Steven Schwartz thought ChatGPT was a faster version of Westlaw. In 2023, the New York attorney asked it to research a personal injury case, and it handed him six cases that sounded perfect: real docket numbers, real-sounding judges. Not one of them existed. That's an AI hallucination: a fabricated citation, quote, or holding that a tool generates with total, unearned confidence. Schwartz filed the brief anyway. When the court came looking for the cases, he asked ChatGPT if they were real. It said yes. That was also made up.
Three years later, AI hallucinations aren't a novelty story anymore. They're a documented, growing pattern with a paper trail. Attorney Damien Charlotin's public tracker of AI hallucination cases had logged more than 1,000 in U.S. courts alone by May 2026, and the count climbs every week.
Why AI Hallucinations Keep Landing Lawyers in Trouble
Schwartz paid $5,000 and got a stern lecture from the bench (Mata v. Avianca, S.D.N.Y. 2023). That looks almost gentle next to what's happened since.
In July 2025, a federal judge in Alabama sanctioned three attorneys at Butler Snow, a firm with more than 400 lawyers nationwide, after they cited hallucinated cases in a civil rights matter. The court disqualified all three from the case, referred them to the Alabama State Bar, and ordered them to send the sanctions order to their clients, every lawyer in their firm, opposing counsel, and every judge presiding over their other pending matters (Johnson v. Dunn, N.D. Ala. 2025). Two other attorneys copied on the filings were not sanctioned; the court found they hadn't drafted or reviewed the offending briefs.
An Oregon federal court went further. In December 2025, Magistrate Judge Mark D. Clarke fined two attorneys a combined $110,000 in fees and penalties in Couvrette v. Wisnovsky, a family dispute over an Oregon winery. Across three summary judgment briefs, the filings included 15 references to nonexistent cases and eight fabricated quotations. When the errors were flagged, one attorney tried to quietly strip them out and refile rather than disclose the problem, and the court found no evidence he had taken any steps to verify the legal arguments before or after. Clarke dismissed the case with prejudice and called the lawyers' conduct "not adequately forthcoming, candid or apologetic."
California delivered its own headline case that same year: a Los Angeles attorney fined $10,000, the largest of its kind in the state, after an appeals court found 21 of the 23 citations in his opening brief were fabricated by ChatGPT. He told the court he never reviewed the output before filing it.
Even the "Legal-Grade" Tools Aren't Clean
Here's the part that should worry firms more than any single sanctions case: switching to a legal-specific AI tool doesn't solve this.
Stanford's RegLab and Institute for Human-Centered AI tested the leading AI legal research tools and found real gaps. Lexis+ AI, the strongest performer, answered 65% of test questions accurately, and it still produced incorrect information on more than 17% of them. Westlaw's AI-Assisted Research hit 42% accuracy and hallucinated on roughly a third of the queries tested. Both are dramatically better than a general-purpose chatbot. Neither is safe to use unsupervised.
The finding is inconvenient for anyone hoping a better subscription fixes the problem. It doesn't. Verification isn't a workaround for a bad tool. It's the job, on every tool.
The Rules Already Require This
The American Bar Association didn't wait for a wave of sanctions to catch up. Formal Opinion 512, issued in July 2024, applies existing rules of professional conduct directly to generative AI. Nothing about it is optional:
- Competence (Rule 1.1): You need a real understanding of what your AI tools can and can't do, including where they fail.
- Candor to the tribunal (Rule 3.3): You can't submit a filing you haven't verified, AI-assisted or not.
- Supervision (Rules 5.1, 5.3): Partners are on the hook for associates' AI-assisted work, and firms need policies to prove they're checking it.
- Confidentiality (Rule 1.6): Client details don't belong in a consumer chatbot that might retain or train on them.
Most of the discipline cases above trace back to a lawyer skipping a step that was already required before AI existed: reading the case before you cite it.
We're not your ethics counsel, and none of this is legal advice; that conversation belongs with your bar counsel. What we can help with is the technology side of the equation: which tools are approved, how they're deployed, and how verification gets built into the workflow instead of left to memory.
The Policy That Worked, and the Lawyers Who Ignored It
The detail that gets lost in the headlines is this: Butler Snow had a written AI policy before any of this happened, one that required sign-off from a practice group leader before AI-generated content could go into a filing. The court in Johnson v. Dunn looked at that policy, along with the firm's response after the fact, and declined to sanction Butler Snow itself. The policy is exactly what kept a 400-plus-attorney firm off the sanctions order.
It didn't save the three attorneys who ignored it. Lunsford, Reeves, and Cranford filed the hallucinated citations anyway, and the court made clear that a fine and a reprimand weren't enough this time. "If fines and public embarrassment were effective deterrents, there would not be so many cases to cite," the court wrote, according to the EDRM analysis of the ruling.
The lesson for a smaller firm isn't "skip the policy, it doesn't matter." It's closer to the opposite: a policy on paper protects the firm's standing, but it does nothing for the individual lawyer who skips a step under deadline pressure. If verification only happens because someone remembers to do it, sooner or later someone won't.
Why This Matters for Law Firms
For a 10-to-150-person firm without BigLaw's compliance staff, the exposure is sharper, not softer:
- Malpractice and E&O exposure. A hallucinated citation in a filed brief is a live malpractice claim waiting for a client to notice.
- Bar discipline follows the individual attorney, not just the firm, and it becomes public record.
- AI use isn't limited to brief-writing. Document review, e-discovery, and first-pass drafting carry the same hallucination and confidentiality risk, often with less scrutiny than anything that lands in front of a judge.
- Deadline pressure makes it worse, not better. Small firms and solo practitioners handling high case volume show up disproportionately in the sanctions record.
We've covered this shift from a couple of angles already: how AI is changing the attacker's side of the equation, and what state-sponsored actors targeting Signal means for law firms specifically. AI hallucinations are a different kind of exposure, but the pattern repeats: the risk lives in the tools your staff already use every day, not in some hypothetical future threat.
Getting This Right Doesn't Require Banning AI
I've read enough of these sanctions orders now to see the pattern, and it's not what you'd expect. The lawyers getting sanctioned aren't unusually careless people. They're busy people who trusted a tool built to sound confident, not to be right, and skipped a step they'd never skip on a slower week. Fixing that isn't about swearing off AI. It's about putting real structure around how your firm uses it: which tools are approved, what gets verified and by whom, and how client data stays protected along the way.
That structure belongs in your IT environment, not in a partner's inbox as a one-time memo. Not sure how AI tools are actually being used across your practice right now? That's worth finding out before a court does it for you. Talk to us about an AI readiness assessment for your firm.