Courts increasingly harsh on AI errors

Invented case laws, false citations, plagiarism: the increasingly uninhibited use of artificial intelligence (AI) by lawyers has led to an explosion in the number of "hallucinations" in pleas, forcing justice to crack down, particularly in the United States, where the Supreme Court of Florida is introducing sanctions.

When an American judge discovered, at the start of the year, seven invented or distorted precedents in a lawyer's brief on a personal injury case, the latter admitted to having used Claude, Anthropic's AI chatbot, to draft the document, then correct it. The lawyer, although experienced, explained that he did not check the content ultimately generated by the AI ​​before filing it, thinking that it would be accurate.

“Ignorance of the risks linked to the use of AI is no longer an excuse,” ruled Judge Jerry Edwards Jr of the Louisiana district court in charge of the case, imposing a fine of 1,000 dollars (864 euros) on the lawyer and the obligation to follow a three-hour training course on legal practice assisted by AI.

Similarly, last week, federal judge Sharion Aycock sanctioned lawyers from both sides whose AI-written briefs in a civil case in Mississippi cited nonexistent cases. In addition to a total fine of $8,000 for four lawyers, two of them were banned from practicing for two years before the same Mississippi court.

Since April 2023, around 1,600 similar cases, including content fabricated by AI models in legal documents, have been identified, mainly in the United States but also in 35 other countries, according to a non-exhaustive database, created by Damien Charlotin, French lawyer and researcher associated with HEC Paris.

- “Damage to reputation” -

Despite the publication of recommendations by numerous orders and associations of lawyers around the world in recent years, emphasizing in particular the need to verify the content generated, the number of hallucinations recorded in legal documents has been multiplied by eight over the last twelve months compared to to the previous year, according to data from Damien Charlotin.

AI is "a technology that likes to follow models", explains the French researcher to AFP. However, legal reasoning is largely based on repetitive structures (quotations, arguments, references) that these systems reproduce very well.

They can thus generate content which appears perfectly credible in form, faithfully imitating the codes of legal language, but which turns out to be misleading or erroneous in substance.

"There are cases where the lawyers still win", despite two or three hallucinations, "because they were fundamentally right", observes Mr. Charlotin.

But these cases lead to many