In partnership with
Last week this newsletter looked at a litigant in person who used AI well. This week the Employment Appeal Tribunal has dealt with one who did not, and in doing so it has set out a minimum standard for anyone, represented or not, who puts AI-assisted documents before it. Elsewhere, OpenAI has launched a legal product which UK firms cannot yet buy, Harvey has started a pro bono programme with a UK arm, and the Legal Services AI Growth Lab closes to applications on Sunday.
AI in Practice
The EAT says AI output is the author's job to check, not the opponent's
On 17 September the Employment Appeal Tribunal handed down judgment in Hancox v Sutherland & Others [2026] EAT 139. Headline is that a litigant in person filed a skeleton argument of 300 pages and almost 132,000 words, created with ChatGPT, two days before a preliminary hearing. The relevant Practice Direction asks for 5–15 pages, with anything over 20 liable to be sent back. In his covering email the appellant said he did not expect the judge to read every page, and invited the respondents to identify in writing, by 4pm the following day, any specific factual inaccuracy, incorrect quotation or erroneous bundle reference, adding that a general complaint that the document was too long would not identify an actual error. At the hearing he told the judge, Marcus Pilgerstorfer KC sitting as a deputy High Court judge, that much of the document lacked credibility, and no reliance was placed on it.
The judge called the document "entirely unacceptable" and used the case to set out what anyone using generative AI to prepare documents for the EAT must do as a minimum. Every document must comply with the procedural rules, must be checked as thoroughly as the litigant or representative is reasonably able so that the factual, evidential and legal position is not misstated, and must contain only relevant points, focused on the central or best arguments and without undue repetition. Paragraph 33 is the sentence to remember: "Simply submitting the product of generative AI to a court or tribunal, or placing the onus to check a document onto an opponent, is not acceptable." The judgment's summary makes clear that this applies to "all litigants and their representatives", and the judge warned that a document of this kind is likely to be rejected in future.
This author sees three practical consequences.
The first answers a question this newsletter raised last week; when the other side serves a vast AI-generated document and asks you to find its errors, the EAT has now said in terms that the burden is not yours, and the sensible response is to raise non-compliance with the tribunal early and in writing rather than spend the night before a hearing proofreading someone else's chatbot.
The second is that the standard clearly reaches well beyond simple fabricated citations. Length, focus and compliance with the practice direction are now part of what the tribunal expects anyone using AI to check, and a fee earner who lets a tool produce a 40-page skeleton where 12 would do is open to the same criticism.
The third is the one this author intends to raise with his litigation team. At paragraph 36 the judge said that the use of AI "may also cause a court or tribunal to be circumspect about the reliance that can properly be placed on a document or evidence", citing Godwin v Godwin [2026] EWHC 923 (Ch), where the use of ChatGPT in preparing witness statements led to a cautious approach to the evidence, and the concerns raised with the Civil Justice Council about AI reshaping or embellishing evidence in ways which are not immediately apparent.
The judgment is fair to the technology. It records the Lady Chief Justice's evidence to a Lords committee in June that many judges find AI-assisted submissions from litigants in person more helpful and easier to digest, and it does not prohibit AI for skeleton arguments.
The decision is a preliminary hearing decision in the EAT and does not bind the civil courts, but it gathers the recent authorities in one place (including Cork & Another v Smith [2026] EWHC 1199 (Ch), where supervising solicitors were held accountable for failing to check a junior's work containing a hallucinated statutory provision), and it seems likely to be cited well beyond employment appeals.
Takeaways
Act: Add a short pre-filing check for any court or tribunal document where AI assisted, covering the practice direction's length and format requirements, verification of every authority against the source, and whether the points have been cut back to the best ones. Record who did the check.
Watch: Whether the civil courts adopt the EAT's formulation, and what the Civil Justice Council's draft report on AI in court documents says about witness statements, which the EAT has now flagged as the area where reliance may be affected.
Risk: A witness statement drafted with AI assistance which the court treats with caution at trial, in a case where that witness's evidence is the part of the case which matters.
Read: Hancox v Sutherland & Others [2026] EAT 139 (GOV.UK), and the reporting at Law Society Gazette
On your radar
OpenAI has launched a legal product, but its research index covers US law only: OpenAI announced Astra for Law on 17 September, combining its GPT-6 Astra model with a legal search index covering US case law, statutes, regulations, court rules and administrative decisions, together with 26 partner plugins from providers including Thomson Reuters, Harvey, Legora and iManage. It is initially offered to selected US law firms, with Latham & Watkins, Ropes & Gray, Cooley and Sullivan & Cromwell named as early adopters, and API access is "coming soon". OpenAI says it was 40% more accurate on research answers than the same model using web search alone, though that is its own testing and there is no comparison with established research providers. Zero data retention is available on the API for eligible firms. Why it matters for UK lawyers: nothing here is available to UK firms yet, and the headline accuracy gains were measured against US law, but Harvey and Legora are both named as able to build on it, so it may reach UK firms through tools they already license. If your firm uses either, ask whether any feature will run on Astra for Law and, if so, what it draws on for English authorities before anyone relies on its research output. (OpenAI, Legal IT Insider)
MPs and peers want an AI Bill and a narrower meaning of "human in the loop": The Joint Committee on Human Rights reported on 14 September that the UK's approach to AI regulation is patchy and called for a wide-ranging AI Bill, a single statutory oversight body with power to test high-risk systems and grant remedies, due diligence duties running down the supply chain to deployers, and a strengthening of the UK GDPR provisions on automated decision-making so that a human in the loop does not, by itself, count as meaningful human involvement. Joshua Rozenberg, writing in the Gazette on 21 September, notes that two years after Labour's manifesto commitment there is still no binding legislation. A select committee report binds nobody, and the government's response is the document to watch. Why it matters for UK lawyers: the Data (Use and Access) Act 2025 made "meaningful human involvement" the hinge of much automated decision-making advice, and a parliamentary committee has now said the phrase is being read too generously. If you advise on automated decisions, check whether your standard wording on human review would survive the committee's test, and diarise mid-November for the government's reply. (Joint Committee on Human Rights, Law Society Gazette)
Harvey is building AI agents for pro bono casework, with a UK firm in the pilot: Harvey announced on 22 September a pro bono programme across the UK, Australia and the US, piloted with Ashurst Perkins Coie, DLA Piper and Nelson Mullins, which will build custom agents for high-volume casework such as housing support and social welfare claims. Harvey says it is covering the cost and will give selected frontline legal services organisations free access to the platform later this year. Why it matters for UK lawyers: pro bono clients are often the most vulnerable a firm acts for, and volume casework handled by agents is exactly where supervision is hardest to evidence. If your firm's pro bono team or a partner law centre is approached, settle before the first matter who reviews agent output, where client data is held, and whether clients are told that AI is being used on their case. (Artificial Lawyer, City AM)
The Legal Services AI Growth Lab closes to applications at 11.59pm on Sunday: The government's advisory sandbox for AI in legal services, run with the SRA, the Legal Services Board, the Council for Licensed Conveyancers and the ICO, will take around 10 to 12 participants for up to nine months, and is open to law firms and conveyancers as well as lawtech businesses. Participation gives coordinated access to regulators but no approval, endorsement or exemption. Why it matters for UK lawyers: this was mentioned here last week without the date, and the window is now days rather than weeks. If your firm has an AI project stalled on a confidentiality, privilege or data protection question which only a regulator can answer, decide by Friday whether to apply, because a rushed application over the weekend is unlikely to be the best version of it. (GOV.UK, SRA)
Ad Break
In order to help cover the running costs of this newsletter, please check out the advert below. In line with my promises from the start, adverts will always be declared.
The AI brief curated by Anthropic and ex-Google engineers
Thousands of AI papers, model releases, and product launches ship every week. You don't need all of them. You need the ten that matter.
TLDR AI is the free daily newsletter curated by Anthropic and ex-Google engineers. People who build AI for a living read everything, then send you what's actually worth your time: the models, research, and tools that will still matter next month.
No hot takes, no hype cycles, no 60-tweet threads. Clear summaries you can read between meetings.
Join 1.1M+ readers. Subscribe for free.
For Review
Guest post: The AI bill many law firms haven't priced in (James Harrison, Legal IT Insider)
Published on 22 September and arguing that per-seat licences are giving way to consumption pricing as vendors meter agents by the token, with current prices subsidised rather than reflecting true cost. The practical value is in the closing list: price protection with rate cards and repricing notice, tracking AI spend by client and matter like a disbursement, and keeping the model replaceable. Worth sending to whoever negotiates your next AI renewal before they sign it.
Read or listen: Legal IT Insider
Artificial Intelligence (AI) Guidance for Judicial Office Holders (Courts and Tribunals Judiciary)
Not this week's news: the current version dates from 31 October 2025. It is quoted in Hancox and is the natural companion to this week's lead, because it tells judges that if a chatbot appears to have been used they may ask what checks were done. Read the section on unrepresented litigants and assume the same questions can be put to a represented party.
Read or listen: Courts and Tribunals Judiciary
Ministers wake up to AI's existential threat (Joshua Rozenberg, Law Society Gazette)
A column from 21 September setting the Joint Committee's call for a ban on developing the most dangerous AI systems against the government's slower progress on binding regulation. It is short and useful context for any client conversation about where UK AI legislation is heading, which on this reading remains uncertain.
Read or listen: Law Society Gazette
Practice Prompt
Try the below prompt to check a court or tribunal document drafted with AI assistance against the minimum standard the EAT set out in Hancox v Sutherland, before it is filed. Ensure you fill in context and constraints and other aspects marked with {}. Remember to adhere to the Golden Rules and do not upload confidential or privileged information to public tools.
You are assisting a solicitor or barrister in England and Wales who is about to file a document with a court or tribunal, where generative AI was used at some stage in preparing it. Your task is to run a pre-filing check against the minimum standard set out by the Employment Appeal Tribunal in Hancox v Sutherland & Others [2026] EAT 139: that a document complies with the applicable procedural rules, has been checked for accuracy so that the factual, evidential and legal position is not misstated, and contains only relevant points focused on the central or best arguments. You are not approving the document for filing and you are not advising on the merits.
Context to apply:
- The court or tribunal, and the type of hearing: {e.g. EAT preliminary hearing, County Court fast track trial, High Court application}
- The type of document: {skeleton argument, grounds of appeal, witness statement, statement of case, application, correspondence to the court}
- The procedural rules and practice directions which govern it, including any length or format limits: {list them, or say "not yet identified" and I will treat that as the first finding}
- The document itself, redacted of names, addresses, figures and anything identifying: {paste}
- How AI was used and by whom: {e.g. first draft by a fee earner using a firm-licensed tool, summarising the bundle, suggesting authorities, restructuring an existing draft}
- What has already been checked, by whom, and against what sources: {describe, or say "nothing yet"}
- Any direction or order already made about this document: {e.g. page limit, filing deadline, format requirements}
Produce the check under these headings:
1. Procedural compliance
List every length, format, structure and content requirement which appears to apply to this type of document in this forum, and mark each as met, not met or unable to tell. Where you are unsure whether a requirement applies, say so and state where the lawyer should look to confirm it. Do not invent page limits or rules.
2. Authorities and sources
List every case, statute, rule, practice direction and guidance document cited. For each, state the proposition the document relies on it for, and mark it as to be checked against the source. Do not confirm that any authority exists or says what is claimed; the lawyer must verify each one against an authoritative source such as legislation.gov.uk, the National Archives Find Case Law service, the official Law Reports or a reputable legal database.
3. Facts and evidence
List every factual assertion and every reference to evidence or to a bundle page. Flag any statement which is not tied to a document or a witness, any page reference which cannot be checked from the material provided, and any passage which reads as though it has been generated rather than drawn from the evidence.
4. Focus and length
Identify the central or best arguments, the points which repeat them, and the points which add nothing. Propose what could be cut to bring the document within any limit, without rewriting the arguments themselves.
5. New points
If the document is for an appeal, flag any point which appears not to have been raised below, so that the lawyer can decide whether it needs to be removed or whether permission must be sought.
6. Witness evidence
If the document is or includes a witness statement, identify any passage whose wording may not be the witness's own, and list questions the lawyer should put to the witness to confirm that the statement reflects their evidence in their own words. Note that the court may be circumspect about reliance on evidence prepared with AI.
7. What this check cannot tell you
Flag every point where the output rests on an assumption about the forum, the rules, the facts or the authorities, and say what the lawyer would need to establish to replace it with a fact.
Constraints:
- {Add matter-specific constraints, for example a hearing listed within days, a supervising partner who must sign off, a client instruction on tone, or a court which has already commented on the length of a party's documents.}
- Apply the law and procedure of England and Wales. Where you are unsure which rule or practice direction applies, say so rather than asserting it.
- Do not invent cases, statutes, rules, practice directions, page limits or court guidance. If you cannot verify something, mark it as unverified.
- Do not rewrite the document or improve its arguments; identify what needs attention and leave the drafting to the lawyer.
- Write in plain English and keep each finding short enough to act on.
- This is a checking aid, not legal advice. The lawyer remains responsible for every document put before the court and for the checks carried out on it.How did we do?
Hit reply and tell me whether your team has started receiving AI-drafted documents from the other side, and how you have dealt with them.
Serhan
UK Legal AI Brief
Disclaimer
Guidance and news only. Not legal advice. Always use AI tools safely.
Recommended Newsletters
Below are a few newsletters that I recommend, for various reasons. Check them out!
