The federal courts' running experiment with AI accountability reached a new phase on June 3, 2026, when the United States Court of Appeals for the Ninth Circuit issued its first precedential published opinion on AI-generated errors in legal filings. The ruling in LNU v. Blanche (No. 24-4790) imposes six-month suspensions and $2,500 fines on each of two Orange County immigration attorneys, refers both to the California State Bar, and — most significantly for the profession as a whole — warns that AI inaccuracies may prove more corrosive than outright fabrications, precisely because they are harder to catch.
What Happened in the Filings
Attorneys Mike Singh Sethi and William Rounds submitted immigration appeal briefs containing four AI-generated errors: two entirely nonexistent cases — Eduardo v. Garland, 28 F.4th 742 (9th Cir. 2022) and Lay v. Holder, 729 F.3d 962 (9th Cir. 2013) — and two real decisions, Kamalthas v. INS and Avendano-Hernandez v. Lynch, to which fabricated quotations had been attributed. The combination matters. Ghost citations are relatively straightforward to detect: any attorney opposing a brief can run the citation through Westlaw and find nothing. Falsely attributed language embedded in a real opinion requires the reader to pull the case and compare the cited passage to the original text — work that is easy to skip under deadline pressure and that has no automated backstop.
When the problem surfaced, the attorneys filed a motion to correct — but characterized their errors as "typographical mistakes" and "copy-paste errors" without disclosing that generative AI had produced them. That characterization transformed a correctable filing error into an independent violation. The panel held that the attorneys' duty ran not just to the accuracy of their submissions but to the truthful identification of the source of any error they discovered and disclosed.
The Holding: Liability Attaches at Signature
The Ninth Circuit's core rule is direct: "the rules are not violated at the point of research and drafting, but at the point of signing and filing." This reframes how the profession should think about AI tools in legal work. An attorney who delegates drafting to an AI and then signs the result without verification is not shielded by the delegation — the act of signing is the certification under Federal Rule of Civil Procedure Rule 11 that the submission is grounded in fact and law. That duty cannot be assigned to a language model.
The panel also established what it called "a continuing duty to correct and to disclose the source of any error." Once an AI-generated error is discovered, the attorney must disclose it as such — not characterize it as a clerical mistake. This is not novel doctrine invented for AI; it is the application of longstanding candor obligations to a new failure mode.
Why "Inaccuracy" Is the Harder Problem
Mata v. Avianca (678 F. Supp. 3d 443, S.D.N.Y. 2023) established the canonical baseline: attorneys who filed six completely fabricated ChatGPT citations received a $5,000 fine after Judge P. Kevin Castel found they acted with "subjective bad faith." The ghost-citation problem, since Mata, has been the profession's primary frame for AI hallucinations in courts.
LNU v. Blanche shifts that frame. The Ninth Circuit panel explicitly observed that AI-generated inaccuracies — real cases with wrong holdings or fabricated quotations — may prove more dangerous to the profession in the long run precisely because they escape notice. A litigant who catches a citation to the nonexistent Eduardo v. Garland immediately knows something is wrong. A litigant who reads a real Ninth Circuit opinion whose quoted language has been subtly altered by an AI tool may not catch the falsification at all. The error propagates into the briefing record, potentially into an oral argument, potentially into a ruling.
As AI tools grow more sophisticated, the failure mode migrates from obvious fabrication toward plausible distortion. The court's warning anticipates that trajectory and signals that verification of quoted substance — not just citation lookup — is now a professional obligation.
A Calibrated Sanction, With Long Teeth
The pro-regulatory case here deserves a fair hearing. Immigration proceedings determine whether someone can remain in the country. The attorneys' clients had no way to know their representation included AI-generated material that no one had reviewed. Courts allocate finite time; fabricated or distorted authority wastes that resource and disadvantages opposing parties whose research turned up accurate law. The judiciary's interest in accurate briefing is not bureaucratic formalism.
That said, the LNU v. Blanche sanctions are proportionate rather than prohibitive. Six months' suspension from Ninth Circuit practice — not disbarment. A $2,500 fine per attorney — not career-ending. No order prohibiting AI tool use. The court explicitly stated that AI assistance in legal practice is not itself sanctionable. What is sanctioned is the failure to verify and the failure to disclose.
The most consequential element of the ruling is prospective: all future filings from the Sethi Law Group must include a certification, under penalty of perjury, identifying whether generative AI was used, naming the specific tool, and confirming that the signing attorney personally reviewed every citation and quotation. That requirement converts a general verification duty into a signed, specific accountability checkpoint on every filing going forward.
The ABA's Formal Opinion 512, issued in July 2024, called for exactly this kind of structural check — requiring attorneys to independently verify AI output for accuracy before presenting it to clients or courts, with no efficiency gain excusing the obligation. LNU v. Blanche is now the appellate precedent that gives that ethical floor binding judicial force throughout the Ninth Circuit.
What Firms Need to Change
The practical implication is structural, not aspirational. Verification of AI-generated citations must be a mandatory workflow step — not a best-efforts practice — before any filing is signed. Boilerplate disclaimers that "AI was used but reviewed by counsel" will not withstand scrutiny if the filing contains errors that genuine review would have caught. The obligation the Ninth Circuit has now codified is substantive: an attorney must read the case, compare the quoted language, and confirm the holding — and must be able to certify under oath that this happened.
Generative AI will reshape legal research and drafting. LNU v. Blanche does not contest that. What it establishes, for the first time in precedential appellate law, is that the accountability architecture surrounding AI use must be as robust as the tools themselves — and that concealment of AI-generated errors is an independent basis for sanction, separate from and in addition to the underlying filing errors.