Critical Reading: What CPR 32 PD 18.1(5) Already Says About AI in Witness Statements (and Why a Blanket Ban Would Reach Further Than the CJC Intends)

The Civil Justice Council is consulting on new rules for AI in trial witness statements. The debate has largely missed an existing rule already sitting in the Civil Procedure Rules. Practice Direction 32 paragraph 18.1(5) requires every witness statement to disclose the process by which it was prepared. Here is what that already covers, what CRM automation is already caught by it, and where a blanket ban on AI would reach further than the drafters probably intend.

The Civil Justice Council's interim report on the use of AI for preparing court documents landed in March 2026, with the consultation closing on 14 April. For trial witness statements the CJC proposes the strictest position across the four document types it considered: a rule that AI must not be used to generate the content, "including by altering, embellishing, strengthening, diluting or rephrasing the witness's evidence", backed by a declaration to that effect. A final report and any rule changes remain pending.

The public debate on this proposal has focused, understandably, on hallucination, integrity and the well-publicised Ayinde line of cases where fabricated citations reached the court. That is important. It is not, however, where the debate needs to land for witness statement work. The witness statement question is a separate and older problem. And the Civil Procedure Rules already contain a rule that addresses part of it. Any critical reading of the CJC proposal needs to start there.

The rule most of the AI debate is ignoring

Practice Direction 32 supplements Part 32 of the CPR. It has been in force for years. Paragraph 18.1 requires every witness statement to comply with a short list of formal requirements. Sub-paragraph (5) says the statement must state:

"the process by which it has been prepared, for example, face-to-face, over the telephone, and/or through an interpreter."

The examples given are illustrative, not exhaustive. The rule requires disclosure of the actual process used, whatever it was. If the process included AI-assisted drafting, or the use of a case management system that composes first-person text from uploaded records, then a witness statement that says only "prepared in a face-to-face meeting" is at best incomplete, and possibly false. Signing a statement of truth over that carries the ordinary CPR consequences.

Two things follow. First, the compliance point that the CJC proposal is trying to address on trial statements is already partly covered by an existing rule. Second, whatever the final CJC position turns out to be, PD 32 paragraph 18.1(5) applies right now, and applies whether or not the CJC's proposal ever becomes a rule.

That existing rule is the one most claimant firms should be reviewing this quarter.

What claimant-firm workflows actually look like in 2026

Ask any handler at a modern claimant firm to describe how a witness statement is produced today and you will get a description that involves automation the profession has been quietly relying on for years.

The intake interview is often recorded and transcribed by an AI transcription service. The initial draft is composed from a structured intake form completed by a handler, sometimes with the CRM populating standard passages from the case type template. Case management systems auto-fill parties, dates, courts, exhibits and standard formulations. In many firms, the "statement of truth" line, the numbered paragraph headers, the tabulation of exhibits, and the exhibit index are all rendered by the CRM without a person choosing the exact words. AI-assisted proofing tools sit on top of every document. Some firms use retrieval-augmented tools that pull the correct case law or procedural boilerplate. Some now use full first-person generation tools that assemble the witness's account from uploaded records into a coherent statement, which is then verified with the witness before signing.

Very little of this is new. The transcription part has been in use for a decade in some firms. The CRM auto-filling parts is older still. The first-person composition tools are new; the pattern of "case management system as first pass at the document" is not.

That is the practical reality against which the CJC proposal will land, and against which PD 32 paragraph 18.1(5) already sits. The disclosure question is not "did AI touch this document". Almost every witness statement produced by a modern firm has had some form of automation in its preparation chain. The disclosure question is "which parts of the preparation process were AI, and where in the drafting chain did they sit".

Where a blanket ban would bite, deliberately and by accident

The CJC's proposed wording for trial statements would prohibit AI use to generate the content, "including by altering, embellishing, strengthening, diluting or rephrasing the witness's evidence". That is a considered list. Each verb in it targets something meaningful.

Some of what it targets is exactly what most reasonable observers would want targeted. A tool that takes a witness's plain account and dresses it in stronger language, or one that rephrases weaker parts of an account to survive cross-examination, is producing content that is not the witness's evidence, and a signed statement of truth over that is not truthful. That is the case the drafters had in mind.

But the wording reaches wider than that. Consider four scenarios that a competent handler would recognise from any live caseload.

Scenario one. A CRM system auto-populates a standard opening paragraph of a housing disrepair statement with the property address, the tenancy start date and the landlord's name. It always writes "The property I have lived at since [date] is at [address]. My landlord is [name]." A witness who has said all three facts in interview is not being altered, embellished, strengthened, diluted or rephrased by this. But the sentence, as it appears on the page, was generated by a system rather than authored by the witness. Under a strict reading of the CJC wording, that is arguably in scope. Under a sensible reading, it plainly is not.

Scenario two. An AI transcription tool has produced a verbatim record of the witness's account. A handler compiles the statement by pulling the witness's own sentences directly from the transcript into a numbered paragraph structure. The paragraphs are the witness's exact words. The order is the handler's. The compilation is done through a template that the CRM offers. Is that AI-generated content? A strict reading might say the transcript is machine-generated, the compilation is template-driven and neither the witness nor the handler chose the exact document layout. A sensible reading says no, this is transcription and assembly, not generation.

Scenario three. A first-person composition tool assembles a claimant's account from uploaded records and produces a coherent draft. The witness reads it, requests changes, adopts what they agree with, corrects what they do not, and the signed statement reflects the witness's confirmed account of their own facts. This is closer to the pattern the CJC drafters had in mind. But it is also the pattern that most claimant firms will need to use to keep pace with the volume of cases coming through the courts under the current cost regime. A blanket ban here catches a lot of legitimate work.

Scenario four. An AI proofing tool suggests that "on the second of March" should be "on 2 March" for CPR consistency. The handler accepts. That is a rephrasing of the witness's evidence, technically. It is also the kind of format normalisation that has been done by paralegals since the introduction of CPR PD 32 paragraph 19.1(6). Under a strict reading, this is caught. Under a sensible reading, it is not what the rule is trying to catch.

The point is not that the CJC's proposal is wrong. It is that the operative distinction the profession needs to work through is between generation of content and assembly of content, and between altering the substance of the evidence and normalising its form. A rule that treats these as one category will catch things it should not catch, and will lead to firms writing declarations they cannot honestly sign.

The two positions a firm can honestly take

We have been thinking about this in the context of our own product for the last month. There are two positions a firm can defend honestly against the current draft of the CJC proposal.

Position one: honest disclosure. The firm's witness statement workflow uses AI-assisted assembly of the witness's account. That use is disclosed on the face of the statement, as PD 32 paragraph 18.1(5) already requires. The statement of truth carries a preparation-method line that names the tool, describes the assembly process, and confirms that the substance of the evidence is the witness's own. The declaration wording does not attempt to say "AI was not used to generate the content" if that is not true. If the CJC's final rule requires that declaration for trial statements, the firm restricts the use of AI-assisted assembly to non-trial statements, or reserves the tool for a different stage of the drafting process.

Position two: engineered transcription. The firm re-engineers its workflow so that the tool functions as a transcription-and-assembly system. It records the witness's account. It reproduces the witness's own sentences under a structured heading pattern. It does not compose new sentences on behalf of the witness. It does not soften, strengthen or reword. Under this position, the "AI was not used to generate the content" declaration is defensible for a wider range of statements. But the firm accepts that some drafting work will need to sit with a human handler that a more generative tool would otherwise have handled.

Both positions are lawful readings of what the CJC is likely to require. The choice between them is a commercial and workflow judgment for each firm. It is not a technology question. It is a question of where the firm wants to sit on the spectrum between "we let the tool compose the first draft and verify it" and "we let the tool transcribe and structure only, and we compose".

The bigger frame the CJC consultation is part of

Sir Geoffrey Vos, Master of the Rolls and Head of Civil Justice for England and Wales, gave a lecture at the Judicial Institute for Scotland on 30 April 2026 on artificial intelligence and the judiciary. His remarks are worth reading in full for anyone taking the CJC consultation seriously.

Two of his points connect directly to this debate.

The first is his repeated warning against assuming that AI is a passing phase that judges and lawyers can wait out. "For the first year or even two years after GPT3 was launched", he said, "the sceptics were in the ascendancy. AI would never be useful or, more commonly, it would never overcome its tendency to hallucinate case names or facts. In fact, these things have already largely been falsified". A rule that treats all AI use in witness statement preparation as suspect will read, in five years' time, like a rule of its time.

The second is his framing of the justice system's task as "to create, alongside those for whom justice is provided, justice systems that are relevant and robust in the new technological environment. This will not be easy, and we will be required to challenge many of our deeply engrained analogue assumptions". A rule drafted from the assumption that a witness statement is an analogue document produced by an analogue process may catch a great deal of the work firms are already doing to bring the cost of civil litigation to a level ordinary citizens can afford.

The Vos framing does not tell us what the CJC's final rule should say. It does tell us that the framing question matters more than the wording question. What the profession needs is a rule that draws the substance / form line clearly, that recognises the existing PD 32 paragraph 18.1(5) disclosure obligation, and that lets legitimate assembly-and-transcription workflows operate without needing to sign declarations that are not truthful.

Where LegalDocs Assist sits, honestly

We think the product sits on the defensible side of the line, and we think we can say so openly.

Our witness statement drafting engine composes the claimant's first-person account from documents the firm already holds. That composition happens through a hard non-invention rule: the system does not invent defects, exhibits, dates, heads of loss or defendants, and refuses content that goes beyond the supplied case facts. Statements are written as the claimant, with no legal conclusions and no defendant-was-negligent language. The correct statement of truth appears, including the contempt of court wording, plus a preparation-method line in the opening paragraph. Third-party information is attributed. Expert opinion sits in the expert's voice. Case law is drawn from a curated UK knowledge base with citation verification. The system refuses rather than fabricates. That is the direct answer to the Ayinde line of cases.

On the CJC's harder question about content generation, we are honest about the position. Even done faithfully, the tool composes the expression of the evidence from the supplied documents. If a rule requiring a no-AI-generation declaration is adopted for trial statements, we are re-engineering the workflow toward genuine assembly-and-transcription of the witness's supplied words, and adding a configurable declaration block to the output so firms can adopt whichever position their counsel signs off. That work is well underway.

We are also renaming the internal Strength Review feature. It is a coverage check, not an evidence-tuner. It scores how well the draft addresses each pleaded allegation and defence point, and it suggests a question to put to the witness or an exhibit to upload. It does not rewrite the evidence. But the name is a gift to the other side, and it is being changed to something neutral such as Coverage Check.

We think this is the right way to respond to the CJC consultation. Not by pretending the exposure is not there, and not by treating the consultation as an existential threat. By acknowledging where the product's design is defensible, fixing the parts that are not, and being transparent about which is which.

What claimant firms should do this week

Three things.

First, look at your current witness statement workflow with fresh eyes. Assume you had to disclose, on the face of every statement, every automated step in the preparation chain. What would you have to say? If the honest answer includes an AI tool, a CRM auto-fill, a transcription service, or a proofing tool, the disclosure required by PD 32 paragraph 18.1(5) is broader than the standard "prepared in a meeting" line most firms currently use.

Second, decide which of the two positions above your firm wants to take. Honest disclosure with a configurable declaration, or engineered transcription with a no-AI-generation declaration reserved for trial statements. Counsel will need to sign off either way. But the decision is one to take now, not after the final CJC rule lands.

Third, treat this as a design opportunity, not a compliance drag. The firms that get their workflow and disclosure discipline right this quarter will be able to demonstrate, on the face of every statement, that they meet the standard the CJC is reaching for. That is a marketing asset, a client-communication asset and a court-relationship asset. Firms that scramble to update their disclosures after the final rule is announced will not have that.

© LegalDocs Assist — www.legaldocs-assist.co.uk