What went wrong: AI hallucinations or AI use?

Another law firm has found itself answering to the court to address AI hallucinations. Cork v Smith caused the judge to express concerns that a “cavalier attitude” was being taken as to the accuracy of the information put before the court.

This article notes salient takeaways from the recent cases of Cork v Smith and Rodney v Gee’z Micro Bar.

• Cork v Smith [2026] EWHC 1199 (Ch): Two misleading letters concerning an inaccurately cited statute in a routine insolvency application were put before the High Court. The law firm in question was directed to explain the circumstances in which the letters came to be generated.
• Rodney v Gee’z Micro Bar [M00DD215]: Three documents were filed by a law firm in support of an application for permission to appeal which contained incorrect citations of legal authorities. The suspicion that these documents had been created using AI was flagged to the court by the respondent to the appeal as “deeply troubling”.

The risks of AI hallucinations

Given that AI has the potential to produce output that can be inaccurate, biased, outdated or entirely fabricated (hallucinations), we highlight some salient points from the above judgments where those using AI have unfortunately fallen foul.

1. Legal representatives are responsible for their work product and material they put before a court/tribunal. Those using AI have a professional duty to check the accuracy of the information by reference to authoritative sources and before using it in the course of their professional work (eg, before advising clients or before a court). Urgency or work pressures will not excuse a failure to check the accuracy of the material.
2. This duty rests on not just the legal representative who may be using AI but also any legal representative who is relying on the work of another who has done so. Relatedly, it is then vital that those relying on a work product that used AI, are informed that AI was in fact used. Comparably this is no different from the responsibility of a legal representative to check a trainee’s work or supervise. Using AI is not yet a replacement for supervision.
3. Placing erroneously cited authorities, including citations used to support a legal proposition which they do not in fact support, before the court creates a huge additional burden and undermines the confidence the court may have in filed written submissions.
4. Text attributed to AI can appear correct and persuasive; however, it can be, and often on a number of occasions is, plainly wrong or in the very least misleading. AI programs can provide warnings of their own limitations, for example flagging that “I am not fully confident that I am reproducing the exact statutory wording…” or “The last thing you want is to cite a provision to the court with inaccurate wording.” A legal representative that fails to act on this type of warning opens themselves up to criticism.
5. Records of how an AI program is used, including instructions provided by the legal representative and the AI program’s responses, can be ordered by the court to examine the accuracy of material put before it.
6. While contempt of court proceedings were avoided in these two cases, they confirm the position in Ayinde that the grounds for assessing contempt of court is a two-stage process. “The first stage is an assessment of whether a contempt may have been committed, the second is an evaluative judgment as to whether contempt proceedings should be initiated.”

The failure to supervise and check AI raises conduct issues and the potential of being held in contempt of court. These judgments provide a broader warning about professional responsibilities, supervision and the importance of the court being able to trust material put before it by officers of the court. They also serve as a broader reminder that whilst AI can be an incredibly useful tool for legal professionals, the accuracy of its outputs should always be thoroughly checked. 

Our content explained