I have read skeleton arguments that ran too long, and until this week I had never read of one that ran to 300 pages. That is the document a litigant in person filed at the Employment Appeal Tribunal on 27 July, a single compressed file holding almost 132,000 words, and he told the judge he had produced it with ChatGPT. The reserved judgment came out yesterday. It belongs on your desk because the person who filed it is the kind of opponent your firm now meets in the tribunal list.
Marcus Pilgerstorfer KC, sitting as a deputy High Court judge, took issue with more than the length. The appellant had invited the respondents to identify the inaccuracies in the document for him. The judge held that it was for the appellant to take responsibility for the contents of his own written argument, and put the principle in a single line. 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 cost of an opponent's AI output is a cost the tribunal expects the opponent to carry.
The floor the judgment sets
What the judge wrote applies to litigants in person and to you in the same terms. Anything produced with AI must comply with the procedural rules that govern it. It must have been checked as thoroughly as the litigant or representative is reasonably able for accuracy, so that the factual, evidential and legal points are correct, including the references to authorities, and so that the tribunal is not misled. It must carry only relevant points, with the focus on the central or best arguments and no undue repetition.
The rules had already given him the measure. A skeleton argument in the EAT should generally run to between 5 and 15 pages, depending on the number of grounds and their complexity. A judge faced with more than 20 pages can require a shorter version or cut the time allowed for oral submissions. The appellant filed twenty times the upper end of that guide. A document of the same kind is likely to be rejected in future, the judge said, because compliance with the tribunal's procedural requirements is mandatory, not optional.
What to do when one arrives
Do not answer it page by page. A firm that meets an opponent's machine output line for line converts somebody else's shortcut into its own irrecoverable cost, and the file runs away from the client long before the hearing. Write to the tribunal instead. Give the page count and the word count, point at the practice direction and at this judgment, and ask for a direction that a compliant skeleton be filed before you are required to respond. The cost of an opponent's AI output is a cost the tribunal expects the opponent to carry.
Keep the record while you do it, because a costs application in the tribunal turns on unreasonable conduct and the time your fee earners burn reading 132,000 words is the evidence of it. Then read this against the other side of the same coin, the litigant in person who used AI to beat counsel at Oxford County Court. The technology that sharpened one opponent buried another, and your firm will meet both this year. The habit that answers each of them is the same, which is to check what a machine produced before anyone else has to.
The judgment is published in full as Hancox v Sutherland and others [2026] EAT 139 on the National Archives case law service, and it is open to read without a subscription.
If you want the questions your team should ask of an AI document, the other side's as much as your own, set down on one page, start with a conversation.
