The Employment Appeal Tribunal in Hancox v Sutherland has stated that anyone using generative AI to prepare EAT documents should ensure, at a minimum, that those documents comply with the procedural rules, have been checked for accuracy as thoroughly as the author is reasonably able, and contain only relevant points

What paragraph 32 says, and where it comes from

The judge poses the question in paragraph 1: generative AI is being used with increasing frequency by litigants in the EAT to prepare notices of appeal, skeleton arguments and correspondence, so what are the responsibilities that fall on those who use such tools? Paragraph 32 is the answer. Litigants who use AI should ensure, at a minimum, that all documents submitted comply with applicable procedural rules, with PD 3.8 for grounds of appeal and PD 11.6 for skeleton arguments given as the EAT examples; have been checked as thoroughly as the litigant or representative is reasonably able for accuracy, so that factual, evidential and legal points, including references to authorities, are correct and the tribunal is not misled; and contain only relevant points, focused on the central or best arguments, presented in an easily comprehensible manner and avoiding undue repetition.

The paragraph begins "In accordance with the guidance and authority I have cited above". That is a deliberate framing. Paragraphs 24 to 30 walk through R (Ayinde) v London Borough of Haringey [2025] 1 WLR 5147, Taiwo v Homelets of Bath [2025] EWHC 3173, the judicial guidance on AI of 31 October 2025, Re D (A Child) (Recusal) [2025] EWCA Civ 1570 and the employment tribunals' Presidential Guidance on interim relief of 22 June 2026, which the judge quotes for the proposition that litigants who use AI have a responsibility to ensure that what is submitted is concise, relevant and accurate.

Paragraph 31 draws the line the whole judgment turns on. A litigant in person will not usually be subject to the professional duties with which Ayinde was primarily concerned, but the underlying principles of personal responsibility and accuracy apply to all persons who submit documents to a court or tribunal. Paragraph 37 then states the message the judge says emerges from all of it: the use of AI is not generally prohibited for tasks such as preparing skeleton arguments, but all litigants must take personal responsibility for any document submitted and it must be carefully checked, as thoroughly as the author is reasonably able. Those two uses of must are the judge's words.

The judgment does not describe itself as the first EAT statement on the point, and we do not either. It cites the Presidential Guidance from June 2026 as earlier employment-specific guidance on exactly this problem.

The 300-page skeleton, and what the judge did with it

The facts are in paragraphs 82 to 87. On Monday 27 July 2026, two days before the preliminary hearing, the Appellant filed a compressed PDF containing his skeleton argument. It ran to 300 pages and a word count revealed almost 132,000 words. The judge's verdict on the label: the word skeleton was entirely inapposite.

Paragraph 84 lists the defects against PD 11.6.6. Not concise, and far beyond the recommended 5 to 15 pages and the 20-page upper limit. Numerous sections each with its own paragraph numbering. No uniform 12-point font. References to documents outside the core bundle when no supplementary bundle had been prepared. The judge's own summary is that the document did not assist the EAT in preparing for the hearing, and that the Appellant himself had recognised in his covering email that there was insufficient time to consider it in full.

At the hearing the Appellant explained that he had used ChatGPT to create the document because he had to prepare it quickly. Paragraph 85 records that he told the judge much of the document lacked credibility and that ultimately no reliance was placed on it. The judge was satisfied that the Appellant had not checked the document against the PD or for accuracy, and said so on the basis of the covering email and the Appellant's own assessment at the hearing.

Two things the judgment does not record deserve to be stated plainly. There is no order formally rejecting the skeleton; the document fell away because it was not relied on. And there is no costs order, regulatory referral, contempt step or other sanction imposed for the use of ChatGPT. What paragraph 87 says is forward-looking: compliance with the EAT's procedural requirements is mandatory, not optional, and in future such a document is likely to be rejected by the EAT with consideration given to the exercise of the other powers the judge had referred to. Likely, in future, with consideration. The judgment preserves the tribunal's discretion and we preserve it here.

Onus on the opponent: the passage respondents' counsel will keep

The covering email quoted at paragraph 83 is the practical heart of the case. The Appellant told the EAT he did not expect the judge to read every page, then invited the Respondents to identify any specific factual inaccuracy, incorrect quotation, mistaken date, inaccurate description of a document, erroneous bundle reference or misrepresentation of their position, in writing, by 4.00 pm the next day. General objections that the document was too long, unread, or should be excluded would, he wrote, not identify an actual error.

Paragraph 86 rejects that approach in terms. It was for the Appellant to take responsibility for the contents of his own written argument, and it is not acceptable for a litigant to attempt to place that burden on their opponent, far less to do so over an unreasonable timescale. Paragraph 33 states the general rule the same way: simply submitting the product of generative AI, or placing the onus to check a document onto an opponent, is not acceptable, and where concerns arise that checks have not been undertaken or were inadequate, judicial enquiries and potential sanctions are likely to ensue.

For anyone acting for a respondent, that is the paragraph to cite when an AI-assisted document arrives with an invitation to find the errors. The judgment puts the checking cost where the drafting choice was made.

Paragraph 34 adds a related trap specific to appeals. The EAT will not generally consider an argument not advanced at first instance, so any AI-produced document needs to be read for new points, which should either be removed or clearly identified with an application under PD 8.13.

Powers the EAT says it has, and what it ordered here

Paragraph 35 borrows the menu from Ayinde: a reference to a regulator or a wasted costs order for regulated professionals; and, regardless of status, strike out, costs sanctions, public admonishment and, in the most serious and egregious cases expected to be rare, a police referral or contempt proceedings. The judge states that the EAT has similar powers and sanctions available to it. Paragraph 36 adds that AI use may cause a tribunal to be circumspect about the reliance placed on a document or evidence, citing Godwin v Godwin [2026] EWHC 923 (Ch) on ChatGPT-prepared witness statements, R v FGD [2026] EWCA Crim 918 on witness coaching by AI, and Lodhia v Twelve Trees [2026] EWHC 1889 (KB) on the argument that false information was purely the invention of an AI tool. These are powers described as available; none was exercised in this judgment.

What was ordered is at paragraph 170 and in the summary. One ground of appeal, on whether the Appellant had fair notice and a fair opportunity to prepare for the strike out application, was found reasonably arguable and proceeds to a full hearing. The other grounds in the notice of appeal were dismissed, and so was the Appellant's application to resile from the position taken at the hearing and pursue a further ground based on articles 6 and 10 ECHR. His disclosure application was dismissed and certified totally without merit, and his applications to review earlier EAT orders were dismissed. None of those outcomes is expressed as a consequence of the use of ChatGPT.

The case management directions at paragraphs 155 to 169 are directions in this appeal, not general rules. The Respondents are to prepare the core and supplemental bundles. Any extra bundle is capped at 50 pages. Paragraph 160 requires both sides' skeleton arguments for the full hearing to comply fully with the PD, importantly as to length, on pain of the sanctions described. Paragraph 163 orders that every application be made on the PD's Annex 2 form or it will not be considered or responded to. Paragraph 169 stops short of restricting the Appellant's email correspondence, citing the available evidence about his mental health and recent apologies, while warning that continued excess will lead the EAT to consider a post-only order.

Who is addressed, and who is not

Paragraph 7 sets the baseline: the EAT Rules and the PD apply to all litigants whether represented or in person, and Barton v Wright Hassall LLP [2018] 1 WLR 1119 is cited for the proposition that lack of representation will not usually justify a lower standard of compliance. Paragraph 37 says the AI message applies beyond legal representatives to those representing themselves. Paragraph 24 notes that since Ayinde regulated professionals are increasingly being held to those standards, with a footnote listing six decisions from 2025 and 2026.

The people directly addressed are parties and representatives who submit documents to the EAT. That includes counsel and solicitors on both sides, ELAAS representatives, and litigants in person. HR teams and in-house employment lawyers are affected in the ordinary way, because they instruct on respondent answers and skeletons, and because paragraph 86 tells them they do not have to audit the other side's AI output on a one-day timetable.

What the judgment does not do is impose any AI duty on employers or HR functions outside tribunal proceedings, require disclosure of AI use, or require any litigant to buy professional verification. It gives no direction to employment tribunals at first instance, but its general observations are expressed more widely: paragraph 22 says the ET and EAT are no exception to the rise in AI use by litigants, and paragraphs 31 and 33 speak of anyone who submits documents to a court or tribunal. The only ET-specific material it quotes is the Presidential Guidance on interim relief. We confine the article to what the EAT said and ordered.

What we did not verify

What we opened: the full approved judgment in Hancox v Sutherland and others [2026] EAT 139 as published on The National Archives Find Case Law service, read from the header and summary through the introduction, Part 1 (paragraphs 6 to 37), the skeleton argument passage (paragraphs 82 to 87), the preliminary hearing narrative, the disclosure and review dispositions (paragraphs 145 to 154), the case management directions (paragraphs 155 to 169) and the conclusion at paragraph 170. The header reads "Date: 17 September 2026" and "Preliminary Hearing date: 29 July 2026".

What we did not open: the EAT Rules 1993, the Practice Direction 2024, the Employment Tribunals Act 1996, Ayinde, Taiwo, Re D, Barton, Godwin, FGD, Lodhia, the Presidential Guidance on interim relief, the judicial AI guidance of 31 October 2025, the Civil Justice Council material, the Employment Tribunal judgment under appeal, the January 2026 order of HHJ Tayler, the 300-page skeleton itself, or the covering email beyond the passage quoted in the judgment. We describe each of those only as this judgment describes them.

What we refuse to claim: we do not say this is the first EAT statement on AI-prepared documents, because the judgment does not say so and cites earlier employment guidance. We do not say the skeleton was rejected, because paragraphs 2 and 85 say no reliance was ultimately placed on it and paragraph 87 speaks of likely rejection in future. We do not say non-compliant AI documents will be rejected automatically, because the judgment reserves that to appropriate cases and lists alternative responses at paragraph 21. We do not say the judgment creates a new rule, practice direction or deadline. We do not say every observation in Part 1 binds every tribunal, employer or HR professional. We do not say the skeleton contained fabricated authorities, because the judgment records no such finding. We do not say the Appellant was referred, sanctioned in costs or found in contempt for using ChatGPT, because he was not in this judgment. We give no date for the full hearing, because none is fixed. Where the word must appears above in the desk's voice, it is quoting or closely reporting the judge or the Practice Direction.

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.

Related briefings

• Ayinde: the Divisional Court on fabricated AI citations and contempt
• Godwin v Godwin: ChatGPT-prepared witness statements and reduced weight
• AI court rules tracker
• Europe AI regulation tracker
• AI Regulation News hub