agentic-ai
"I didn't know AI could hallucinate" just stopped working as a legal defence
A New Mexico defence lawyer offered that explanation this month after citing fabricated testimony from made-up witnesses. Courts in Mississippi and California have already rejected the same excuse in writing, and the pattern is what matters, not the individual case.
"I didn't know that AI could hallucinate facts," a New Mexico defence lawyer told a court on 11 September, after citing fabricated testimony from witnesses who do not exist, according to Ars Technica. The explanation is not new. What has changed is that courts have stopped accepting it.
The excuse used to work
Early hallucination cases, starting with the 2023 Mata v. Avianca sanctions in New York, treated ignorance of the technology as a mitigating factor: the lawyer was fined, admonished, and allowed to continue practising once he expressed remorse. That template held for two years. A database tracking documented instances of AI hallucinations in filed legal documents counted 51 in December 2025 alone, and the monthly count has not fallen since.
The excuse stopped working
In June, a federal judge in Mississippi sanctioned attorneys on both sides of a contract dispute for filing briefs with fabricated citations. One attorney said she was unaware AI could produce hallucinated cases; the court called the explanation "insufficient and incredulous" and barred her from the district for two years. A California appellate panel, in a separate opinion published the same year, went further: it published its ruling specifically "as a warning" to the profession, and noted that the same firm involved had already been sanctioned once before for the identical failure.
A defence that worked in 2023 because the risk was novel does not survive 2026, when the risk has a name, a case law, and a court that already published a warning about it.
What changes for anyone signing off on AI-assisted work
The professional standard has moved from "verify AI output because it might be wrong" to "verify AI output because a court has already told you, in writing, that not knowing is no longer a defence." This applies beyond law: any regulated profession where a practitioner signs their name to AI-assisted work product, medicine, accounting, engineering, architecture, is accumulating the same body of precedent that a New Mexico or Mississippi court is building for law. The organisations still relying on individual diligence rather than a mandatory verification step in the workflow are the ones a future sanctions order will name first.
Sources
Written by the Devence Lab research team.
