Ninth Circuit on AI Hallucinations

Some excerpts from today’s long opinion in LNU v. Blanche, decided by the Ninth Circuit by Judge Richard Paez, Carlos Bea, and Danielle Forrest:

Attorneys Mike Singh Sethi and William Rounds filed briefs in this Court with multiple nonexistent cases, misattributed quotations, and gross misrepresentations of real cases. Sethi and Rounds claimed that the errors were the product of innocent typographical mistakes. And they repeatedly denied the possibility that generative artificial intelligence (“AI”) might have produced the errors.

Having identified other cases in which Sethi or Rounds filed briefs that presented similar problems, we ordered them to show cause why they should not be sanctioned, suspended, or disbarred from practice before this Court. We have considered their response, and we now impose discipline….

We issue this disciplinary order, and explain our reasoning at some length, as a warning to the members of this Court’s bar: be aware of the risks of overreliance on generative AI, read everything cited in a court filing—whether drafted by generative AI or not—and disclose quickly and transparently generative AI hallucinations that are inadvertently included in court filings….

There’s a lot of factual detail in the opinion, but here are a few general observations from the court:

Two types of [generative AI] mistakes, or “hallucinations,” are most relevant: fabrications and inaccuracies. Fabrications are instances in which the generative AI tool provides cases or quotations that do not exist at all.

Inaccuracies are more subtle. The generative AI tool might cite to real authorities but provide an answer that is legally or factually inaccurate or not supported by the citation.

Fabrications are the most notorious hallucinations, but inaccuracies may prove more dangerous to our profession in the long run. Inaccuracies are more likely to go unnoticed by attorneys and judges because they are not always susceptible to facial checks. Rather, “[i]dentifying these misunderstandings often requires close analysis of cited sources.” With close analysis, some inaccuracies might be clear—for instance, claiming that a case expressly stands for a proposition about a certain topic when the case does not discuss that topic at all. Others may be difficult to distinguish from poor legal reasoning.

And inaccuracies are common, even in newer generation models that produce fewer fabrications. Including inaccuracies, legal-specific generative AI tools from Westlaw and Lexis hallucinated 17% and 33% of answers, respectively, to a representative set of queries run in 2024. [Citing Magesh et al., Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, 22 J. Empirical Legal Stud. 216 (2025).] {Popular legal AI tools have apparently adopted a definition of hallucination that only includes fabrications. We agree with Magesh et al. that this is “plainly irrational,” as such a definition “would require us to conclude that a tool that links only to Brown v. Board of Education on every query … has provided ‘hallucination-free’ citations.”}

The most common error modes of the latest generation tools include misunderstanding holdings, failing to distinguish between legal actors (e.g., presenting a rejected party argument as the holding of the court), and failing to respect the hierarchy of authorities. In other words, the sort of errors that we might expect a first-semester law student to make, but certainly not licensed attorneys appearing before this court.

As we will explain, filing briefs with hallucinated fabrications and inaccuracies violates procedural and ethical rules. Lawyers using generative AI must thus be aware of the tendency of generative AI to make these mistakes and guard against them….

However legal papers are prepared, and however legal technology develops, our procedural and ethical rules apply with equal force. Just as faithful adherence to those rules would prevent the submission of generative AI hallucinations, such adherence would also prevent the submission of similar human-generated errors….

[A] competent and diligent attorney must do more than prompt generative AI, check that the citations provided by the AI are real and the subject matter roughly on point, and call it a day. A competent and diligent attorney must also read and reason….

It is no excuse that Sethi entrusted substantive cite checking to subordinates, and it is no excuse that Sethi purportedly did not know his subordinates had used generative AI. It was Sethi’s signature, and his alone, on the briefs. So, it was Sethi who “present[ed]” the unwarranted contentions.

And Sethi’s signature was an attestation that he personally reviewed the contents of the brief, including the cited authorities, and that they were accurate. Sethi did no such review, and his attestation was not conditional on the tools that his subordinates might have used to prepare the first draft, nor could it be.

To the contrary, Sethi attests that he and Rounds “do not normally vet citations used by the Brief writer during our review.” That is an extraordinary confession. A competent and diligent attorney cannot decline to “vet” citations, in a brief he signs, for substantive validity—in other words, to read the cited authorities and ensure that they are on point.

We do not suggest that every minor typographical error in a citation gives rise to a violation of the ethical and procedural rules. The errors we identify, however, are not plausibly typographical….

We stress that when an attorney learns of any error in a filing—including generative AI hallucinations—he should immediately alert the court and opposing counsel of the error and disclose its source. There is no upside to denying the use of generative AI or to passing off an AI hallucination as an innocent typographical error. The other rules discussed in this order—competence, diligence, meritorious arguments, citations to authority, attestations to accuracy—do not turn on the source of the error.

If, in the Motion to Correct, Sethi and Rounds had disclosed that AI was used in the opening brief against firm policy and apologized for failing to check the brief, lesser sanctions may have been warranted. But that is not what they did. The gravity of discipline we impose, including the temporary suspension of practice, is owed to this repeated failure of candor….

And here are a few excerpts as to the factual details, though there’s a lot more on this in the opinion:

Sethi filed an opening brief in this case with multiple fabricated citations and quotations. Sethi cited two cases that do not exist and never existed: “Eduardo v. Garland, 28 F.4th 742 (9th Cir. 2022),” and “Lay v. Holder, 729 F.3d 962 (9th Cir. 2013).” And Sethi twice attributed quotations to real opinions in which the quoted language does not appear: Kamalthas v. INS, 251 F.3d 1279, 1284 (9th Cir. 2001), and Avendano-Hernandez v. Lynch, 800 F.3d 1072, 1080 (9th Cir. 2015). The Attorney General did not flag the fabricated citations in the answering brief.

After we denied the parties’ joint motion to submit Lnu on the briefs, Sethi filed a Motion to Correct the Record Re: Errata to Petitioner’s Opening Brief (“Motion to Correct”). The Motion to Correct represented that the two nonexistent cases—”Eduardo v. Garland” and “Lay v. Holder“—were “typographical errors.” Sethi sought to replace those cases with two cases that have similar names, different reporter numbers, and in the case of “Lay,” a different year: Udo v. Garland, 32 F.4th 1198 (9th Cir. 2022), and Lai v. Holder, 764 F.3d 1098 (9th Cir. 2014). He also sought to “correct” the holding for which “Lay” was cited. Sethi did not explain how such significant typographical errors might have occurred. Nor did he address the quotations misattributed to Kamalthas and Avendano-Hernandez.

Sethi did not appear for oral argument. Rounds appeared on behalf of Petitioners instead. At oral argument, we asked Rounds to explain the errors identified in the Motion to Correct. Rounds stated that the intended citations were “somewhat garbled” and reiterated the claim that Sethi had intended to cite real cases. Rounds claimed that the real cases “stand for the same proposition.” As for the source of the errors, Rounds asserted that he was “not sure” but that “it looks like it was a copy and paste error or something like that.”

We then asked Rounds whether the errors might have been the product of generative AI, to which Rounds said that AI “was not used.” Rounds explained that the quotation misattributed to Kamalthas appeared in a different case, but did not explain how the misattribution occurred. We again asked Rounds whether generative AI might have been used to supplement the briefs, to which Rounds again said: “No. AI was not used ….” We raised the other quotation misattributed to Avendano-Hernandez and again asked whether the error was the product of generative AI, to which Rounds again said “No.”

After further questioning on the source of the errors, Rounds finally conceded that it was “possible” that AI might have been used by the individual who drafted the briefs. He clarified that although Sethi’s name was on the briefs, Sethi did not draft the briefs, but only “reviewed” them. Rounds later revealed that the brief writer was not yet licensed to practice law, and that no licensed attorney read the cases cited by the unlicensed brief writer…

After oral argument, we ordered Sethi and Rounds to show cause (“Order to Show Cause”) why they should not be sanctioned, suspended, or disbarred from practice before this court for “conduct unbecoming a member of [this] court’s bar,” and for “violating applicable rules of professional conduct.” In the Order to Show Cause, we identified additional issues in the Lnu reply brief…. We also identified similar issues in briefs filed by Sethi in other cases pending in this Court….

The misconduct in this case did not end with the initial filing of the Lnu briefs. At every subsequent step—including the Motion to Correct, oral argument, the Response to our Order to Show Cause, and more recent filings in other cases—Sethi and Rounds have knowingly or recklessly made false statements to this Court….

Instead of transparency, and despite our Order to Show Cause, Sethi has apparently chosen to engage in more subtle subterfuge. In one of the matters we identified in our Order to Show Cause, Contreras Pelayo v. Bondi, No. 24-5168, Sethi filed a “Notice of Errata” that identified two hallucinated citations and requested to replace them with real citations. The motion did not represent the errors as “typographical.” But nowhere in that motion did Sethi disclose that the prior citations were hallucinations. Instead, Sethi merely said that the brief “contains errors in two of the case citations.” He then simply listed the fabricated citations and said that each was “an error” or “incorrect,” and then identified what the real citation “should be.”

This is not sufficient disclosure. By citing the hallucinations in the opening brief and signing the brief, Sethi previously attested to the accuracy, and thus the existence of, the hallucinated citations. Swapping a hallucination out for a real case does not correct the prior false claim that “this is a real case.” By failing to notify the Court that Sethi previously cited cases that do not exist, Sethi “fail[ed] to correct false statements of … law previously made to the tribunal.” …

The court imposed the following disciplinary measures:

1. Sethi and Rounds are hereby suspended from practice before this Court for a period of six months starting ten days after this Order is filed.
2. Sethi and Rounds are ordered to provide a copy of this Order to their clients, opposing counsel, and the presiding judge in every pending state or federal case in which they are counsel of record. They shall also provide a copy of this Order to every attorney in their law firm….
3. Sethi, Rounds, and all attorneys at the Firm are ordered to include in all future filings a statement, made under penalty of perjury, addressing whether generative AI was used, disclosing the name of the tool used, and certifying that the attorney signing the brief or other filing has personally reviewed the filing and that all citations and quotations therein refer to existing authority….
4. The Court orders the Clerk of Court to serve a copy of this Order on the State Bar of California and any other applicable licensing authorities for further proceedings as appropriate.