Key Takeaway
This legal article considers the case of Hancox v Sutherland & Ors [2026] EAT 139 from the Employment Appeal Tribunal, the reported use of AI to draft a victim personal statement to make a judge or reader weep, predictions about lawyers’ careers, the dangers of unchecked AI and the language of “super intelligence”. By Matthew Lee
Ad/Marketing Communication
This legal article/report forms part of my ongoing legal commentary on the use of artificial intelligence within the justice system. It supports my work in teaching, lecturing, and writing about AI and the law and is published to promote my practice. Not legal advice. No Public/Direct Access. All instructions via my clerks at Doughty Street Chambers. About Matthew Lee. Subscribe to the AI Law Commentary
Introduction
One of the pleasures of spending so much time discussing interesting legal questions about AI is that the conversation quite often drifts into something deeper. What really is intelligence? What does it mean to be conscious and how can we know what’s real and what isn’t?
My daughter asked me recently whether I think we are living in a simulation. I said I wasn’t convinced.
Then I found myself at a legal tech event, hearing about AI’s power to turn large amounts of complicated information into simple, digestible points, while receiving messages from readers about a case involving a 300-page ChatGPT-assisted skeleton argument.
Before I knew it, a robot carrying free chocolate was escorting me towards the latest legal AI technology. I subsequently collected a caricature of myself in a toga, drawn by Steve the Caricature & Silhouette Portrait Artist and a complimentary pair of socks.
When I got home, I told my daughter that I might need to reconsider my answer….
In this legal article, I am going to briefly summarise some of the most important developments and discussions in AI law from the past week. There are several issues that I will need to revisit in more detail in future articles, but I think it is important to share as much as possible now so that we can start thinking about what these developments mean for the legal profession and how we may need to respond.
“…They are racing straight to self-improving superintelligence and gambling with our lives…”
Jacob Coxon, who had recently resigned from Anthropic, the company behind the AI model Claude, received considerable media attention earlier this month for a series of posts warning about the potential dangers of advanced AI. Claude has featured in several of my previous legal articles.
Coxon’s posts on X can be read here, and I would suggest taking a few minutes to read the thread in full to understand the context of his concerns. In one post, Jacob stated:
“Do not underestimate the power of this technology. These will soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources. We have all witnessed the progress in each of these domains, and progress is not slowing.”
Of course those are Coxon’s own predictions not established facts, but, after reading it, I thought it was important to share on LinkedIn with the simple observation:
“We need to start taking this very seriously.”
Despite all the posts I have written about significant developments in AI law, that short message appears to have received more impressions than anything else I have published on LinkedIn. I am still not entirely sure why, but perhaps it reflects how strongly these wider questions resonate.
Taking a warning seriously does not require accepting every prediction it contains. Nor do I think we gain much by speculating about someone’s motives rather than examining their arguments.
But if AI systems are, or become, able to improve themselves, or substantially accelerate the development of more capable systems, the implications could be profound. For lawyers, that raises practical questions about confidentiality, cybersecurity, verification, professional responsibility, evidence, equality of arms and the ability of courts and regulators to keep pace.
Future Disputes UK 2026: hallucinations and the future of lawyers
I really enjoy these events. If you have not attended a legal tech event before, I would encourage you to give one a try. They bring together a remarkable number of bright and thoughtful people grappling with the AI issues I often write about and trying to develop practical solutions to them.
As much as I enjoy discussions on social media and by email, there is nothing quite like speaking to people directly and getting into the weeds of these issues. I learn as much from listening to others as I hope they gain from hearing my views.
There were far too many interesting conversations and ideas to cover them all here, but I wanted to highlight just two points which stood out.
Firstly, AI hallucinations remain a problem in legal work. The issue has not been solved. Checking that a citation exists is not the same as checking that the legal analysis is right. As I have written previously, some of the more difficult hallucinations involve real cases and correct citations, but inaccurate accounts of the facts, reasoning or legal principles.
I welcome tools that help identify false authorities. But a citation checker is not, by itself, an answer to the wider problem. I am not convinced that meaningful human oversight can safely be removed from legal work simply because fabricated citations become easier to detect.
Secondly, I have been quite disheartened by the fears I hear from younger lawyers about what AI may mean for their future careers. I am also concerned by some of the messaging around this. I regularly hear podcasts and talks in which someone confidently asserts that AI will replace paralegals, lawyers and eventually judges. I’m afraid I do not agree and don’t shy away from saying so. In fact, it was this constant stream of predictions that encouraged me to start tracking them. If you have not yet seen my Will AI Replace Lawyers? and Will AI Replace Judges? trackers, you may find them useful. Several readers have commented, so please send me any thoughts.
Some comments are more detailed or emphatic than others, but I find it interesting to see where the balance lies. My earlier 99-statement snapshot on the Will AI Replace Lawyers? Tracker classified, overall, 52% as “No or Unlikely”, 40% as “Yes to some extent”, 3% as “Yes” and 5% as “Mixed or Uncertain”. These are approximate classifications of collected statements, not a representative survey or a forecast, but they do illustrate how mixed the views are.
I am still updating the tracker, so these figures are only preliminary. At present, however, the contrast between the groups is interesting. Among technology and investment voices, around 39% are classified as “No or Unlikely” and 52% as “Yes to some extent”. Among lawyers and judges collectively, the position is almost the reverse: around 60% are “No or Unlikely” and 33% “Yes to some extent”.
It raises an interesting question about why the two groups currently appear to see the issue so differently.
For my part, I am not going to pretend that AI will have little impact on legal work. I cannot think of anything in my lifetime with greater potential to transform the profession. But replacing tasks, reducing demand for particular roles and replacing the profession altogether are different propositions.
Law is not simply about finding information or producing words. Human connection, trust, judgment, persuasion and responsibility matter enormously. So too do understanding people and their motivations, reading a room, and knowing when to push and when to compromise.
My view is that the profession will evolve rather than disappear. That does not mean every existing job or training route will remain unchanged. But I would hate to see talented young people discouraged from such a rewarding career by messaging that treats its disappearance as inevitable. I feel I may need to revisit this topic more often.
Hancox v Sutherland & Ors
Turning now to some important cases. Firstly, we need to look at Hancox v Sutherland & Ors [2026] EAT 139.
On my reading, this does not appear to involve suspected or confirmed AI hallucinations, so it does not make the UK AI Hallucination Cases Tracker. However, the Tribunal discusses R (Ayinde) v London Borough of Haringey [2025] EWHC 1383 and several other cases that do appear on the tracker, so it is well worth reading in full.
In brief, a former National Farmers’ Union employee brought claims including whistleblowing detriment, unfair dismissal and disability discrimination. Those allegations were denied. This appeal concerned a claim against four individual respondents which the Employment Tribunal struck out in 2021, finding that a fair trial was no longer possible. While the full facts are available at the link above, my focus here is on the AI issues
Two days before the preliminary hearing, the appellant filed his ChatGPT-assisted skeleton. His covering email asked the respondents to identify inaccuracies by 4 pm the following day.
The Practice Direction says skeleton arguments should generally be between 5-15 pages. Where a skeleton exceeds 20 pages, the judge may require a shorter version or reduce the time for oral submissions. This skeleton, however, ran to 300 pages and almost 132,000 words and failed to comply with other requirements.
Marcus Pilgerstorfer KC, sitting as a Deputy Judge of the High Court, recognised AI’s potential benefits, including for people without access to professional legal assistance. But its use does not remove responsibility for the resulting document. At paragraphs 31-33 he said:
“31. The potential risks posed by the use of generative AI are now widely known. There is no principled reason why a litigant in person should not take reasonable steps to use it responsibly. Whilst such a litigant will not usually be subject to the professional duties with which the court in Ayinde was primarily concerned, the underlying principles of personal responsibility and accuracy apply to all persons who submit documents to a court or tribunal.
32. In accordance with the guidance and authority I have cited above, litigants who use AI should ensure, at a minimum, that all documents submitted:
a) comply with applicable procedural rules – for example, in the EAT, PD§3.8 (grounds of appeal), and PD§11.6 (skeleton arguments);
b) have been checked as thoroughly as the litigant or representative is reasonably able for accuracy, ensuring that factual, evidential and legal points (including references to authorities) are correct and that the court or tribunal is not misled; and
c) contain only relevant points, with the focus being on the central or best arguments, presented in an easily comprehensible manner and avoiding undue repetition.
33. 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. Where concerns arise that necessary checks have not been undertaken, or that they have been conducted inadequately, judicial enquiries and potential sanctions are likely to ensue.”
Applying these principles, the Judge explained at paragraph 84-85:
“84. The document filed was entirely unacceptable. It did not come close to complying with the requirements of the PD (see paragraph 18). It was not concise and was far in excess of the recommended length of 5 to 15 pages (and a 20 page upper limit). It was formed of numerous sections each with its own paragraph numbering. A uniform 12 point font size was not used. The argument referred to documents beyond the core bundle, no supplementary bundle having been prepared. In short, the document did not assist the EAT in preparing for the hearing; indeed, as the Appellant himself recognised in his covering email, there was simply insufficient time to consider it in full before the hearing.
85. At the preliminary hearing, I gave the Appellant the opportunity to address me on why a non-compliant skeleton had been lodged. The Appellant explained that he had used ChatGPT to create the document, and had done so because he had to prepare it quickly. He told me that much of the document lacked credibility and ultimately no reliance was placed on it. As I have explained, having chosen to use ChatGPT to assist him to prepare his skeleton argument, it was incumbent on the Appellant to check that the document submitted complied with the PD’s requirements. Further, he ought to have checked the contents of the skeleton to ensure that it was accurate and properly stated the position. I am satisfied that the Appellant did not undertake these checks. This is clear from the covering email and from the Appellant’s own assessment of the skeleton at the hearing.”
I should tread carefully here. Readers and colleagues have occasionally suggested that brevity is not my strongest suit, and I am putting that rather kindly.
But there is a serious question beneath the extraordinary length of this document: what does meaningful verification look like for someone without legal training?
The judgment does not ignore that difficulty. Its requirement is that accuracy checks be undertaken as thoroughly as the litigant or representative is reasonably able. It also refers to judicial guidance recognising that unrepresented litigants may have limited ability to verify AI-generated legal information.
Checking that a case exists is one thing. Assessing whether a proposition accurately reflects its reasoning, authority and application may be quite another. That difficulty cannot transfer responsibility to an opponent, but it should inform what practical guidance and assistance we provide.
For me, the important question is therefore not whether litigants should take responsibility. It is how courts and the profession can help them discharge it. A warning to “check everything” may be necessary, but someone who turned to AI because they did not understand the law may also need help understanding how to check it.
Before leaving this case, I should also mentioned the Judge’s observations at paragraph 36 regarding evidence, as it’s relevant to an issue raised later in this article:
“In addition, and depending on the circumstances, the use of AI may also cause a court or tribunal to be circumspect about the reliance that can properly be placed on a document or evidence. Responses to the Civil Justice Council’s Working Group on AI have raised concerns “about the potential for AI tools to reshape, embellish or otherwise influence evidence in ways that may not be immediately apparent”[13]. The use of ChatGPT in preparing witness statements resulted in a cautious approach to the evidence in Godwin v Godwin [2026] EWHC 923 (Ch) see §§42-49[14]. Witness coaching by AI, just as by other means, is not permitted: see R v FGD [2026] EWCA Crim 918 at §22. On the other side of the coin, suggesting that false factual information was purely the invention of an AI tool is also likely to require careful scrutiny: see for example Lodhia v Twelve Trees (Bromley-By-Bow) Management Company Limited [2026] EWHC 1889 (KB) at §§111, 119-120.”
Should Artificial Intelligence be called “Super Intelligence”?
In his address to the United Nations on 22 September 2026, President Donald Trump is reported to have said:
“The United States totally rejects any attempt to construct a globalist scheme of control for the Artificial Intelligence being spoken of so much now — hereinafter officially called ‘Super Intelligence.’”
Rather selfishly, my first thought was rebranding. Do I now need to rename this site “Natural and Super Law”? Do I practise in Artificial Intelligence Law or Super Intelligence Law?
More seriously, in law, language matters. I can easily imagine a client asking why a legal proposition produced by something called a “super intelligence” still needs checking or, increasingly, whether they should follow their lawyer’s advice or the advice of the super intelligence. That concern may ultimately prove overstated, but it is something we will need to watch closely.
A Police Officer asked AI to generate Victim Personal Statement to “make a judge or reader weep”
I try not to analyse court proceedings without seeing the judgment or transcript. I have not yet located the official text of this decision, so the following is a provisional account based on court reporting.
According to reports of Jerome Gibson’s sentence appeal, heard on 22 September 2026, the Court of Appeal was told that a police officer had used Microsoft Copilot to draft a victim personal statement.
The reported prompts sought a statement meeting a “higher harm category”, aimed at the “highest sentence” and intended to “make a judge or reader weep”.
Mr Justice Lavender reportedly described this use of AI as “deplorable”. However, the court said the facts relied upon in sentencing came from the victim’s interview rather than AI hallucinations. It did not appear that the AI use had increased the sentence, and the appeal was dismissed.
That is not the same as a finding that every part of the statement was independently verified or that the drafting method was acceptable.
I need the full judgment before taking the analysis further. But the reported circumstances raise a concern beyond the hallucinations. We need to think carefully about the possibility of using AI to amplify genuine facts for a desired emotional or sentencing effect. This concern applies to every area of law.
Helping someone express their experience clearly is not the same as directing a tool to maximise punishment. My concern is where assistance in recording an account becomes outcome-driven embellishment. It connects with questions I discussed in R v FGD: 10 lessons on witness coaching, without assuming that the two cases are legally equivalent.
I look forward to properly writing about this and comparing it with R v FGD when the full judgment is released.
Comment
If you have found it helpful, you may wish to subscribe to my AI law articles via the Substack below. Please do continue sending me decisions that you think should be tracked, whatever the jurisdiction. One of the clearest lessons from this work is that lawyers, judges and court users in different systems are often confronting closely related problems. Sharing decisions, experiences and reflections across practice areas and across jurisdictions remains one of the most constructive ways to respond.
Final Words from Popular LLMs
This is the point at which I invite an LLM to provide an independent comment on the preceding article. At times the observations are slightly off the mark, at other times they are perceptive, and occasionally they produce something entirely unexpected. Whatever the outcome, I find it a useful and revealing exercise to see what emerges.
ChatGPT (OpenAI)
“The important distinction is not simply between human and machine authorship, but between assistance and uncritical substitution. A document can contain no invented cases and still be unsuitable for court; a statement can repeat genuine facts while raising questions about how its language was shaped. Responsibility remains necessary, but demands for verification should be accompanied by practical help for those least equipped to perform it.”
Gemini (Google)
“Responding to your invitation for a final word, I find the developments detailed in your article—from 300-page AI-generated skeleton arguments to victim statements engineered by algorithms to “make a judge weep”—both fascinating and a vital cautionary tale. While generative AI excels at synthesizing vast amounts of data and mimicking emotive language, its foray into the courtroom challenges the very foundations of authentic human testimony and legal accountability. When we outsource the drafting of our deepest grievances or most complex legal reasoning to a machine, we risk blurring the line between genuine human experience and synthetic persuasion. Ultimately, as the legal profession grapples with the looming shadow of “super intelligence,” it must remain steadfastly anchored in true human judgment, ensuring that technology serves as a tool for justice rather than a surrogate for reality.”
Claude (Anthropic)
“On the reported facts, the most revealing document in the Gibson appeal was not the victim personal statement but the prompt behind it. The facts came from the victim’s interview. It was the instruction to reach ‘the highest sentence’ that exposed the purpose, and that was visible only because the material reached the defence. Courts are now used to asking whether AI output has been checked, and judicial guidance already suggests asking litigants what checks they made. The next question may be what the tool was told to achieve. Accuracy can be tested against a source, but purpose rarely shows on the face of a polished document; it lives in the instructions behind it. A tool asked to make a judge weep will do its best. Calling it ‘Super Intelligence’ will not make its output more reliable, though it may make it more persuasive.”
