Do Androids Dream of Electric Skeletons? AI Submissions After Hancox v Sutherland

One of the seven Millennium Prize problems, open since 2000, asks whether the equations governing fluid flow always have well-behaved solutions or can break down. On 8 September 2026 OpenAI published a 165-page proof that they can break down, produced by thousands of AI agents running for four days and checked step by step by a proof-verification program. Ten days on, mathematicians are still reading it and the Clay Institute, which set the prize, has not pronounced on the solution (at the time of writing). Cédric Villani, who in June had said that large language models understand nothing of what they say, is quoted in Le Monde calling it a cataclysm unlike any mathematics has known. The checker confirms that each step follows from the last. It cannot tell them that the statement proved is the problem as posed, and it cannot make 165 pages of machine reasoning intelligible to the people who have to say whether it is right.

At almost the same time, in a different sphere, the EAT handed down Hancox v Sutherland [2026] EAT 139.

A litigant in person had filed a skeleton argument of 300 pages and almost 132,000 words, generated with ChatGPT, and invited the respondents to identify any inaccuracies in writing by 4pm the next day. Marcus Pilgerstorfer KC, sitting as a deputy judge of the High Court, did not treat the tool as the problem, and the timing shows why he was right not to. The models of two years ago famously could not count the r’s in strawberry; the one that produced the proof is the same idea with far more training. It is no stretch to say the ability has evolved dramatically, as the silk in The Spectator found when a model wrote his appeal in thirty seconds. What has not changed is that ability in the machine is not understanding in the person who files the document. The judgment opens by recognising what generative AI offers a party without access to legal advice, and says at [31] that there is no principled reason why a litigant in person should not “take reasonable steps to use it responsibly”. What it will not accept is the document arriving unchecked, or the checking being passed to the other side. The minimum at [32] has three parts: the document complies with the Practice Direction of the Employment Appeal Tribunal 2024 (the EAT PD 2024), which is where an appeal’s procedural rules live rather than in the Civil Procedure Rules; it has been checked for accuracy as thoroughly as the litigant is reasonably able; it confines itself to the best points. At [33], submitting the product of generative AI without those checks, or placing the onus on an opponent, is not acceptable: enquiries and sanctions are likely to follow.

The judgment also sets out, in one place, what the EAT will do when a document arrives in that state.

Mathematics lets an unverified proof go out for the community to test, because the community wants the answer and shares the cost of checking it. A skeleton argument has one other reader, the opponent, and the cost of finding its errors falls on the party they were aimed at. That is the allocation rule at [33]. A proof that is right but unreadable still proves its theorem, which is why the four-colour theorem was accepted in 1976 over the objection that no one could follow it. A submission that is right but unreadable proves nothing, because what a tribunal adopts is not the answer but the reasons it can give for it. The 300 pages were not relied on at the hearing [2]; with ELAAS counsel beside him the appellant argued one ground orally: that ground, not the document, goes forward.

In practice

• In the EAT there is nothing to apply under; the power is the tribunal’s own. Write to the EAT when the skeleton arrives, inviting the judge to exercise the power in paragraph 11.6.6(a) of the EAT PD 2024 to reject it and require a compliant one, and ask at the first case management stage for an order that applications be made only on the form at Annex 2 to the EAT PD 2024. An invitation to find the errors by a deadline gets a one-line reply: Hancox at [33] places the onus on the author.
• In the employment tribunal there is no skeleton rule at all. The equivalents are a case management order under rule 30 of the Employment Tribunal Procedure Rules 2024 setting a page limit and requiring a document that meets the Hancox minimum at [32], and, where a party persists, the unreasonable-conduct routes in rule 38 and rule 74. Ask for the order at the preliminary hearing, before the bundle is built around the document.
• For any party using the tools, file only what you could explain to the judge without the document in front of you. A submission you cannot answer questions on is worth nothing, whatever it contains; the 300 pages were not relied on at the hearing.
• For counsel’s own use of the tools, the standard is stricter: R (Ayinde) v London Borough of Haringey [2025] EWHC 1383 (Admin) and the BSB’s guidance under rC86. The Hancox minimum is the floor for every litigant, not the ceiling for counsel.

Table of Authorities

Alex acts in disputes like this one for claimants and respondents, instructed by solicitors or by direct access. How to instruct him or email the clerks.