A federal judge in Mississippi recently handed the legal profession another reason to ask an increasingly uncomfortable question: How is this still happening?
AI hallucinations are no longer a mystery. Lawyers know — or certainly should know by now — that generative AI can produce confident, polished, and completely false answers. The warnings have been everywhere: court opinions, bar guidance, ethics alerts, CLE programs, law firm policies, and plenty of public embarrassment. Yet lawyers keep putting AI-invented cases into court filings.
What's new, in the cases piling up this spring, is that the courts have stopped treating this as a technology story and started telling us why it keeps happening. The answer is more damning than "they didn't understand the tool."
An AI scandal with an unusual twist
The latest example comes from Mississippi, where The New York Times reported that U.S. District Judge Sharion Aycock sanctioned all four lawyers on opposing sides of a civil case after attorneys on both sides cited fake legal authority in their filings. The consequences were severe. All four were removed from the case and fined a combined $8,000. Two were barred for two years from appearing in the U.S. District Court for the Northern District of Mississippi. And the judge canceled the civil trial.
"This case presents the court with an unusual scenario — attorneys for both litigants engaged in similar sanctionable conduct," Judge Aycock wrote. That detail turns the Mississippi case from just another AI cautionary tale into something closer to a profession-wide warning sign.
One of the disciplined lawyers, a solo practitioner, told the court she used First Drafts – an AI-powered legal drafting program — on a motion that included nonexistent cases. When confronted, she said she did not know AI could produce hallucinated cases and did not know what a hallucinated case even was. The judge found that explanation, in her words, insufficient and incredulous. So do I. In 2026, such claims of ignorance from a licensed attorney don't strain credulity so much as snap it. (I'd add that the lawyer might have taken the tool's name more literally. First Drafts makes a first draft – not a document ready for filing.)
But here is the part the "she didn't know" framing leaves out: this was not a lawyer caught off guard by an unfamiliar risk. She had already been disciplined weeks earlier, in April, by a federal bankruptcy court in Louisiana for the very same conduct. Judge Aycock had flagged her concerns about fictitious authorities back in December. The warnings didn't land. The filing went out anyway.
The other side fared no better, and arguably worse. Aycock reserved her sharpest language for an out-of-state partner whose own firm had an AI policy requiring lawyers to verify their research. The court found it "particularly egregious" that the partner — "a partner and presumably a leader in her law firm" — disregarded that policy and relied blindly on an AI research tool anyway.
That's the more important lesson here, and it's not just for law firms: A policy is not a practice, a training session is not a safeguard, and a warning label is not a verification system.
The mistake is in signing off
For the record, there's nothing wrong with lawyers experimenting with AI. Legal work is document-heavy, deadline-driven, and expensive; the appeal is obvious and the upside is real. But that's also where the profession keeps getting the diagnosis wrong — by treating these episodes as a problem of technology rather than a problem of lawyering.
The most clarifying statement of the year on this came not from Mississippi but from the Ninth Circuit, which suspended two California lawyers in June over fabricated cases and misquoted authorities in an immigration appeal. The court went out of its way to say what it was not punishing: not the use of AI, either by these lawyers or by their subordinates. The rules, the panel explained, aren't broken in the research or the drafting – they're broken "at the point of signing and filing." And whether a fake citation springs from an artificial intelligence or, as the court put it, from a lawyer's own natural intelligence, makes no difference.
The duty to read what you put your name on never changed. AI didn't create a new obligation; it just made it far easier to manufacture authority that looks real.
The reason this keeps happening isn't that lawyers don't understand hallucinations, but that the verification step – slow, tedious, and unglamorous – is precisely the work AI promises to make disappear. The tool's whole value proposition is in tension with the one duty it can't touch.
The Ninth Circuit case shows what that looks like up close. The firm assigned brief-writing – including the legal arguments and the supporting authority – to law school graduates who were not yet licensed attorneys, and no licensed lawyer read the cases they cited.
The managing attorney conceded that the firm doesn't normally check the citations its brief-writer supplies. The court called that an extraordinary confession, and it is. A signature on a brief has always meant "I have read this and I am responsible for it." Outsourcing that to an unlicensed assistant who outsourced it to a chatbot doesn't change what the signature certifies.
Not just a small-firm problem
If the comforting story is that this happens only to solo lawyers and under-resourced practices handed powerful tools without enough training, that story collapses at Sullivan & Cromwell.
In April, the elite Wall Street firm — 900-plus lawyers, offices on four continents — apologized to a federal bankruptcy judge in Manhattan after submitting a motion riddled with AI-generated errors, including fabricated case citations. As the NYT described it, a partner in the firm sent the judge a three-page ledger cataloguing roughly three dozen errors, some clerical, others involving imagined passages from real cases. The mistakes were caught not by the court, and not by the firm, but by opposing counsel.
Sullivan & Cromwell is not a startup in a strip mall. According to the firm, lawyers must complete a training course before they're given access to AI tools, and one of that training's instructions is to "trust nothing and verify everything." That is exactly the right rule. The uncomfortable lesson is that the right rule, at one of the most sophisticated firms on the planet, still didn't prevent the wrong filing.
There's an irony here too sharp to leave out: Sullivan & Cromwell is the firm that advises OpenAI on the safe and ethical deployment of artificial intelligence. Yet the counselor to the AI company couldn't keep hallucinations out of its own brief.
So the question for any firm using AI isn't whether it has a policy. It's whether the policy is actually followed before the filing goes out the door. Who used the tool? Who checked the result? Who read the cases? Those are no longer abstract governance questions. They're practical ones for every law office in the country.
The cover-up, not the crime
There's a second thread running through these cases, and it's the one that should worry the profession most: the hallucination is rarely what earns the harshest punishment. The dishonesty afterward is.
The Ninth Circuit was explicit about it. One of the suspended lawyers denied at oral argument, three separate times, that AI had been used — then conceded minutes later, under continued questioning, that it was "possible."
There was no upside to the denial, the court observed, and dressing a hallucination up as an innocent typo only made things worse. Had the lawyers simply admitted their AI use and apologized, the panel said, they might have drawn a lighter sanction. The suspension, in the end, was the price of a "repeated failure of candor." In other words, the cover-up cost more than the mistake.
The Fifth Circuit told the same story in February, sanctioning a Texas attorney whose reply brief contained 21 fabrications and misrepresentations — 16 invented quotations and five serious misstatements of law or fact. What pushed the court from annoyance to sanction was her shifting, evasive account of how the brief was prepared. And, like the Mississippi solo practitioner, she was a repeat actor: The prior year, the court noted, she had been caught doing the same thing and escaped sanctions only by filing a sworn declaration affirming her firm had taken steps to reinforce diligence and candor. Then she did it again.
That's the recurring shape of these cases. Warned. Promised to reform. Repeated. And then, when caught, less than forthright. It's why "I didn't know" rings so hollow by now: In case after case, they did know, and the failure was honesty – or, more accurately, the lack thereof.
A problem with no end in sight
The scale is no longer anecdotal. Damien Charlotin, a lawyer and data scientist, maintains a database of court decisions addressing AI-hallucinated content. When the Fifth Circuit referenced it in February, the database counted 239 such cases involving lawyers in the United States. Today it holds roughly 1,600 decisions worldwide – involving lawyers, pro se litigants, and judges alike – with more than 1,100 in the U.S. alone, and it keeps climbing, both as new filings surface and as older ones get added to the record.
Perhaps not surprisingly, self-represented litigants account for more of the entries than lawyers do. The real surprise is the company those pro se filers keep: at least some Wall Street partners are making the identical mistake.
And the usual escape hatch – I used a real legal tool, not ChatGPT – doesn't hold. The Ninth Circuit pointed to a 2024 study finding that even the purpose-built legal tools hallucinate at startling rates – roughly 17% of answers for Lexis and 33% for Westlaw. The Mississippi lawyers' "First Drafts" and "in-house research tool" defenses rest on an assumption a federal court has now rejected on the record.
Notice, too, who is actually catching these. Not the courts, as a rule, but adversaries and a handful of obsessive trackers who comb filings for fabricated cites. That ad hoc net works for the obvious fakes. It's far worse at catching what the Ninth Circuit warned may prove "more dangerous to our profession in the long run" — real citations attached to subtly wrong propositions, the kind that pass a glance and slip through unnoticed. The fake case is the easy problem. The plausible misquote is the one that will quietly corrupt the record.
The Fifth Circuit put the bottom line plainly. It found that, despite years of news coverage, CLE presentations, scholarly articles, and judicial warnings, the problem "shows no sign of abating" – it's getting worse, not better. And then the line that ought to retire the ignorance defense for good: If pleading ignorance of AI's risks was ever an excuse, "it is certainly no longer so."
The case that should have changed everything
The case that made all of this famous was Mata v. Avianca, a 2023 personal-injury suit in New York. Lawyers there submitted a filing citing made-up cases generated by ChatGPT — including a fabricated opinion, Varghese v. China Southern Airlines – complete with a real federal judge listed on the panel of a case that itself was entirely fictional. The judge fined two lawyers and their firm $5,000 jointly, and the episode became the profession's first major public warning that AI could conjure legal authority out of thin air.
That was three years ago. Back then, the profession could still claim it was learning the shape of a new risk. The opinions stacking up in 2026 are the sound of that excuse expiring.
Colorado has seen this, too
There are Colorado chapters of this story, too, and they're worth telling because they show how close to home the problem runs – and how pernicious it can be. As Chris Brown discussed in a Colorado AI News column nearly two years ago, Colorado attorney Zach Crabill was disciplined in 2024 after using ChatGPT in a court filing that included fake cases. To his credit, Crabill tried to turn the experience into something constructive — a push to use AI more responsibly and to help underserved clients.
That could be seen as a positive turn, but in the bigger picture, it's two years later and far too few attorneys seem to have gotten the memo. As recently as last month, in Garcia v. City of Monte Vista, a federal judge in Colorado flagged a fabricated quotation in a motion and announced an order to show cause, writing that he was "highly suspicious" the false language was "a product of the un-verified work of generative artificial intelligence."
Tellingly, the offending citation wasn't an invented case – it was a real Tenth Circuit decision quoted as saying something it never said. That's the subtler, more dangerous version of the problem: not a phantom case a careful reader can't locate, but a genuine case with words put in its mouth. The warning signs here aren't a 2024 artifact. They're current – and they're getting harder to spot.
A simple rule for a complicated time
Courts aren't asking lawyers to become AI engineers. They're not demanding that litigators understand model weights, training data, or retrieval-augmented generation. They are asking lawyers to do the job lawyers have always done: Read the case, confirm the quote, make sure the authority supports the proposition, and verify that the filing says what the law actually says.
That's what makes these cases so baffling – and why the courts' reframing matters. A hallucinated citation is not a typo but a representation to a court that legal authority exists for a proposition. When the authority is invented, the damage runs past embarrassment: It wastes judicial time, misleads opposing counsel, harms clients, and erodes the court's ability to trust lawyers as officers of the court.
AI did not create the duty of candor. It created a faster, more tempting way to violate it – tempting precisely because the one safeguard it can't automate is the one that feels like wasted time. An AI-generated case citation isn't authority. It's a lead, a starting point. That's it. Before that citation appears in a filing, a lawyer has to do the thing a signature has always promised: Read it, and confirm it's real.
That obligation didn't change in 2026. What changed is that the courts have run out of patience for pretending anyone could have missed it.
