The country's largest federal appeals court has fined and suspended two attorneys for filing immigration briefs containing citations to cases that did not exist and quotes that had been fabricated. The lawyers, Mike Sethi and William Rounds of Orange County, told the Ninth Circuit at first that the errors were typos. They later conceded the actual source was an unlicensed law school graduate at the firm using an unauthorised AI tool. The penalties: $2,500 each, a six-month suspension from practice before the Ninth Circuit, and an order to send the ruling to every client, opposing counsel and presiding judge in their other matters.
If that were the whole story, it would be one more entry in an already long ledger. Earlier this year, the Fifth and Tenth Circuits sanctioned attorneys for similar AI-related fabrications. What makes the Ninth Circuit's ruling different is that the three-judge panel used it as a doctrinal opportunity. Judges Richard Paez, Carlos Bea and Danielle Forrest wanted to lay out, for the first time, what the proper consequences look like when a brief contains hallucinations or inaccuracies rather than being fabricated wholesale.
The court was specific about what it was not doing. "We do not sanction Sethi and Rounds for the simple fact that they or their subordinates used generative AI," the panel wrote. The procedural rules do not regulate how lawyers produce filings, only what they put their name on. The duty of candour to the court is the same whether the brief was written by a partner, a paralegal, or a language model. What matters is that the attorney of record signed it, and so owns its accuracy.
The most useful passage in the ruling is the one that distinguishes types of error. Fabricated citations, where the case simply does not exist, are the obvious failure mode and the one that has generated most of the press. But the court was more worried about something quieter. "Inaccuracies may prove more dangerous to our profession in the long run," the panel wrote, "because they are likely to go unnoticed." A citation that resolves to a real case but mischaracterises its holding is harder to catch in a review, and easier for a busy court to take on trust. The opinion cited a 2024 study that found Westlaw's AI tool hallucinated answers 17 percent of the time, and Lexis's 33 percent of the time, against a representative set of legal queries. Those numbers are for purpose-built legal AI, not general-purpose chatbots.
The market context matters. Both Anthropic and OpenAI have pushed in recent months to add legal plug-ins to their flagship chatbots, and legal research vendors have been racing to ship AI features for years. The ruling does not slow any of that down. It does, however, sketch the legal duty that follows from using these tools, and it puts the obligation squarely on the signing attorney. Subcontracting accuracy to a model is not a defence. Subcontracting it to a junior, who then subcontracts it to a model without permission, is not a defence either.
For the rest of the profession, the practical signal is that the federal bar is starting to converge on a standard. Use the tools, but verify the citations. Read the cases. Be aware, as the panel put it, that "lawyers using generative AI must guard against" the technology's tendency to make mistakes that look plausible at a glance. That guidance reads obvious only because it is. The fact that it has now been written down by an appellate court is the part that will matter when the next disciplinary case arrives.