Part of the AI Regulation News hub.
A High Court judge approached two witnesses' evidence with caution because he could not be sure their statements were in their own words after both had run drafts through ChatGPT
The witnesses said a digital assistant had helped with grammar, spelling and presentation. The judge observed that people do not ordinarily call an AI tool a digital assistant, could not rule out that ChatGPT had supplied the phrase itself, and had not been shown the first drafts.
Bottom line: No sanction and no exclusion. The consequence was evidential: the judge could not be sure the statements were in the witnesses' own words, which reinforced his decision to approach the defendant's evidence cautiously and required him to approach his former wife's evidence cautiously too.
Who this affects: Litigators preparing witness evidence in England and Wales, particularly in the Business and Property Courts where PD57AC applies, and any party or litigant in person using an AI tool to tidy a draft statement.
Issue date: Heard over several days; judgment dated 22 April 2026.
What changed: No new rule. The judgment applies the existing PD57AC machinery, under which non-compliant statements risk exclusion or reduced weight, to a statement that passed through an AI tool.
Analysis: The reasoning does not depend on anything being fabricated. Nothing in the statements is found to be invented. The problem is narrower and harder to cure: a witness statement is supposed to be the witness's own account, and once a model has been in the loop the court may be unable to satisfy itself of that, particularly where the first drafts are not produced.
Primary sources: Judgment, [2026] EWHC 923 (Ch), Find Case Law
- Case
- William Quentin Godwin v Jason Peter Godwin, in the matter of Michael Godwin deceased
- Neutral citation
- [2026] EWHC 923 (Ch)
- Court
- High Court of Justice, Business and Property Courts in Leeds, Property, Trusts and Probate List (ChD)
- Judge
- His Honour Judge Klein, sitting as a High Court judge
- Representation
- The Claimant appeared in person. The Defendant appeared in person.
- Jurisdiction
- England and Wales
- Rules engaged
- CPR Part 32 and Practice Direction 57AC, including the statement of truth and the certificate of compliance required by paragraph 4.1 of PD57AC
- Status
- Judgment delivered 22 April 2026.
- Bindingness
- A first-instance judgment. Binding between the parties; persuasive on the treatment of AI-assisted witness statements.
- Editorial Note
- Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.
- Primary source
- https://caselaw.nationalarchives.gov.uk/ewhc/ch/2026/923
The rule the statements had to satisfy
This was a probate dispute between two brothers, including a contested question about where their father should be buried. Both parties appeared in person.
An earlier order of District Judge Bond had set out precisely what the witness statements had to do. They had to be prepared in accordance with CPR Part 32 and Practice Direction 57AC, contain a complete statement of truth in proper form, and include a signed certificate of compliance in the form required by paragraph 4.1 of PD57AC, which the District Judge had set out in full.
The order also spelled out the consequence of failure. Non-compliant witness statements were at risk of being excluded from the evidence or of having reduced weight attributed to them.
That framework matters, because it means the court did not need any new AI rule to reach the result. The existing regime already ties the admissibility and weight of a statement to whether it is properly the maker's own account.
What the witnesses disclosed, and what the judge made of it
The defendant and Lucie Godwin, his former wife who gave evidence for him, said in their witness statements that they had limited assistance from a digital assistant for grammar, spelling and presentation. By digital assistant they meant an AI tool, ChatGPT.
Their account was specific: each had prepared the first draft of their own statement, each had uploaded that draft to ChatGPT, and ChatGPT's product did not add, remove or re-arrange any words compared with the first draft.
The judge was not persuaded there was any need for it. He found there was no good reason for either of them to use ChatGPT in this case. Both struck him as sophisticated people, the defendant being a key account manager who travels in Europe giving presentations for a company in the lift industry, and both were technically adept enough to have used the spelling and grammar checking facility in the word processing software they had used for their first drafts. The District Judge had in any event set out in detail how the statements should be structured.
Then the observation that carries the reasoning. It is not conventional, in spoken English, to refer to an AI tool as a digital assistant. The judge received no satisfactory explanation for the choice of phrase, and said he could not rule out either that the phrase was one ChatGPT had contributed to the statements, or that the defendant might have been trying to obscure his use of AI in producing his statement.
He also records, in a single line that does a great deal of work: he had not been provided with the first drafts.
The consequence, which is about weight rather than punishment
The conclusion follows directly from the certification requirement. For all those reasons, the judge could not be sure that the defendant's or Ms Godwin's witness statements were in their own words. That reinforced his decision to approach the defendant's evidence cautiously and required him to approach Ms Godwin's evidence cautiously too.
Nothing was struck out and nobody was sanctioned. There is no finding that anything in either statement was untrue or invented, which distinguishes this from the fabricated-citation and fabricated-exhibit cases entirely.
The judgment also shows the ordinary operation of the same rules on a different statement, for comparison. A third witness's statement was given no weight at all, partly because it was procedurally non-compliant, lacking a complete statement of truth in proper form and containing no certificate of compliance, and more significantly because she did not attend to be cross-examined.
So the AI point sits on a spectrum the court was already applying. Non-attendance and missing certification cost that witness all of her weight. Unexplained AI assistance cost the defendant and his witness the court's confidence that the words were theirs.
Why this is the harder category to manage
A fabricated citation can be found and removed. A fabricated exhibit can be disproved. Neither is easy, but both are about the truth of a specific item.
This is about provenance, and it does not resolve by checking the content. The statements may be entirely accurate. The court's difficulty was that it could not satisfy itself whose words they were, and the evidence that would have resolved it, the first drafts, was not before it.
That points to two practical steps. If an AI tool touches a witness statement at all, keep and be ready to produce the pre-AI draft, because a comparison is the only thing that can substantiate a claim that nothing was added, removed or re-arranged. And describe the tool plainly. The judge's suspicion attached to the euphemism, not to the software.
There is also a proportionality point for advisers. The judge's view was that the spelling and grammar tools already in ordinary word processors were sufficient for what these witnesses said they needed. Where that is true, using a language model on a statement introduces a provenance question in exchange for very little.
The exposure here is not sanction, it is weight, and weight decides cases. Three things to act on. First, a witness statement in the Business and Property Courts already carries a certificate of compliance under PD57AC, and reduced weight is an express consequence of non-compliance, so no AI-specific rule is needed for AI use to cost a party. Second, if a model touches a statement, preserve the pre-AI draft: the witnesses here asserted that ChatGPT changed nothing, and the court could not test that because the first drafts were never produced. Third, name the tool. The judge could not rule out that the phrase digital assistant was itself contributed by ChatGPT, or that it was chosen to obscure the tool's use, and that doubt fed directly into how the evidence was received.
Source File
https://caselaw.nationalarchives.gov.uk/ewhc/ch/2026/923
Open the judgment and confirm four things: the neutral citation [2026] EWHC 923 (Ch) with the date 22/04/2026 and HH Judge Klein sitting as a High Court judge; paragraphs 42 and 43, setting out the PD57AC requirements and the risk of exclusion or reduced weight; paragraphs 44 to 47, recording the digital assistant explanation, the judge's comment on that phrase and the absence of the first drafts; and paragraph 48, in which the judge states he cannot be sure the statements are in the witnesses' own words.
I cannot be sure that the Defendant's, or Ms Godwin's, witness statements are in their own words, reinforcing my decision to approach the Defendant's evidence cautiously and requiring me to approach Ms Godwin's evidence cautiously too. Godwin v Godwin, [2026] EWHC 923 (Ch) at [48]
FAQ
Was anyone sanctioned?
No. Nothing was struck out and no sanction was imposed. The consequence was evidential: the judge approached both witnesses' evidence cautiously because he could not be sure the statements were in their own words.
Was anything in the statements found to be false?
No. There is no finding that anything was invented or untrue. That is what distinguishes this from the fabricated-citation and fabricated-exhibit cases; the issue was provenance, not accuracy.
What exactly did the witnesses say they did?
That they had limited assistance from a digital assistant, meaning ChatGPT, for grammar, spelling and presentation; that each prepared their own first draft; that each uploaded that draft; and that ChatGPT's product did not add, remove or re-arrange any words compared with the first draft.
Why did the judge not simply accept that?
Because he had not been provided with the first drafts, so the assertion could not be tested. He also found there was no good reason to use ChatGPT at all given that ordinary word-processor spelling and grammar tools were available, and he could not rule out that the unusual phrase digital assistant was contributed by ChatGPT or chosen to obscure the tool's use.
Which rules were engaged?
CPR Part 32 and Practice Direction 57AC. The District Judge's order required a complete statement of truth in proper form and a signed certificate of compliance under paragraph 4.1 of PD57AC, and expressly warned that non-compliant statements risked exclusion or reduced weight.
What should a party do if AI has touched a witness statement?
On this judgment, keep the pre-AI draft so any claim that the tool changed nothing can be substantiated, and describe the tool by name rather than by a euphemism. The court's doubt here attached to the unexplained description and the missing drafts.
Related briefings
Sponsored Training
Practical AI training for regulated professionals, built around verification, documentation and a defensible process. See the courses.