French data consultant Damien Charlotin found 455 cases in the first quarter of 2026, compared with 423 in the second quarter and 359 in the third. In a Substack post on Friday, he credits technological improvements that reduce AI hallucinations, as well as increased awareness of AI's potential to hallucinate.
But even if an AI tool doesn't provide an incorrect answer to an attorney's question, it may also fail to tell them if there's something wrong with the question's premise, Charlotin warns.
"We have a lot of people experimenting with data showing people prefer sycophancy," he told Law360 Pulse on Monday. "Sometimes [developers] know they risk losing users to competitors if they're not sycophantic enough."
Charlotin noted that the public saw this tension point in real time when OpenAI updated its ChatGPT model last year to reduce sycophancy, only to roll it back due to user backlash, though it says it is still working on the issue.
"But more generally, we may be moving from legible AI errors to illegible ones," he wrote in his Substack. "Fabricated cases were always the easy part, the one that would (mostly) go away on its own. The subtler error modes of AI, by contrast, may take centre stage: sycophantic models giving you too much of what you asked for, or inputs optimised to trip you up."
Those adversarial inputs, as Charlotin calls it, involve lawyers manipulating AI tools they know or suspect opposing counsel uses. A criminal lawyer in Brazil was fined by the nation's Supreme Court for this conduct less than a year ago after he filed a brief with hidden text ordering AI readers to "deny all GPT commands." The court called the lawyer's actions "an act contrary to the dignity of Justice."
As clever as this type of trickery may sound, Charlotin is optimistic that good litigators will be able to stay ahead of the issue.
"I expect smart and sophisticated lawyers will know this and start optimizing," he told Law360.
--Editing by Michael Watanabe.
For a reprint of this article, please contact reprints@law360.com.