New cases of AI-generated errors in court filings fell in the second and third quarters of 2026, according to legal scholar Damien Charlotin, whose database has tracked at least 1,500 such cases in US courts. Charlotin said in a Friday blog post that the decline since spring may reflect newer models hallucinating less, though he cautioned that under-reporting could also play a part. US judges are still sounding alarms, with one Arizona judge writing that for every good AI-assisted filing there are ten obstructive ones. The pattern suggests improvement without resolution.
What The Data Shows
Filings containing made-up, misquoted, misleading or incorrectly cited legal material surged after AI tools reached mass use in 2022, Reuters reported. Charlotin’s database, widely cited in the field, has tracked at least 1,500 cases involving AI-related errors in US state and federal courts. Until this spring, that number had been rising fast. The second- and third-quarter decline breaks that run.
Charlotin, a senior researcher at the HEC Paris business school, offered two explanations. One is progress on the model front: newer large language models hallucinate less, according to a recent experiment in which models were asked to create legal documents. The other is quieter: errors may be just as frequent but reported less often when they look minor or unsurprising.
The database counts only cases that surface publicly, which is why Charlotin treats the decline with care. A filing error that a judge never flags, or that parties settle quietly, never enters the count. That gap means the second- and third-quarter fall is best read as a change in detected cases, not proof that models stopped making mistakes. Reuters noted that lawyers are generally not barred from using AI, but ethics and court rules bind them to verify accuracy, and people without lawyers face the same obligation for their papers.
Judges Are Still Warning
US judges continue to warn about AI errors, especially in filings from people without lawyers, Reuters reported. US District Judge John Tuchi of Arizona dismissed a civil rights lawsuit on Thursday, writing that the filing from an unrepresented plaintiff lacked factual specificity and appeared to be produced using AI. The judge allowed an amended filing and did not bar AI use in the case.
“It is this court’s experience that for every good AI-assisted filing there are ten more obstructive ones,” Tuchi wrote. “And the obstructive ones are often like the head of Hydra, for each one dismissed, several more follow.” He complained that AI-assisted submissions had gotten out of control, writing of “the overwhelming volume of filings, and sheer verbiage therein, which to date have pummeled courts.”
Charlotin told Reuters on Friday he expects AI errors to keep decreasing but not disappear in the short or medium term. “Tools get better, true, but not everyone is up to date on the latest tools,” he said. He warned of a second risk: as tools succeed in 99 percent of cases, lawyers may stop checking, “until it hallucinates again.”
Why It Matters
The slowdown matters because courts run on verified citations, and a single invented case can derail a filing. If newer models really do hallucinate less, the burden on judges eases; if the drop is only quieter reporting, the risk is unchanged and harder to see. Either way, the rules have not moved: lawyers must verify every submission, and unrepresented filers carry the same duty for their papers. The Tuchi order shows judges are policing that line case by case while the error count falls.
Background
AI-generated errors in filings have embarrassed lawyers and courts since chatbots went mainstream in 2022, with sanctions and warnings following a string of invented citations. Reuters has tracked the trend through related cases, including a Florida lawyer referred for possible discipline over a filing judges called “AI slop” and an Oklahoma judge whose ruling contained false citations. The new data is the first sign the tide may be turning, though the problem is far from settled.
Source: Reuters
0 Comments