By Jay Kt
A database maintained by Damien Charlotin, a legal researcher tracking AI hallucination cases in courts worldwide, listed 1,457 documented incidents as of mid-May 2026, with 1,006 of them in the United States. Lawyers accounted for 556 of the cases. Pro se litigants accounted for 865. The numbers grow weekly.
Todd Mensing of AZA Law, a Chambers Band One-rated commercial trial lawyer who has tried more than 60 cases, watches that growth with the practical skepticism of someone whose work runs on jury persuasion rather than document drafting.
“I think about AI as being a young, overeager, extremely intelligent second-year associate sitting in the corner of your office that you can throw questions to,” Mensing said. “But that genius, highly intelligent second-year associate does not have the experience to understand how things will land with human beings. We’re not writing software code. In our job, we are not engineering pipelines. We are convincing people.”
Where AI Has Moved the Practice
Adoption of generative AI in legal work has shifted from pilot projects to baseline infrastructure across the profession. The American Bar Association’s 2024 Artificial Intelligence TechReport found that 31% of legal professionals report personal use of generative AI at work, up from 27% the prior year, and 21% of firms report firm-wide adoption. AI assistants now sit inside Westlaw, LexisNexis, and most major drafting platforms. Categories of work where adoption has run furthest are document review, deposition analysis, summarization, and routine drafting.
Those categories share a feature: the human in the loop can verify the output against a known record. A lawyer reviewing an AI-summarized deposition can pull the underlying transcript. A lawyer using AI to draft a motion can check each citation against the actual case. An error’s cost is bounded by the diligence of that verification step.
Charlotin’s database documents what happens when the verification step fails. Cases range from $500 monetary sanctions and CLE requirements to a $110,000 sanction issued by a federal judge in Oregon against two lawyers who submitted 23 fabricated citations and eight invented quotations in a single filing, the largest hallucination-related penalty in U.S. legal history as of its issuance. Bar suspensions, public reprimands, notices to clients, and case dismissals as a sanction have followed in cases across at least 33 U.S. jurisdictions.
The Justia 50-state survey of AI ethics rules tracks how state bars have responded. Most have adopted some version of the position that lawyers retain professional responsibility for the contents of AI-generated work product, regardless of which tool produced it. The verification step is the lawyer’s responsibility, embedded in the ethics rules themselves.
Where AI Has Not Moved the Practice
Mensing’s argument is that the categories where AI works well are not the categories where commercial or criminal trials are won and lost.
“I think AI is a useful tool to organize and summarize information,” Mensing said. “But it cannot incorporate the human element an experienced trial lawyer introduces that ultimately means the difference between winning and losing.”
That distinction tracks the structure of how a trial actually moves. Pre-trial preparation involves enormous volumes of documents, depositions, exhibits, and prior witness statements: the kind of work AI processes well. Trial itself is a real-time persuasion exercise in front of twelve human beings whose decisions about credibility, motive, and narrative coherence will determine the verdict.
Mensing’s multiple felony trials illustrate the boundary precisely. A capital murder mistrial that broke a 75-0 Harris County prosecution record (State of Texas v. Johnson) required reading a jury’s nonverbal cues during cross-examination of a deeply prepared state witness. An embezzlement acquittal in a case with three signed confessions (State of Texas v. Petrie) required pivoting the defense theory mid-trial when the worst facts began to flip in the jury’s perception. None of those moves can be delegated to a model that has not watched the jury for the past two days.
“An appreciation of how people absorb information and make decisions at trial is lost when a lawyer is building a case like it’s a debate: point, counterpoint, counter-counterpoint. Jurors don’t think like that. They think about who is telling the truth about what really happened,” Mensing said.
A model that organizes legal arguments into a logical sequence may produce a brief that holds up on appeal. Deciding when to slow down on a cross-examination, so a juror has time to register a contradiction in the witness’s testimony, requires the lawyer’s read of the room.
What the Hallucination Cases Document
A pattern across Charlotin’s database is that hallucination sanctions concentrate in two populations: lawyers who used AI without verifying citations, and pro se litigants who don’t have the legal training to spot fabricated case law. Both groups bring the same basic failure mode. Confidence in the model’s output was substituted for verification of the underlying record.
Defense and plaintiff briefs have been struck. Cases have been dismissed with prejudice as a sanction. Several courts have required client notification, in writing, of the lawyer’s reliance on AI in the filing. A Michigan appellate panel admonished counsel from the bench in oral argument. New York’s First Department suspended an attorney from the bar in part on AI hallucination grounds.
Charlotin’s database covers research and drafting failures, not trial failures. Trial work, where AI penetration remains lower and the human-in-the-loop verification step is structural rather than optional, has produced far fewer documented incidents. That asymmetry tracks Mensing’s framing. The same tools that struggle when no one checks them produce useful drafting and organization output when the lawyer verifies the result.
Why Todd Mensing, AZA Law’s Lead Trial Lawyer, Holds the Line
Mensing’s position is calibrated by what trial work demands. Todd Mensing, one of AZA Law’s lead trial lawyers, has tried more than 60 cases across state and federal courts, and he has learned, in his own description, that the difference between winning and losing rests in the moments AI cannot model: the read of a juror’s body language, the recognition that the witness is improvising, when to go for the kill on cross-examination based on a wavering in the witness’s behavior, or the decision to abandon a planned line of questioning because a different opening has appeared.
His pro bono criminal trials sharpened that read because the stakes left no room for the model’s confidence to substitute for the lawyer’s. A capital murder jury does not deliberate on a brief. It deliberates on a story built across days of testimony, and the lawyer who has not read the jury cannot guess at which version of the story has landed.
“I can tell clients, ‘I’ve been in front of a lot of juries, and I know from experience, this is how this will play,’” Mensing said. “I would not be able to do that with confidence if I had not gone outside the normal channels to gather as much trial experience as possible.”
Technology will keep moving. Whether AI develops a genuine read of human persuasion is a question the trial bar does not need to settle in 2026. The question that matters now is which categories of legal work can be delegated to a confident, brilliant junior associate who has never tried a case, and which cannot.
Mensing’s answer is the same in 2026 as it would have been in 2006, before generative AI existed: trial work runs on the human element, and the human element is where AI’s confidence does not extend.
