© 2026 JSH Law Ltd. AI-assisted original artwork created for JSH Law using supplied JSH Law branding and portrait photography. All rights reserved.AI Can Write 300 Pages. That Does Not Make It Advocacy.
A litigant filed a 300-page ChatGPT-assisted skeleton argument. The real lesson from Hancox v Sutherland is not that AI has no place in court, but that technology cannot replace verification, evidential discipline, relevance or human judgment.
Legal AI | Access to Justice | Court Documents | Evidence
AI Can Write 300 Pages. That Does Not Make It Advocacy.
What Hancox v Sutherland tells us about ChatGPT, litigants in person, court documents and the much harder question of what happens when artificial intelligence starts shaping evidence.
JSH Law | 1 October 2026
Artificial intelligence has made it extraordinarily cheap to produce legal words. It has not made legal judgment cheap.
A litigant in person recently filed a 300-page skeleton argument containing almost 132,000 words in the Employment Appeal Tribunal. It had been created using ChatGPT.
The problem was not simply that artificial intelligence had been used.
The Employment Appeal Tribunal did not say that litigants in person must avoid AI. Nor did it decide that every AI-assisted court document is inherently unreliable.
The real problem was much more fundamental.
A court document still has to do the job of a court document.
It must comply with procedure. It must be accurate. It must identify the issues that actually matter. It must distinguish strong arguments from weak ones. It must help the judge understand the case rather than bury the case beneath hundreds of pages of generated text.
In Hancox v Sutherland & Ors [2026] EAT 139, the tribunal said the 300-page document had served to “obscure rather than to illuminate”.
That phrase matters.
Because this is not really a story about whether AI will replace lawyers.
It is a story about something much more important:
What remains uniquely human when machines can generate almost unlimited legal material?
The quick answer
AI can be extremely useful in legal work. For litigants in person in particular, it may help explain unfamiliar terminology, organise information, identify questions, create timelines, improve readability and turn an intimidating blank page into a workable first draft.
But the person putting a document before a court remains responsible for what it says, whether it is true, whether the law is accurate, whether it complies with the rules and whether the material actually assists the court.
1. What happened in Hancox v Sutherland?
Hancox v Sutherland & Ors [2026] EAT 139 was an Employment Appeal Tribunal case concerning an appeal following the strike-out of claims against individual respondents.
Shortly before a preliminary hearing, the appellant, who was representing himself, filed what was described as a skeleton argument.
It was anything but skeletal.
The appellant explained that he had used ChatGPT because he had needed to prepare the document quickly.
He also invited the respondents to identify factual inaccuracies, incorrect quotations, mistaken dates and other problems with the document by the following afternoon.
That attempted transfer of responsibility was significant.
It is not the other party’s job to fact-check a document before you put it before a judge.
Nor is it the judge’s job.
The EAT emphasised that litigants who use generative AI remain personally responsible for ensuring that documents:
- comply with the applicable procedural rules;
- have been checked as thoroughly as reasonably possible for factual, evidential and legal accuracy; and
- contain only relevant points, concentrating on the central or best arguments in an intelligible way without unnecessary repetition.
That is a remarkably useful summary of responsible legal AI use.
Hancox is an Employment Appeal Tribunal decision. Its particular Practice Direction and page requirements should not simply be transplanted into family proceedings. But its central principles of personal responsibility, accuracy, relevance and procedural compliance have much wider significance.
2. The danger is not just AI inventing cases
Much of the legal profession’s early discussion about generative AI focused on hallucinations.
Fake cases. Invented quotations. Authorities that sound completely plausible but do not exist.
Those risks are real.
In R (Ayinde) v London Borough of Haringey; Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin), the Divisional Court dealt directly with serious concerns arising from actual or suspected use of generative AI in legal work where material had not been properly checked.
The message from Ayinde was clear: the availability of AI does not dilute professional or personal responsibility for material placed before a court.
But Hancox exposes another problem.
A document does not need to contain a fake case to be a bad court document.
It can contain real authorities and still be useless.
It can contain technically correct propositions while missing the actual issue.
It can repeat the same argument twenty different ways.
It can identify every conceivable point while failing to distinguish the decisive point from the irrelevant ones.
And it can make an already complicated case almost impossible to understand.
Volume is not analysis.
Complexity is not sophistication.
More law is not necessarily better advocacy.
3. AI can assist with the work. It cannot take responsibility for it.
This may become one of the most important principles governing the use of AI in the justice system.
Artificial intelligence can help perform a task.
It cannot become the accountable person behind the task.
If an AI-generated document contains a false citation, ChatGPT does not appear before the judge to explain it.
If a witness statement stops reflecting the witness’s genuine memory, the language model is not cross-examined.
If confidential information has been entered into an inappropriate public AI system, the machine is not the person who owed the duty of confidentiality.
And if 300 pages of generated argument obscure the point the judge actually needs to determine, the technology does not bear the procedural consequences.
The person using it does.
The governing idea is simple
Delegating the drafting does not delegate the responsibility.
4. The uncomfortable access-to-justice paradox
There is another side to this debate that should not be ignored.
Litigants in person do not generally turn to AI because they have unlimited access to legal advice.
They use it because the law is complicated, the procedure is intimidating, legal representation may be unaffordable, and they are trying to navigate a system largely designed by professionals for professionals.
Generative AI can therefore be a genuinely important access-to-justice tool.
It can explain terminology.
It can help organise disordered information.
It can suggest a structure for a chronology.
It can turn a distressed stream of consciousness into something clearer.
It can help someone work out the difference between evidence, allegation, submission and opinion.
It may give somebody who has never written a formal document enough confidence to begin.
Those benefits should not be dismissed.
But they create a difficult question:
If somebody is using AI because they do not know the law, how are they supposed to know when the AI has got the law wrong?
Telling an unrepresented person simply to “check the answer” does not fully solve that problem.
Checking requires a reliable source.
It requires knowing which source is authoritative.
It requires recognising when an answer is legally incomplete.
And sometimes it requires knowing that the question asked of the AI was the wrong question in the first place.
That is where legal judgment still matters enormously.
5. The harder problem: when AI starts shaping evidence
Drafting legal argument is one problem.
Evidence is potentially much more sensitive.
In September 2026, the Court of Appeal reportedly heard a criminal sentence appeal in which a police officer had used Microsoft Copilot when preparing a victim personal statement.
According to press reporting of the hearing, prompts included requests aimed at achieving the highest sentence and making a judge or reader emotional.
The reported circumstances prompted serious judicial criticism.
At the time of writing, JSH Law is treating this part of the discussion as a report of Court of Appeal proceedings rather than relying upon it as a published judgment or authoritative statement of law.
Nevertheless, the issue it exposes is important.
There is a difference between using technology to help somebody communicate what they genuinely remember and using technology to optimise their account towards a desired legal result.
Consider the difference between these prompts:
Potentially useful organisational assistance.
Potentially legitimate, but care is needed to preserve the witness’s own evidence and language.
Now the technology is being directed towards an outcome rather than simply helping communicate evidence.
That distinction matters.
The Civil Justice Council has already identified witness statements as an area requiring particular consideration.
Its June 2026 update on the use of AI in preparing court documents recorded concern about preserving the authenticity, integrity and personal recollection of the witness, including the possibility that AI may reshape or embellish evidence in ways that are not immediately obvious.
That is exactly the right concern.
AI should help a witness communicate their evidence. It should not manufacture the evidence the case needs.
6. Why this matters particularly in the Family Court
Family proceedings make these questions especially difficult.
The evidence is often deeply personal.
Parents may be writing statements while frightened, traumatised, angry or exhausted.
Domestic abuse cases may involve patterns of conduct rather than one easily identifiable incident.
There may be hundreds of messages, years of history, safeguarding referrals, school records, medical information, police material, Cafcass assessments and competing explanations of the same events.
A litigant in person may understandably ask AI to help make sense of it.
Used carefully, that can be valuable.
But the Family Court also depends heavily upon the court being able to distinguish:
- what actually happened;
- what a witness personally remembers;
- what is supported by contemporaneous evidence;
- what somebody else has reported;
- what is disputed;
- what is inference;
- and what is legal argument.
Generative AI can blur those categories if it is used carelessly.
It is particularly good at producing smooth, coherent prose.
Real life is often not smooth or coherent.
Memory is messy.
Contemporaneous messages contain contradictions.
Trauma may affect how events are recalled and communicated.
Relationships develop over time.
Evidence does not always arrive neatly packaged around the legal test.
That untidiness may itself matter.
The danger of retrospective perfection
AI can make a statement sound more consistent, more articulate and more legally focused than the witness’s underlying material really is.
That may look helpful. But if the process removes uncertainty, introduces language the witness would never use, strengthens an allegation beyond the source material or unconsciously aligns the evidence with a legal test, the resulting document may become less reliable rather than more.
7. Child-focused proceedings make outcome-driven prompting particularly dangerous
The same problem becomes even more acute when the material concerns a child.
Imagine asking an AI system:
“Rewrite these notes so Cafcass can see that the child is being alienated.”
That instruction already contains the conclusion.
The technology is then being asked to organise the facts around that conclusion.
A safer question would be:
“Organise these events by date. Separate what I personally observed from what the child said, what another person reported and what the records show. Do not add facts or draw conclusions.”
That is a fundamentally different use of the technology.
One asks AI to prove a theory.
The other asks AI to help expose the evidence.
In safeguarding work, those are not interchangeable.
8. Useful AI, risky AI and AI that should set alarm bells ringing
| Use | Example | Risk |
|---|---|---|
| Generally useful | Turning supplied dates into a chronology | Still check every date and event |
| Generally useful | Explaining unfamiliar procedural terminology | Verify against current rules and guidance |
| Useful with care | Improving the structure of a witness statement | Must preserve the witness’s actual recollection and evidence |
| Higher risk | Asking AI to identify legal arguments | Law may be incomplete, outdated or fabricated |
| High risk | Asking AI to rewrite evidence to satisfy a particular legal test | Evidence may become outcome-driven or embellished |
| Do not do this | Filing generated cases, quotations or factual assertions without checking them | Risk of misleading the court and procedural consequences |
| Do not do this | Uploading confidential or identifying case material indiscriminately into public AI tools | Confidentiality, privacy and data-protection risk |
9. There is another problem: confidentiality
Accuracy is not the only issue.
The Courts and Tribunals Judiciary’s current guidance on artificial intelligence also emphasises confidentiality and privacy.
That should matter enormously to anyone using AI in family proceedings.
Family cases can contain highly sensitive information about:
- children;
- domestic abuse;
- health and mental health;
- schools;
- social care;
- police involvement;
- addresses and contact details;
- medical records;
- financial information;
- and allegations concerning third parties.
Before putting case material into any AI system, the user needs to understand what system they are using, how the information is handled, whether it is retained and whether entering the information is appropriate.
10. The JSH Law AI Court Document Check
Before filing or sending a court document that has been prepared with significant AI assistance, ask six questions.
1. SOURCE — Where did this proposition come from?
If the AI gives you a case, statute, rule, quotation or guidance document, have you found and read the real source?
2. ACCURACY — Is it actually right?
Have names, dates, quotations, factual assertions, legal tests and procedural requirements been checked?
3. EVIDENCE — Can I prove this?
Does each important factual assertion come from the witness’s own evidence, a document, a record or another identifiable source?
4. PROCEDURE — Is this the right document in the right form?
Does it comply with the relevant rules, Practice Direction, court order, page limit, filing requirement and deadline?
5. RELEVANCE — Does the judge actually need this?
Does the point help determine an issue the court has to decide, or has AI simply generated another plausible paragraph?
6. HUMAN VOICE — Whose evidence is this now?
If this is a statement or personal account, does it still genuinely reflect the person’s own recollection, experience and meaning? Or has the AI begun to improve, embellish or reshape the evidence?
This sits naturally alongside the wider JSH Law Six-Question Check.
Because whether the problem is poor professional practice, procedural overload or indiscriminate use of artificial intelligence, the underlying discipline is the same:
identify what happened, identify the evidence, identify the issue, and make the court’s job easier.
11. If you are a litigant in person using ChatGPT or another AI tool
You do not need to be frightened of using AI.
But you should use it as an assistant rather than an authority.
Before anything goes to court:
- read every word yourself;
- remove anything you do not understand;
- check every case and legal authority from the original source;
- check every date and factual assertion;
- make sure the document complies with the relevant court order and procedural rules;
- separate facts from allegations and submissions;
- make sure a witness statement remains the witness’s own evidence;
- delete repetition;
- ask whether every section helps the judge decide something;
- and do not upload sensitive case material without understanding the privacy implications of the tool you are using.
There is also a useful final test:
If the judge asks, “Why is this paragraph here?”, can you answer?
If not, it probably should not be there.
12. What? So what? Now what?
What?
Courts are now confronting AI-generated legal material in real proceedings. Hancox demonstrates that the problem is not confined to invented authorities. Scale, relevance and procedural discipline matter too.
So what?
Generative AI dramatically increases the amount of material a person can produce. Courts therefore need human users to become better at verification, selection, judgment and restraint.
Now what?
The next frontier is likely to be evidential integrity: witness statements, expert evidence, personal accounts and the point at which AI assistance begins changing rather than merely communicating the underlying evidence.
13. AI will change legal work. That makes judgment more valuable, not less.
Artificial intelligence is not going away.
Nor should the legal system respond by pretending that the only safe legal document is one written entirely without technology.
Lawyers already use search tools, document automation, transcription, databases, templates, electronic disclosure systems and increasingly sophisticated legal technology.
Generative AI is another major step in that progression.
But it changes the economics of legal information.
Producing words is becoming extraordinarily easy.
Producing the right words, supported by the right evidence, addressing the right issue, in the right document, at the right time remains much harder.
That is why the skill that becomes more valuable in an AI-saturated justice system may not be drafting.
It may be judgment.
Knowing what matters.
Knowing what does not.
Knowing when something needs checking.
Knowing when the evidence does not support the argument you hoped to make.
Knowing when an apparently persuasive paragraph should be deleted.
The future lawyer’s advantage may not be the ability to produce more material.
AI has already solved that problem.
The advantage will be knowing what to leave out.
And for litigants in person, the same principle applies.
The goal is not to create the most impressive-looking document.
It is to help the court see the case clearly.
Because a 300-page document can look sophisticated.
It can even contain a great deal of law.
But if the judge cannot find the issue that matters, the technology has not improved the advocacy.
It has simply automated the noise.
Sources and further reading
- Hancox v Sutherland & Ors [2026] EAT 139, Employment Appeal Tribunal, 17 September 2026.
- R (Ayinde) v London Borough of Haringey; Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin), Divisional Court.
- Courts and Tribunals Judiciary, Artificial Intelligence (AI) Guidance for Judicial Office Holders, updated October 2025.
- Civil Justice Council, Use of AI in Preparing Court Documents — update on consultation findings, June 2026.
- Matthew Lee, Doughty Street Chambers, Will AI Replace Lawyers? A 300-page ChatGPT Skeleton Argument and a Police Officer asks AI to create a VPS to make “a judge or reader weep”, September 2026.
- Press reporting concerning the Court of Appeal proceedings in the Jerome Gibson sentence appeal, September 2026. This is referred to as reported proceedings rather than a published authority.
