contemporaneous records of the witness’s responses remain central. The tool can assist the interviewer; it should not become the unseen author of the evidence.
10. Repeated redrafting can create a false sense of certainty
The Statement of Best Practice warns that repeatedly revisiting a draft may corrupt rather than improve recollection. Generative editing makes repeated redrafting extremely easy. Each pass may make the prose cleaner while moving it further from the witness’s original language or uncertainty.
A controlled workflow should therefore preserve the evidence obtained from the witness and treat language editing as subordinate to that record.
11. The preparation process may itself matter later
Where the applicable rules require the process of preparation to be described, teams should be able to explain what role technology played. That does not mean every use of spell-checking or administrative assistance needs elaborate disclosure. It means the process should be sufficiently understood that the legal representative can comply honestly with the governing requirements.
This is another reason to avoid invisible AI use. If the team cannot reconstruct how the witness’s words became the final statement, professional review becomes difficult precisely where evidential integrity matters most.
12. A useful system should preserve the witness/document boundary throughout the matter
The distinction should not disappear after the statement is signed. When the legal team later reviews an issue, it should still be possible to tell whether a proposition comes from contemporaneous documentary evidence, the witness’s recollection or an analytical inference.
This becomes particularly important at cross-examination and trial preparation. A witness may be challenged with a document that appears inconsistent with the statement. If the matter workspace has already separated those evidential layers, the lawyer can investigate the conflict without confusing the system’s earlier summary with the witness’s own account.
Preserving that boundary also improves professional review. The system can assist in mapping what evidence exists around the witness while leaving the ultimate assessment of credibility, recollection and significance to the lawyer and court.
Conclusion
AI can help lawyers find and organise the material around a witness. It should not make the boundary between the documentary record, the witness’s own recollection and analytical inference disappear. In witness evidence, preserving that distinction is part of preserving the evidence itself.
Related LegalRAG Pro Insights
Professional context. This article discusses legal-technology workflow and professional-risk questions. It is not legal advice and should not be treated as a substitute for checking the current procedural, regulatory and factual position in a particular matter.