Use of AI in court documents: the rules in England and Wales

The Civil Justice Council published its final report on using AI to prepare court documents on 29 October 2026. It makes fifteen recommendations, and for solicitors and barristers the main one is a refusal: no new AI declaration on pleadings, skeletons or advocacy.

The declarations land instead on witnesses of fact and on experts. None of it is in force until the rules are changed. Meanwhile the TCC Guide, three judgments and the SRA already set expectations, and our page for solicitors covers the client-file side.

What the Civil Justice Council’s final report on AI recommends

The working group was chaired by Sir Colin Birss. Its approach is functional and risk-based: the question isn’t which tool you used, but whether AI generated or influenced the substantive content, meaning the facts, evidence, opinion, reasoning or argument.

Regulation by the SRA, BSB and LSB stays the starting point. The Civil Procedure Rule Committee is asked for a plain-language Practice Direction codifying what the courts have said on AI use and sanctions. Knowing which details identify someone is part of the regulators’ confidentiality guidance.

Of the fifteen, Recommendation Six is the one I’d flag to anyone who isn’t a lawyer: it changes what happens when a client sits down to write their own statement.

Do you have to disclose AI use? It depends on your role

The interim report of 17 February 2026 had floated a declaration from legal representatives on trial witness statements. The final report moves that duty onto the witness, whether or not the party has a lawyer, and likens AI that reshapes a witness’s evidence to witness coaching.

For lawyers, it went the other way. Consultation responses strongly backed professional responsibility for statements of case, and warned that a declaration would breed satellite litigation over whether AI was used. Until the Practice Direction lands, my preference is to act as if the witness and expert declarations already existed.

From interim report to Practice Direction: what’s in force now

There’s a timeline problem with this subject. Articles written between February and October 2026 quote the interim report’s proposals as if they were the plan, and several were dropped. Quoting the interim report as current law is the mistake I’d most expect this winter.

What binds you today is narrower than the commentary suggests: court guides, the Upper Tribunal’s amended form, the existing CPR and your regulator.

The SRA’s warning notice of 17 August 2026 is the piece most firms will be measured against, and our checklist for choosing an AI tool works through its conditions on contracts and training.

The TCC Guide on AI: if in doubt, disclose

Section 1.6 of the TCC Guide, published on 1 July 2026, runs to five short paragraphs. Paragraph 1.6.2 says “any doubt as to whether the fact that AI has been used should be mentioned must be resolved in favour of disclosing such use”.

Paragraph 1.6.3 gets less attention. Legal representatives must uphold the confidentiality, and protect the privacy, of “any underlying data” used with AI. Much of it sits in scanned PDFs, where a black box drawn over a name can leave the text underneath for any tool to read.

This one irritates me slightly: two judicial texts eight months apart start from opposite defaults. Embedded AI in practice software makes the TCC default harder to apply, and the leaks that happen through those built-in model calls rarely reach the fee earner’s screen.

AI fake citations in UK courts: Ayinde and Munir

The case every litigator now knows is Ayinde v Haringey, heard with Al-Haroun by a Divisional Court led by Dame Victoria Sharp. At paragraph 6 it said freely available generative AI tools “are not capable of conducting reliable legal research”. They may cite sources that don’t exist.

The Upper Tribunal went further in UK and Munir, promulgated on 17 November 2025. Its judicial review claim form now has a statement of truth that every authority cited exists, can be found from its citation and supports the point. Sign it over a false case and expect, ordinarily, a regulator referral.

Munir is principally about supervision: a supervisor who misses a junior’s false citation is likely to be more culpable. The Law Society lists court sanctions from public censure to contempt or referral to the police. Put the checking rule in your firm’s AI policy, not in each fee earner’s memory.

AI in witness statements: Godwin v Godwin

Godwin v Godwin, [2026] EWHC 923 (Ch), was decided in Leeds on 22 April 2026 by HHJ Klein: two brothers, both in person, disputing how their father’s body should be disposed of. The defendant and his former wife said a “digital assistant” had helped with grammar, spelling and presentation.

It was ChatGPT, and they said it changed no words. The judge saw no good reason to use it and couldn’t rule out that ChatGPT supplied the phrase “digital assistant” itself. Without the first drafts, he couldn’t be sure the statements were in the witnesses’ own words.

Godwin is the case I’d hand to any client drafting their own statement, with one instruction: keep the first draft, bearing in mind that a Word draft carries its tracked changes to whoever receives it. Recommendation Six would turn the lesson into a declaration.

AI use in expert reports under Practice Direction 35

An expert’s report must already carry the statement of truth in paragraph 3.3 of PD 35: the expert has made clear which facts are within their own knowledge, and the opinions are true and complete. It says nothing about AI, a gap the CJC thinks matters more for experts than lawyers.

AI can properly be part of an expert’s method, and the other side has to see the method to test it. The interim report cited the Bond Solon Expert Witness Survey 2025: 20% of the 525 experts who took part had used AI in that role.

My bet is that experts meet this first through Recommendation Ten, in the CJC’s revised 2014 Guidance for the Instruction of Experts, well before any CPR amendment. Medical records they work from are special category data under UK GDPR, whichever tool reads them.

Litigants in person using AI in court

The CJC’s terms of reference excluded litigants in person, but consultation responses pushed hard. Recommendation Two asks the CJC and the rule committee to consider their use of AI “as a matter of urgency”, with no date attached.

Until then, the judges’ guidance of October 2025 is the nearest thing. It accepts that chatbots may be the only advice some litigants receive, and tells judges to ask what checks were made and to remind the litigant they’re responsible for what they put before the court.

I find it odd that the people most likely to lean on a chatbot are the ones the CJC has left for later. What OpenAI keeps from those conversations is set out in our ChatGPT guide.

Can what you say to AI be used in court? Privilege and disclosure

The sharper risk is what leaves the firm. In Munir, the Upper Tribunal said uploading confidential documents into an open-source tool such as ChatGPT puts them in the public domain. It named closed tools such as Copilot as an alternative, so check what Copilot does with confidential work.

The SRA’s notice speaks of “the permanent waiver of privilege that can occur”. Paid tools aren’t exempt: client data must stay secure, not train models without explicit authority, and not be kept longer than necessary. Firms building their own face a choice between a private LLM and a privacy proxy.

Treat every prompt as something you might one day have to show the other side. A saved chat is information recorded, which is how CPR 31.4 defines a document. If your systems call a model, what went in may also sit in the logs of an AI gateway.

If the model calls come from your own software, redacting client data at your own gateway keeps names out of those logs and out of the provider’s hands.

Preparing court documents with AI, without the client in the prompt

Recommendation Six asks whether AI shaped a witness’s words. It says nothing about the bundle someone wants summarised, the chronology built from emails or the first draft of a letter to the other side, and those are what carry the client’s identifiers.

Under UK GDPR, swapping names for labels is pseudonymisation, not anonymisation: the facts can still point to someone. The step nobody signs for is the one I’d make routine. Paste a passage into the box below to see which fixed-format identifiers it carries.

Before a chronology or a bundle extract goes into an AI tool, Nonimo replaces the names, phone numbers and postcodes in the passage you select with labels on the fee earner’s own computer, and restores them in the reply. See how it works for UK litigation teams.

Sources

• Civil Justice Council, Use of AI for Preparing Court Documents: Final Report, 29 October 2026: the fifteen recommendations, the witness and expert declarations, litigants in person, translations.
• Technology and Construction Court Guide 2026, edition 4.1, published 1 July 2026 and updated September 2026: section 1.6, paragraphs 1.6.1 to 1.6.5.
• Godwin v Godwin, [2026] EWHC 923 (Ch), 22 April 2026: paragraphs 44 to 48.
• R (Ayinde) v London Borough of Haringey, [2025] EWHC 1383 (Admin), 6 June 2025: paragraph 6.
• UK and R (Munir) v Secretary of State for the Home Department, [2026] UKUT 81 (IAC): headnotes 2 to 4, paragraphs 21 and 37.
• SRA, Misuse of AI: warning notice, 17 August 2026: safeguards for client data and the permanent waiver of privilege.
• The Law Society, Generative AI: the essentials, updated June 2026: verifying citations and the range of court sanctions.
• Civil Procedure Rules, Practice Direction 35: paragraph 3.3, the expert’s statement of truth.
• Civil Procedure Rules, Part 31: rule 31.4, the meaning of document.
• DCMS, Information Commission succeeds the ICO (GOV.UK): the transfer of functions on 30 September 2026.

Accessed 8 November 2026.