Judges discuss legal AI hallucinations and risks

More than 2,000 hallucination cases are recorded

According to a report published on the Relativity blog on November 3, 2026, the event featured five judges with a combined total of more than 85 years of experience in federal, state and British courts. The discussion was moderated by David Horrigan of Relativity.

One of the central topics was AI hallucination, when a system generates false legal or factual content. The post cited a database maintained by Damien Charlotin, which listed slightly more than 400 such cases in October 2025. One year later, the number had risen to 2262.

U.S. federal judge William Matthewman said errors must be acknowledged quickly. According to Relativity's summary, this could include proactively notifying the other side, paying any related costs, withdrawing incorrect documents and then continuing with the case. The report said shifting blame, using AI again to conceal the error or showing no remorse could make the situation worse.

Judge Allison Goddard cautioned that the number of hallucination cases alone does not show how safe AI is. She compared it to a numerator without a denominator. The judge said the phenomenon has made some judges wary of using AI, while some legal communities have banned it. Judge Xavier Rodriguez said models are improving and that a total ban on AI could lead people to use these tools at home.

Judges use AI, but to varying degrees

The report also presented the results of a 2026 survey. In the study, which surveyed 500 federal judges, 38.4 percent of respondents said they never use AI, while 5.4 percent use it daily. The rest fell between the two extremes.

Goddard said the share of non-users was not surprising because judges are supported by clerks and other staff. Matthewman, however, also noted that not every judge necessarily knows when they are interacting with an AI-based tool. Rodriguez said it was surprising that many judges use the technology monthly, weekly or daily.

Michelle Rick, a judge on the Michigan Court of Appeals, said the situation is similar in state courts. According to the report, she had this message for state judges who say they do not use AI: “Yes, you do.” Rick said many people use the technology every day without realizing it. Victoria McCloud, a former judge in England and Wales, described the British approach as urging caution, while more prescriptive rules apply in Ireland and judges are discouraged from using AI.

The panel members agreed that judges and courts need training and clear rules. Rodriguez said courts need practical guidance and specific use cases alongside policies. Matthewman put it this way: “We need to provide guidance. Just as law firms need guidance, courts need guidance too.”

Faster discovery, unsettled protection

The discussion also covered proportionality in civil litigation, particularly Federal Rule of Civil Procedure 26(b)(1). Matthewman supports technology-assisted review and any technology that speeds up the production of relevant materials. He said legal cases take too long and that the tools available could move litigation forward more quickly.

Goddard said the rules are unlikely to change, so how they are applied will be decisive. In discussing British practice, McCloud emphasized targeted discovery and suggested that parties begin with, for example, keyword searches rather than handing all materials to AI at once.

There is still no uniform court position on the protection of conversations conducted with AI tools and work product. Relativity lists three recent decisions: protection was not granted in United States v. Heppner, while it was granted in Warner v. Gilbarco, Inc. and Morgan v. V2X, Inc..

The panel's message to users is that legal applications of AI require oversight and training. Incorrect content must be checked before every filing, while courts and law firms need to establish clear guidance on how the technology may be used. Based on the report, the technology can support the work, but it will not replace judicial decision-making.