Another cautionary tale about AI “hallucinations” in legal research
4 June 2026
Cork and another v Smith [2026] EWHC 1199
This was an unfortunate case of delegation, where the ultimate outsourcing ended up with an AI system. It started as a simple block transfer application to the Chancery Division. Such applications are almost invariably determined without a hearing but the judge gave directions for the matter to be listed as a result of concerns about “misleading statements” made by the applicants’ former solicitors, Pinsent Masons LLP in two letters to the court in connection with the application,
A junior associate solicitor (“Lawyer A”) at the firm had used AI to research a point of insolvency law. The AI generated false text purporting to be Insolvency Rule 12.37(5), suggesting that the court had an express power to grant release to outgoing liquidators. The problem is that this text did not exist.
Lawyer A had asked the AI what the relevant statutory provisions said, rather than checking the legislation directly. The AI itself apparently flagged that the answer might be inaccurate, but that warning was not properly acted upon., and the AI-generated material was not checked against an authoritative source.
There was a further problem; when the court identified the false wording, the response did not simply acknowledge what had happened. A second letter to the court attempted to explain the earlier error as a “summary conclusion” rather than squarely accepting that AI had generated wording which had not been verified. The judge considered that this had made matters worse, because it suggested not just an initial mistake but a continued failure to correct the record promptly and frankly:
“The Purported Text in the 30th March Letter had caused me to be concerned that a cavalier attitude was being taken as to the accuracy of the material that Pinsent Masons were putting before the court. As I say, it struck me as likely to be an AI hallucination, which had not been checked. The attempt to explain it away in what appeared to be an untruthful manner in the 14th April Letter only heightened my concerns.” [para 25]
“it is concerning”, the judge continued,
“that Lawyer A appears to have asked the AI what these sections said, rather than reading an authoritative online resource or book, and does not appear to have checked the AI’s response. Had Lawyer A done so, it would have been readily apparent that the AI was producing nonsense and was unreliable. It perhaps betrays a misguided faith in the AI on Lawyer A’s part but, if that is so, I cannot understand why Lawyer A did not heed the AI’s later warnings about the need to check the provisions to which it referred.” [para 38]
The judge concluded that the conduct looked more like serious lack of care and judgment than deliberate dishonesty, but it was still troubling because the material was presented in a misleading way.
The case reinforces a basic professional rule: AI may assist drafting, but it does not displace the lawyer’s duty to verify accuracy. That is particularly important where the output is used in correspondence or submissions to court and may be read as an authoritative statement of the law. It also sits in the same developing line of authority as earlier AI-related judicial warnings, especially Ayinde v London Borough of Haringey EWHC 1383 (Admin), which had already highlighted the risk of misleading courts with unverified AI-generated citations:
“Freely available generative artificial intelligence tools, trained on a large language model such as ChatGPT are not capable of conducting reliable legal research. Such tools can produce apparently coherent and plausible responses to prompts, but those coherent and plausible responses may turn out to be entirely incorrect. The responses may make confident assertions that are simply untrue. They may cite sources that do not exist. They may purport to quote passages from a genuine source that do not appear in that source.” [para 6 of Haringey]
For practitioners, the case is less about AI itself than about process. The key failures were the absence of source-checking, the failure to heed the tool’s own caution, and the failure to escalate or correct the problem once it was identified.
The practical lesson is that any AI-assisted legal document should be treated as a draft only. A competent lawyer must still verify quotations, authorities, and propositions against primary sources before anything goes out under their name or the firm’s name.
The judgment is a strong reminder that courts expect lawyers to remain personally accountable for what they file or send. Even if the error arises from carelessness rather than bad faith, the consequences can still be serious because the justice system depends on reliability and honesty from legal professionals.
In short, Cork v Smith is now an important cautionary case on AI use in legal practice: it shows that generative tools can increase efficiency, but only if they are controlled by careful human checking rather than treated as a substitute for legal judgment.
As the newsletter LEXOLOGY says,
“Misleading the court, even unintentionally, strikes at the heart of legal ethics. The judge referred to a “cavalier attitude” to accuracy and emphasised that AI does not remove the need for proper research, thought and verification.
The fallout has been significant: public criticism in a judgment, embarrassment for the firm, a referral to the SRA, and the firm agreeing to meet additional costs caused by the issue.
So whilst we must all accept that generative AI can be a valuable tool to support legal work, it cannot think, reason or take responsibility. Lawyers remain the decision-makers, the fact-checkers and the guardians of professional integrity.
This case shows that AI is not, in itself, the problem. Uncritical reliance on its output, compounded by lack of senior supervision, are bigger issues.
For everyone else, this is a useful moment to pause. How is AI being used in your firm? Is it built into your case management system? Are individuals using external AI tools without permission or supervision? How would you know? Do you have an AI policy that explains what is and is not acceptable? Do supervisors know what to look for? Does your culture allow people to admit capacity issues or ask for help before they take shortcuts?
Nobody wants to be the next cautionary tale.”
