The cases that never existed.
A New York lawyer filed a brief citing half a dozen court decisions. None of them were real. Three years on, courts across the country are still working out what to do when a lawyer's research was written by a machine.
The brief that started it
In the spring of 2023, a lawyer in a routine personal-injury case in the Southern District of New York filed a brief that looked ordinary. It cited a string of federal decisions, with names, reporters, and quotations. There was only one problem, and it was a large one: several of the cases did not exist. They had been produced by ChatGPT, which invented the case names, the citations, and the quotes, and presented them with complete confidence.
The matter became Mata v. Avianca, and it turned into the most-discussed footnote in the legal profession's first year with generative AI. When opposing counsel could not find the cases and the court asked for copies, the fabrications unraveled. In June 2023, Judge P. Kevin Castel sanctioned the two lawyers and their firm five thousand dollars under Federal Rule of Civil Procedure 11, and ordered them to send the ruling to each real judge who had been falsely named as the author of a fake opinion. That last remedy, rare and pointed, is part of why the story traveled around the world.
Why it was never really about the AI
The lesson most people took from Mata was "do not use AI." The court's actual reasoning was narrower and more useful. Judge Castel was careful to say that using a new tool, and even unknowingly submitting fake cases, would not by itself have crossed the line into sanctionable conduct.
What drew the sanction was not the machine. It was filing research no one had checked, and then standing behind it after there was every reason to doubt it.
The lawyers had chances to catch the problem and withdraw the citations. Instead the fabricated cases were defended and even resubmitted. Rule 11 asks a filer to certify that the claims and citations in a paper are warranted, and that certification has always required actually reading the authority you cite. The tool was new. The obligation was not.
What courts have done since
Three years later, the isolated embarrassment has become a genre. A public database maintained by the legal researcher Damien Charlotin now tracks well over a thousand US court decisions in which a party relied on AI-fabricated material and a court had to respond, with the count still climbing through 2026. What began in trial courts has moved upward: federal appellate panels and state high courts have now weighed in.
The responses fall into two buckets. The first is disclosure. Many judges have issued standing orders that require lawyers to certify whether generative AI was used to prepare a filing and, if so, that a human checked every citation. The second is sanctions, and they have grown teeth. In 2026 the US Court of Appeals for the Ninth Circuit did not just fine two attorneys whose immigration brief cited nonexistent opinions and invented quotations; it suspended them from practice before the court and ordered them to notify their clients, the opposing side, and the presiding judge. Reported penalties elsewhere have climbed well past the original five-thousand-dollar figure, and suspensions are no longer unusual.
Notably, courts have kept drawing the same line Mata drew. The heaviest consequences have tended to follow not the mere use of AI, but the same aggravating pattern: a citation no one verified, followed by a lawyer who was not forthcoming once the problem surfaced.
The duty is old
Rule 11 has always required a lawyer to confirm that a cited case exists and says what the brief claims. Nothing about that changed.
Disclosure orders spread
Many courts now have standing orders requiring lawyers to disclose or certify any use of generative AI in a filing.
Sanctions are escalating
From a $5,000 fine in 2023 to fines plus suspensions in 2026, and from trial courts up to federal appellate panels.
It's the cover-up, not the tool
Courts have been clearest that the sanctionable failure is not checking, and then not owning it, more than using AI at all.
The rule underneath
For all the novelty, the doctrine here is old and settled. A lawyer signing a filing vouches for it. The duty of candor to the court forbids passing off a fabrication as authority. The duty of competence has always included knowing what your sources actually say. None of that was written with chatbots in mind, and none of it needed to be. What generative AI changed is not the standard but the ease of falling short of it: a confident, well-formatted paragraph of citations can now be produced in seconds by something that does not know, and does not care, whether any of it is true.
That is why the guidance emerging from bench and bar is strikingly consistent, and strikingly boring. Use the tools if they help. Then read every case before you cite it, the same way a careful lawyer always had to.
The technology in these stories is three years old. The obligation to stand behind what you put your name on is a great deal older.
The tool that writes the citation still cannot promise the case is real. That part is still the lawyer's job.
Archivar
The Archivar Editorial Desk
Researched and fact-checked against primary sources. General information for attorneys, not legal advice.
Sources: Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023); the US Court of Appeals for the Ninth Circuit's 2026 sanctions and suspension order in an immigration matter; the AI-hallucination case database maintained by legal researcher Damien Charlotin; and analysis by Norton Rose Fulbright and Bloomberg Law.
