This is a published attorney-discipline order arising out of an immigration appeal. The underlying petitioners (identified as Malkeet Lnu, Sunita Rani Lnu, and Jaivin Lohan) sought review of a Board of Immigration Appeals decision dismissing their appeal from an immigration judge’s denial of asylum, withholding of removal, and protection under the Convention Against Torture. The merits are not the focus of the order; the discipline concerns the conduct of their two attorneys, Mike Singh Sethi and William Rounds of the Sethi Law Group in Orange County, California.
Sethi filed an opening brief containing two cases that do not exist — “Eduardo v. Garland” and “Lay v. Holder” — and twice attributed quotations to real opinions (Kamalthas v. INS (9th Cir. 2001) 251 F.3d 1279 and Avendano-Hernandez v. Lynch (9th Cir. 2015) 800 F.3d 1072) in which the quoted language never appeared. The reply brief and the attorneys’ briefs in other pending cases contained additional fabricated citations and gross misrepresentations of real authority. The court later learned that the firm used unlicensed law-school graduates (“Brief Writers”) to draft legal arguments and identify authority, that no licensed attorney read the cited cases, and that the errors were almost certainly the product of unauthorized generative-AI use, despite a purported firm policy against it.
This was an original disciplinary proceeding by the Ninth Circuit itself rather than an appeal from a lower court. After the court denied a joint motion to submit the case on the briefs, Sethi filed a “Motion to Correct,” recharacterizing the two nonexistent cases as “typographical errors” and substituting two real but unrelated cases (Udo v. Garland (9th Cir. 2022) 32 F.4th 1198 and Lai v. Holder (9th Cir. 2014) 764 F.3d 1098) that did not support the propositions asserted. At oral argument, Rounds (appearing in Sethi’s place) attributed the errors to “copy and paste” mistakes and three times categorically denied that generative AI was used — only later conceding under questioning that AI use was “possible.” The court then issued an Order to Show Cause why the attorneys should not be sanctioned, suspended, or disbarred. The attorneys’ Response ultimately conceded it was “more likely than not” that the errors resulted from unauthorized AI use by their Brief Writers.
The panel imposed discipline. It sanctioned Sethi and Rounds $2,500 each, suspended both from practice before the Ninth Circuit for six months, ordered them to notify clients, opposing counsel, presiding judges, and every attorney in their firm, required a perjury-backed AI-disclosure-and-verification statement in all future filings for two years, and directed the Clerk to forward the order to the State Bar of California.
In the published case of Lnu, et al. v. Blanche, Acting Attorney General, (U.S. Court of Appeals, Ninth Circuit, June 3, 2026) No. 24-4790, the 9th Circuit did not sanction the attorneys merely for using generative AI, which is not inherently unethical. (See Noland v. Land of the Free, L.P. (2025) 114 Cal.App.5th 426.) The rules, the court explained, are violated not at the point of research and drafting but at the point of signing and filing. Whether a fabricated citation comes from an AI tool or “his own natural intelligence,” the procedural and ethical rules apply with equal force.
Sethi’s signing and filing of briefs with nonexistent cases, misattributed quotations, and gross misrepresentations violated the requirement that contentions be supported by citations to authority (Fed. R. App. P. 28(a)(8)(A)), his duties of competence and diligence (Cal. R. Prof. Conduct 1.1, 1.3), and his duty to present only meritorious claims (Cal. R. Prof. Conduct 3.1(a)(2)). The court emphasized that a signature is an attestation that the signer has personally reviewed and is responsible for the filing’s accuracy, an obligation that cannot be delegated to subordinates or AI. It relied heavily on Noland, which held that attorneys must personally read and verify every citation, and noted that a fabricated opinion is not “existing law.” (Citing Mata v. Avianca, Inc. (S.D.N.Y. 2023) 678 F. Supp. 3d 443.) It rejected the excuses that other real cases might support the same propositions and that the citations went to minor points, observing that it is counsel’s job — not the court’s — to find supporting authority.
The more serious failing, and the reason for the suspension, was the repeated lack of candor (Cal. R. Prof. Conduct 3.3(a)(1)). The court found that no plausible typographical or copy-paste error could transform the real cases into the hallucinated citations, and inferred that the attorneys knew the “typographical error” explanation was false as early as the Motion to Correct. Rounds’s flat denials of AI use at oral argument, followed by his concession that AI use was “possible,” supported an inference of actual knowledge of falsity; the court stressed that candor requires admitting uncertainty, not just disclosing known facts. The court also faulted Sethi for a later “Notice of Errata” in another case that swapped out fabricated citations without disclosing that the originals were hallucinations. The duty of candor, the court held, requires an attorney who discovers a hallucination to promptly alert the court and opposing counsel, describe the nature of the error, and disclose its source. Had the attorneys done so candidly at the outset, lesser sanctions might have sufficed; the gravity of the discipline was owed to their repeated failure of candor.
In explaining the risks, the court drew on empirical research showing that even legal-specific AI tools from Westlaw and Lexis produced hallucinations — including subtle “inaccuracies” that misstate holdings rather than outright fabrications — in a significant share of queries, warning that inaccuracies may prove more dangerous because they are harder to detect. (Citing Magesh et al., Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools (2025) 22 J. Empirical Legal Stud. 216.)
