Terence Hancox v Kenneth Sutherland & Ors [2026] EAT 139 is an important legal-AI decision because it widens the conversation beyond fabricated authorities. The Employment Appeal Tribunal was confronted with a 300-page skeleton argument created using ChatGPT. The document did not comply with the EAT Practice Direction and, in the judge’s words, served to obscure rather than illuminate during preparation for the hearing.
The case is therefore useful for any litigation team thinking seriously about AI. The quality question is not simply “did the model hallucinate?” It is also whether the document is relevant, concise, procedurally compliant, based on the real evidential and legal position, and checked by the person who submits it.
1. The document was part of a procedural problem, not merely an AI anecdote
The EAT’s summary records that the appellant filed a 300-page skeleton argument created using ChatGPT. The Practice Direction generally expects skeletons to be concise and explains their purpose: to help the EAT and the parties focus on the points of law raised.
That context matters. A generated document can contain individually plausible paragraphs and still fail as litigation work because it does not perform the procedural function of the document.
2. Hancox identifies several distinct failure modes
The judgment discusses risks including output that is unfocused, excessive or irrelevant; non-compliance with procedural requirements; inadequate or inaccurate input material; misleading statements about the factual, evidential or legal position; new points raised at the wrong stage; bias; and hallucinated authorities.
That list is valuable because it resists a simplistic accuracy metric. Litigation documents are instruments within a procedure. They have a purpose, format, scope and audience.
3. Personal responsibility remains central
The EAT states that all litigants and representatives using generative AI must take personal responsibility for documents they submit. At a minimum, they must ensure compliance with applicable procedural rules, check accuracy as thoroughly as reasonably possible and focus the document on relevant points and the central or best arguments.
The court also makes clear that it is not acceptable to submit AI-created material without checking it or to shift the checking burden to the opposing party.
4. A source-correct document can still be professionally poor
Suppose every authority in an AI-generated skeleton exists and every quotation is accurate. The document can still be defective if it is 200 pages too long, misses the actual ground of appeal, cites documents outside the permitted bundle, repeats itself or forces the judge to reconstruct the argument.
This is why legal AI quality needs multiple layers: source accuracy, factual accuracy, issue relevance, procedural fitness, concision and professional judgment.
5. Hancox strengthens the case for matter-centred drafting
A chat interface naturally encourages the production of text. A litigation workspace should instead help the lawyer control what propositions are relevant to the live issue, which evidence supports them, which authorities establish the legal test and what the procedural document actually needs to achieve.
That does not mean software can decide the advocacy strategy. It means the system should make it easier for the professional to work from the matter rather than from an expanding transcript of generated prose.
6. A practical pre-filing check for AI-assisted court documents
- Purpose: what is this document for under the applicable rule or Practice Direction?
- Scope: which issues and grounds are properly before the court or tribunal?
- Evidence: do factual propositions match the actual record and bundle?
- Authority: have cases, statutory provisions and quotations been checked from authoritative sources?
- Relevance: does each section advance the live issue?
- Concision: can repetition, secondary points and background be removed?
- Procedure: do page limits, bundle references, format and filing requirements comply?
- Ownership: can the person submitting the document explain and defend every material proposition?
7. The case is not an anti-AI ruling
The judgment recognises that AI may assist people who do not otherwise have access to professional legal advice and states that AI is likely to have a continuing role in litigation. The message is not “AI is forbidden”. It is that the person submitting the work remains responsible for what the court receives.
8. Poor input material can produce professionally dangerous output
The judgment expressly identifies inadequate or inaccurate input material as one of the risks associated with AI. This matters because a model can faithfully synthesise a bad premise. If the user supplies an incomplete procedural history, an incorrect bundle reference or a mistaken description of the ground of appeal, fluent drafting may reinforce rather than expose the error.
For litigation teams, the lesson is to control the inputs as carefully as the generated text. The live issues, evidence set and procedural posture should be established before asking the system to draft around them.
9. Procedural fitness is part of legal quality
Lawyers often separate substantive correctness from formatting and procedure. Hancox shows why that division can be artificial. Page limits, bundle requirements and the need to focus on the actual grounds of appeal exist to make adjudication manageable and fair.
An AI system that produces a comprehensive but unusable document has not solved the professional task. The correct optimisation target is not maximal completeness of prose; it is a document that performs its procedural function.
10. Concision is a form of professional judgment
Generative systems make expansion cheap. Litigation often requires the opposite skill: deciding which points are central, which authorities are necessary and which background can be omitted. That selection involves judgment about relevance and persuasion.
AI can assist by identifying duplication, mapping propositions to grounds or checking whether each section has a purpose. It should not be allowed to substitute volume for analysis.
11. What firms should take from Hancox
A sensible internal rule is that no AI-assisted court document should leave the firm unless a responsible lawyer can explain the issues it addresses, identify the evidential and legal source for each material proposition, confirm procedural compliance and defend the choice of what was included and omitted.
That standard is deliberately human. The technology may generate, retrieve, compare or restructure. Ownership of the filed document remains with the person and team responsible for it.
12. The judgment also has a knowledge-management lesson
One reason generated litigation documents can expand uncontrollably is that the drafting system is being asked to rediscover the case from scratch inside the prompt. A matter-centred workspace offers a different pattern: maintain the live issues, key evidence, authorities and reviewed propositions separately, then draft from that controlled matter state.
This does not guarantee a good skeleton. It does reduce the temptation to treat a long conversation with a model as the case file. The lawyer can focus the drafting task on the grounds actually pursued and the material that belongs in the permitted bundle.
In that sense, Hancox is relevant not only to responsible AI use but to the architecture of legal work. The safer design objective is to make the case understandable before asking the system to produce the court document.
Conclusion
Hancox should move legal-AI governance beyond the narrow question of fake cases. A litigation document is good only if it is accurate, relevant, proportionate, procedurally fit and professionally owned. AI can assist with preparation; it cannot absorb the responsibility for the finished work.
Related LegalRAG Pro Insights
Professional context. This article discusses legal-technology workflow and professional-risk questions. It is not legal advice and should not be treated as a substitute for checking the current procedural, regulatory and factual position in a particular matter.