In Hancox v Sutherland & Others [2026] EAT 139, the Employment Appeal Tribunal (EAT) considered the increasing use of generative AI by litigants to prepare notices of appeal, skeleton arguments and correspondence. The case provides useful guidance for lawyers and litigants alike on the level of human oversight expected when AI is used in litigation.
What happened?
The appellant, who was representing himself, submitted a 300-page skeleton argument containing almost 132,000 words which had been created using ChatGPT. This was dramatically longer than the EAT’s usual expectation that a skeleton argument should generally be between five and 15 pages. The document also failed to comply with other requirements of the EAT Practice Direction. At the hearing, the appellant acknowledged that much of the document lacked credibility and ultimately did not rely on it.
Perhaps more strikingly, when submitting the document, the appellant invited the respondents to identify any factual inaccuracies, incorrect quotations, dates, bundle references or other errors. The EAT rejected this approach. It was for the appellant to take responsibility for checking his own written argument and not for his opponent to quality-control AI-generated material on his behalf.
What did the EAT say about using AI?
The EAT did not suggest that generative AI should be prohibited in litigation. It recognised the technology’s potential to assist litigants, particularly those who do not have access to professional legal advice. Used properly and responsibly, AI can help with the preparation of legal documents. However, the EAT identified a range of risks. AI-generated material may be inaccurate, excessively long, irrelevant or procedurally non-compliant. It may misstate the factual or legal position, introduce arguments that were not raised before the lower tribunal or, in the familiar phenomenon of AI “hallucinations”, invent cases, citations or quotations.
The EAT considered the High Court’s decision in R (Ayinde) v London Borough of Haringey [2025] 1 WLR 5147, and emphasised that users cannot simply assume that plausible-sounding AI output is correct. The EAT distilled the position into three minimum requirements. Anyone using AI to prepare a document for a court or tribunal must ensure that it:
- Complies with procedural requirements. The document must comply with applicable rules, practice directions and court orders. AI use is not an excuse for non-compliance.
- Has been properly checked. The user must check the document as thoroughly as they reasonably can for factual, evidential and legal accuracy, including checking authorities and citations and ensuring the court is not misled.
- Is relevant, focused and comprehensible. AI output should be edited so that it concentrates on the central or strongest arguments and avoids irrelevant material and undue repetition.
What are the consequences of getting it wrong?
The judgment makes clear that inadequate oversight of AI-generated material can have real procedural consequences. The EAT warned that a document such as the 300-page skeleton submitted in this case would, in future, be likely to be rejected. Where concerns arise about whether appropriate checks have been carried out, judicial enquiries and potentially sanctions may follow.
More serious misuse may attract other consequences depending on the circumstances. Drawing on existing case law, the EAT referred to potential costs sanctions, striking out, professional regulatory consequences and, in the most serious cases, contempt proceedings or referral to the police.
