"The rules are violated at signing, not research": the Ninth Circuit's precedential AI-hallucination order

July 9, 2026 · Every factual claim in this piece links to its source.

By mid-2026 an attorney sanctioned for AI-fabricated citations is barely news. What makes Lnu v. Blanche, No. 24-4790 (9th Cir. June 3, 2026), worth reading is the header on page one: FOR PUBLICATION. The Ninth Circuit didn't issue a one-off show-cause slap. It wrote a precedential order — the panel called it a deliberate "warning to the members of this Court's bar" — that lays out exactly which rule a hallucinated brief breaks and when it breaks it. The reasoning is the most useful thing to come out of these cases so far, and it happens to draw the exact line that separates what a citation checker can do from what it can't.

What happened

Two Orange County immigration attorneys, Mike Singh Sethi and William Rounds of the Sethi Law Group, filed briefs whose opening brief cited two cases that never existed — "Eduardo v. Garland, 28 F.4th 742" and "Lay v. Holder, 729 F.3d 962" — and twice attributed quotations to real opinions (Kamalthas v. INS and Avendano-Hernandez v. Lynch) in which the quoted language does not appear. When the panel flagged it, the attorneys called the fake cases "typographical errors" and, at oral argument, denied three times that AI had been used — before conceding, minutes later, that it was "possible." It later emerged that the briefs were written by an unlicensed brief-writer whose citations no licensed attorney at the firm read before filing.

The discipline: each attorney personally sanctioned $2,500, suspended from practice before the Ninth Circuit for six months, ordered to serve the order on every client, opposing counsel, and presiding judge in their pending cases, and — for two years — required to file a sworn statement in every filing disclosing whether AI was used, naming the tool, and certifying the signer personally verified every citation. The order was referred to the State Bar of California. Details per the slip order and Metropolitan News-Enterprise.

The holding: the rule breaks at the signature, not the search bar

The court was emphatic that it was not sanctioning anyone for using generative AI. "The rules are not violated at the point of research and drafting," it wrote, "but at the point of signing and filing." A signature under FRAP 32(d), the panel explained, "is an attestation that the signer has reviewed the filing and is responsible for the accuracy of its contents" — and that duty doesn't turn on how the draft was produced, "whether by delegation, partnership, solo work, or generative AI." It quoted the California Court of Appeal's line from Noland v. Land of the Free, L.P.: no filing "should contain any citations — whether provided by generative AI or any other source — that the attorney responsible for submitting the pleading has not personally read and verified."

That framing matters for practice because it collapses a distraction. Whether the fake came from ChatGPT or a careless paralegal or a copy-paste error is beside the point; the obligation is a pre-filing verification pass on every cite you put your name to. The court made candor the aggravator: had the attorneys disclosed the AI use in their motion to correct and apologized, "lesser sanctions may have been warranted." The suspension was the price of the cover-up, not the hallucination.

The distinction worth building your workflow around: fabrications vs. inaccuracies

The most quotable line is about signatures. The most useful passage is where the court, citing the Stanford/Magesh empirical study (Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, 22 J. Empirical Legal Stud. 216 (2025)), splits AI errors into two kinds:

Fabrications — "cases or quotations that do not exist at all." Notorious, embarrassing, and — critically — facially checkable. You can ask a database whether the case is real.
Inaccuracies — a real citation offered for a proposition the case doesn't support, or a real case grossly misrepresented. The court warned these "may prove more dangerous to our profession in the long run" precisely because they "are not always susceptible to facial checks. Rather, identifying these misunderstandings often requires close analysis of cited sources."

And the base rates aren't small. The panel cited the same study's finding that legal-specific AI tools from Westlaw and Lexis hallucinated 17% and 33% of answers respectively on a representative set of 2024 queries — figures the court flagged include inaccuracies, not just fabrications. The lesson the opinion draws is the one every verification workflow should internalize: the flashy failure (a made-up case) is the easy one to catch; the quiet failure (a real case cited for something it never held) is the one that slips through and does the damage.

Which of these can an automated check actually catch? Honest answer: the first kind

This is where an existence-and-name check earns its place — and where it has to be honest about its limits.

Fabricated cases (the "Eduardo" and "Lay" problem) — yes. LegalCite checks every citation against public court records (CourtListener). A case that never existed in a public reporter fails the lookup outright. This is the failure mode that a machine catches faster and more reliably than a tired human at 11 p.m.
Citations that resolve to the wrong case — yes. When a reporter cite is real but the name attached to it belongs to a different case, a name-to-citation match flags the mismatch — the error that survives a quick "does something pull up?" glance because something does.
Misattributed quotations and misrepresented holdings (the Kamalthas, Avendano-Hernandez, and Singh problems) — no. These are the court's "inaccuracies." The case is real; the quote or the holding is not in it. Catching them requires a human reading the opinion. No existence check honestly claims otherwise — and the Ninth Circuit's own point is that these are the more dangerous category. The value of automating the first two failure modes is that it frees your reading hours for this third one, which is the only one that ever really needed them.

One nuance the opinion underscores: the court noted that popular legal-AI tools define "hallucination" to mean fabrications only, calling that definition "plainly irrational" because it would let a tool that answers every query with Brown v. Board claim to be "hallucination-free." An honest checker shouldn't play that game either. Flagging that a case exists is not the same as vouching that it says what the brief claims — and a tool that blurs the two is selling the same false comfort the court just warned about.

Before you sign and file — run the brief through LegalCite. It checks every citation against public court records, flags cases that don't exist, flags citations that don't match their case names, and reports Westlaw/Lexis database cites as unverifiable rather than guessing. It won't read the holdings for you — nothing honest can — but it clears the fabrications so you can spend your time on the cites that need a human. Free, no signup, and the document never leaves your browser: run it through Brief Check.

LegalCite verifies that citations exist in public court records (CourtListener) and match the case names attributed to them. It does not assess good-law status, verify quotations, or confirm that a case supports the proposition it's cited for — the "inaccuracies" the Ninth Circuit describes require human review — and nothing here is legal advice. Citations to unpublished databases (Westlaw/Lexis) are reported as unverifiable, not fabricated. Sources: Lnu v. Blanche, No. 24-4790 slip order (9th Cir., June 3, 2026, For Publication); Reason / Volokh Conspiracy, June 3, 2026; Metropolitan News-Enterprise, June 4, 2026; Magesh et al., Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, 22 J. Empirical Legal Stud. 216 (2025), as cited in the order. All URLs verified live July 9, 2026.