The Firm Said It Banned AI. Its Unlicensed Brief Writers Used It Anyway. The Lawyer’s Suspension Now Reaches Every Immigration Court, and His Petition Is Scheduled for the Supreme Court’s November 6 Conference.

1. What happened in Lnu v. Blanche, in one paragraph

In Lnu v. Blanche, a published order filed June 3, 2026, a Ninth Circuit panel of Judges Paez, Bea and Forrest disciplined two lawyers from an Orange County immigration practice. Their briefs cited cases that do not exist, attributed quotations to opinions that do not contain them, and described real cases as holding things they never discuss. Each lawyer was ordered to pay $2,500 within 21 days and was suspended from practice before the Ninth Circuit for six months. Both were also ordered to send the order to every client, opposing counsel and presiding judge in their pending cases and to every attorney at their firm, and to certify that they had done so. Only the managing attorney’s name appeared on the briefs; the other lawyer argued the appeal.

The Ninth Circuit’s Lnu v. Blanche order is a published disciplinary decision holding that a lawyer who signs a brief attests that he personally reviewed every cited authority, whoever or whatever drafted it. The court sanctioned the signature, not the AI tool, and tied the severity of the discipline to the lawyers’ repeated failure to disclose that generative AI was the likely source of the fabricated citations.

The facts read like a case study written for small firms. The firm used law school graduates who were not yet licensed to write the legal arguments in its appellate briefs. The managing attorney signed the briefs. By the firm’s own account, its attorneys “do not normally vet citations used by the Brief writer” during review. The court called that “an extraordinary confession.”

It is not an isolated problem. Damien Charlotin’s AI Hallucination Cases database listed 2,241 decisions worldwide when we checked it on October 29, 2026. Its country filter shows 1,514 from United States courts, and its party filter, which covers the whole database rather than only the US entries, attributes 886 to lawyers. The database tracks court decisions rather than every incident, it is still growing, and some entries rest on inferred rather than confirmed AI use. Our earlier piece on the 2026 state bar AI ethics landscape covers the opinions that sit behind these cases.

2. How a firm with a no AI policy filed invented cases

The firm filed invented cases because its workflow put research in the hands of unlicensed writers and put no licensed reader between their drafts and the court. According to the response the lawyers filed to the court’s order to show cause, attorneys prepared the statement of facts, and the brief writers were “left to write the legal arguments and identify supporting authority for those arguments.” The lawyers said the firm prohibited AI and required the writers to check their own citations.

The fabrications were not confined to one brief. In Lnu itself the opening brief cited two nonexistent Ninth Circuit cases, “Eduardo v. Garland” and “Lay v. Holder,” and misattributed two quotations. The reply brief misquoted a third real case and cited two more for propositions they do not address. In two other appeals the court found five more nonexistent cases, three in Anguiano Alvarado v. Bondi and two in Contreras Pelayo v. Bondi. The court said its list was not exhaustive.

One detail should worry any firm that reuses brief templates. The same invented citation, “Eduardo v. Garland, 28 F.4th 742,” appeared in two different clients’ briefs. Counted across the three appeals, the order identifies seven fabricated case citations, which are six distinct invented authorities because “Eduardo” recurred in two appeals. The order does not say how that repetition happened, but the operational risk is plain: an error saved in a shared bank of arguments can travel with every brief that draws on it.

Opposing counsel did not catch the problem either. The court noted that the Attorney General’s answering brief “did not flag the fabricated citations.” The panel found them on its own. That is the practical reality for a small firm: you cannot count on the other side, or on a clerk, to be your second reader.

The lawyers eventually conceded it was “more likely than not” or “probable” that the errors came from unauthorized AI use by the brief writers and a failure to cite check. They also told the court they had hired a licensed attorney to read every brief the writers prepare. That fix arrived after the damage, not before it.

3. Why the court says the violation happens at the signature

The Ninth Circuit held that the rules “are not violated at the point of research and drafting, but at the point of signing and filing.” It said it did not sanction the lawyers “for the simple fact that they or their subordinates used generative AI.” A fabricated case is a violation whether it came from a chatbot, a tired associate or a brief writer working from memory.

The legal hook is the signature requirement in Federal Rule of Appellate Procedure 32(d). The court relied on an advisory committee note to Ninth Circuit Rule 32-1, dated December 1, 2025, which says that “regardless of how the filing is prepared, the signature is an attestation that the signer has reviewed the filing and is responsible for the accuracy of its contents.” In the panel’s words, the attorney who submits a filing “must read the authorities cited therein to ensure that they are real, properly attributed, and accurately represented.”

Because the managing attorney practiced from California, the court applied the California Rules of Professional Conduct. It found violations of the duties of competence and diligence in Rules 1.1 and 1.3, the duty to present only meritorious claims in Rule 3.1(a)(2), and the citation requirement in FRAP 28(a)(8)(A). It also found violations of the duty of candor in Rule 3.3(a)(1), discussed in the next section.

The order draws a distinction that matters for how you check work. It separates “fabrications,” meaning cases or quotations that do not exist, from “inaccuracies,” meaning real cases cited for propositions they do not support. The court warned that inaccuracies “may prove more dangerous to our profession in the long run” because a quick existence check will not catch them.

The court backed that warning with data. Citing a 2025 study in the Journal of Empirical Legal Studies, it noted that legal specific AI research tools from Westlaw and Lexis “hallucinated 17% and 33% of answers, respectively” on a representative set of 2024 queries, counting inaccuracies. Those are historical results for particular tools on 2024 queries, not current error rates, and the court noted that performance may have improved since. A citator check that confirms a case exists covers only part of the risk. Someone still has to read the case.

4. Why the repeated candor failures drove the severity of the suspension

The repeated candor failures drove the severity of the discipline, and the court said so expressly: the gravity of the discipline, “including the temporary suspension of practice, is owed to this repeated failure of candor.” The filings themselves were already violations; the lack of candor aggravated the outcome. “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 record shows four separate candor failures. First, the managing attorney filed a motion to correct that called the two invented cases “typographical errors” and swapped in real cases with similar names. The court found no plausible way a typo turns Udo v. Garland, 32 F.4th 1198, into “Eduardo v. Garland, 28 F.4th 742,” or Lai v. Holder, 764 F.3d 1098, into “Lay v. Holder, 729 F.3d 962.”

Second, at oral argument the other lawyer said three times that AI “was not used,” then conceded minutes later that it was “possible.” The court wrote that “candor requires admissions of uncertainty just as much as it requires revealing facts known for certain.” Third, the lawyers’ written response continued to insist on their innocent belief at the time. Fourth, in another appeal the managing attorney filed a notice of errata that replaced two fabricated citations without saying they were fabricated.

That fourth point is the one most firms get wrong. The court held that “swapping a hallucination out for a real case does not correct the prior false claim that ‘this is a real case.'” The required response has three parts. You notify the court and opposing counsel immediately, you describe the nature of the error, and you disclose how it came about. State what is known and what remains uncertain, and do not claim AI caused an error unless the available facts support that conclusion.

The Fifth Circuit reached the same place about three and a half months earlier. In Fletcher v. Experian Information Solutions, No. 25-20086, decided February 18, 2026, Chief Judge Elrod’s panel identified 16 fabricated quotations and 5 serious misrepresentations of law or fact in a reply brief and sanctioned the signing attorney $2,500. The court wrote that “had Hersh accepted responsibility and been more forthcoming, it is likely that the court would have imposed lesser sanctions.”

The pattern across both circuits is consistent. The fabricated citations set the floor, and the response to being caught sets the ceiling. A small firm cannot control whether a hallucination ever slips through, but it fully controls what it says the day one is found.

5. How Matter of Sethi carried the suspension into every immigration forum

The Board of Immigration Appeals extended the Ninth Circuit suspension through reciprocal discipline, under regulations that make an immediate suspension mandatory on a qualifying petition and create a strong rebuttable presumption in favor of final reciprocal discipline. In Matter of Sethi, 30 I&N Dec. 112 (BIA 2026), decided October 1, 2026 and designated as precedent after an unpublished decision on September 15, the Board suspended the managing attorney from practice before the Board, the Immigration Courts and DHS for six months, effective June 13, 2026.

The timeline was fast. The Ninth Circuit suspension took effect June 13. Disciplinary counsel for EOIR and DHS filed a joint notice of intent to discipline and a petition for immediate suspension on June 15. The Board granted the immediate suspension on July 1, 2026, roughly four weeks after the circuit order.

The speed comes from the regulation, which separates two steps. First, under 8 C.F.R. § 1003.103(a)(4), once disciplinary counsel files a petition with a certified copy of a federal court suspension, the Board “shall forthwith enter an order immediately suspending the practitioner” pending a final decision. Second, in the summary proceeding that follows, § 1003.103(b)(2) provides that the certified order creates “a rebuttable presumption of the professional misconduct.” The presumption can be rebutted only by clear and convincing evidence of a due process failure, an infirmity of proof, or grave injustice.

The Board considered the three available grounds for rebuttal and found none established. It found no showing of a due process deprivation, and the lawyer’s principal arguments concerned infirmity of proof and grave injustice. He argued that no direct evidence showed he knew AI had been used. The Board answered that the Ninth Circuit had inferred that knowledge from the circumstances, and that his arguments were “effectively a collateral attack” it had no authority to entertain. He argued that reciprocal discipline for AI misconduct was an “unprecedented sanction” and an “outlier penalty.” The Board rejected that framing outright.

The Board’s key sentence is broader than immigration practice: “An attorney must conduct himself or herself in accordance with governing rules of professional conduct regardless of the tools he or she uses to prepare and present cases.” It also repeated its long standing rule that the financial hardship of a suspension does not amount to grave injustice.

The practical reach is the point for immigration practitioners. A suspension before DHS covers practice before the department itself, not just appearances in court. A sanction that began in one appellate court now also bars practice before EOIR and DHS, although by its terms it does not suspend him from other federal courts. The Board’s regulations also require a practitioner to notify EOIR disciplinary counsel within 30 days of any discipline under § 1003.103(c). Here that notice was given on July 2, 2026, according to the amended charging document.

Two limits are worth stating plainly. The precedent decision addresses the managing attorney only. It also says nothing new about AI as such. It applies ordinary reciprocal discipline rules to a federal court suspension that happened to involve AI. That is exactly why it matters: there is no special, softer track for AI mistakes.

6. What the Supreme Court petition does and does not change

The petition does not suspend or alter the Ninth Circuit order while it is pending. The managing attorney filed a petition for a writ of certiorari on September 1, 2026, and the Supreme Court docketed it as No. 26-296, Sethi v. United States Court of Appeals for the Ninth Circuit on September 8. The federal respondents filed a waiver of their right to respond on September 18, and the petition was distributed on October 14 for the Court’s conference of November 6, 2026.

A waiver of response is not a concession on the merits, and the Court can still call for one. Distribution for conference means the justices will consider whether to grant review, deny it, or ask the government for a response. The docket shows no stay. The petition is a request for review, not a grant of review or relief from the existing orders. The BIA said in Sethi that a stay or other relief from the Ninth Circuit order could warrant reconsidering the immediate suspension, but it rejected the grounds presented on the existing record, and its final order lets the lawyer petition for reinstatement.

For a small firm, the petition should not change any planning. Even if review were granted, the supervision and candor principles in the order track rules that every state already has. California’s own Court of Appeal said much the same in Noland v. Land of the Free, 114 Cal. App. 5th 426 (2025), which the Ninth Circuit quoted: no filing should contain any citation “that the attorney responsible for submitting the pleading has not personally read and verified.”

California has since written a version of that line into statute. SB 574, signed September 30, 2026 and effective January 1, 2027, amends Code of Civil Procedure Section 128.7 so that a California court filing may not contain a citation that “an attorney responsible for submitting” it has not personally verified. That wording is not identical to the Ninth Circuit’s description of the signer’s duty, so treat the two as separate rules. We covered the statute in detail in our SB 574 explainer.

7. What Rule 5.3 asks of a lawyer whose staff use AI

California Rule of Professional Conduct 5.3 requires a lawyer with managerial authority to “make reasonable efforts to ensure that the firm has in effect measures giving reasonable assurance” that nonlawyer assistants act compatibly with the lawyer’s professional obligations. A lawyer with direct supervisory authority must make reasonable efforts as to that person’s conduct. The rule took effect November 1, 2018, long before generative AI was in law offices, and it already covers the problem.

The Ninth Circuit flagged the issue in a footnote. It said that before the firm’s corrective actions, “the Firm’s system of delegation may have violated California’s responsibilities for supervising nonlawyers and its prohibition on the unauthorized practice of law,” citing Rules 5.3(a) and (b) and 5.5(a). Without reading the cases in the drafts, the court said, the lawyers could not have “merged” the brief writers’ work into their own. That is an express warning about supervision and possible unauthorized practice issues, not a separate adjudicated Rule 5.3 violation, and it is about supervision rather than AI.

Rule 5.3(c) adds personal responsibility in two situations. A lawyer is responsible for a nonlawyer’s conduct that would violate the rules if the lawyer orders it, or ratifies it “with knowledge of the relevant facts and of the specific conduct.” A lawyer with managerial or direct supervisory authority is also responsible if he knows of the conduct “at a time when its consequences can be avoided or mitigated but fails to take reasonable remedial action.” Discovering a hallucinated citation in a filed brief is exactly that moment.

The comment to the rule says a lawyer must give assistants “appropriate instruction and supervision concerning all ethical aspects of their employment,” and that the measures “should take account of the fact that they might not have legal training.” A law graduate may have substantial legal training but is still a nonlawyer for these supervisory purposes, and the duty to instruct applies in full. Instruction must cover which tools are approved, what may be entered into them, and what must be checked before work goes to an attorney.

The American Bar Association’s Formal Opinion 512, released July 29, 2024 as its first formal opinion on generative AI, frames the same duties nationally around competence, confidentiality, communication and fees. Your own state’s version of Rule 5.3 and its AI guidance control your practice. California’s rule is the one the Ninth Circuit applied, so treat it as a worked example rather than a universal text.

8. How the Ninth Circuit order compares with California, New York and the Fifth Circuit

The Ninth Circuit order is stricter on disclosure than any general rule a small firm is likely to face, because it was a sanction tailored to one firm. Item 4 of the order requires the two lawyers “and all attorneys at the Firm” to include in all future filings, for two years, a statement made under penalty of perjury addressing whether generative AI was used, naming the tool, and certifying that the signing attorney “has personally reviewed the filing and that all citations and quotations therein refer to existing authority.”

Other authorities take different approaches to disclosure, but they converge on verification. New York’s 22 NYCRR Part 161, effective June 1, 2026, says that attorneys “should not be required” to disclose AI use, while letting individual judges adopt a model rule that requires careful review and treats the signature as a certification. California’s SB 574 goes the other way: from January 1, 2027, an attorney who uses generative AI must disclose that use to the court for documents submitted. It does not require a “no AI used” statement on every filing, and it does not set filing procedure for federal courts.

If your firm files in more than one court, the safest design is to meet the strictest verification standard everywhere and handle disclosure court by court. Verification is a quality control that costs the same in every forum. Disclosure is a rule that varies. Check applicable statewide rules, local or part rules and judge specific orders, and resolve any apparent conflict rather than assuming a standing order overrides superior authority.

9. What a defensible AI drafting workflow looks like in a small firm

A defensible workflow puts the submitting attorney’s own reading of every authority between any draft and any signature, and keeps evidence that the reading happened. Nothing in the Ninth Circuit order requires a specific tool or prohibits delegation. It requires that whoever signs actually reviews. The controls below are our recommended baseline, built from what the courts said went wrong. They are not a safe harbor, and no court has endorsed a specific checklist.

The cite check log deserves emphasis. In Sethi, the lawyer argued there was no draft history, no prompt and no metadata tracing the brief to AI. The Board was unmoved because the Ninth Circuit could infer knowledge from the circumstances. The lesson cuts both ways. A firm cannot defend itself by the absence of records, but it can defend itself with records showing a real review took place.

Confidentiality belongs in the same policy. SB 574 bars a California attorney from entering confidential or personal identifying information into a generative AI system unless access is restricted to the attorney and persons the attorney has authorized who are under obligations to protect its confidentiality, and the ordinary duty of confidentiality applies to client data typed into a consumer chatbot. Brief writers working from a client file need to know which tools may see that file.

10. A 30 day plan for a small firm

A solo or small firm can close the gaps the Ninth Circuit identified in about 30 days without buying new software. The plan below is a practical baseline we recommend, not a statutory requirement or a guarantee against sanctions.

Days 1 to 5: map who drafts what. List every person who drafts legal argument or selects authority, including law clerks, contract writers, paralegals and unlicensed graduates. Record which AI tools each one actually uses, including features built into research platforms and word processors. A policy written without this map will miss the real workflow.

Days 6 to 12: write the policy and the escalation rule. Name the approved tools and the data that may go into them. Require a note on any draft AI touched. Write down, in one paragraph, what happens the moment anyone finds a bad citation in a filed paper: notify the court and opposing counsel, describe the error, disclose its source, and say what is still uncertain. The Ninth Circuit’s guidance is specific enough to copy almost word for word.

Days 13 to 20: install the reader. Make the submitting attorney’s personal reading and verification of cited authorities a precondition to filing, and start a cite check log. Use another reviewer as an additional control, not as a substitute for that responsibility; the Ninth Circuit said it is “no excuse” that cite checking was entrusted to subordinates. Under SB 574 in California courts from January 1, 2027, verification must be personal to an attorney responsible for the filing.

Days 21 to 25: sweep open matters. Recheck the authorities in every pending brief prepared under the old process, starting with any argument reused from a template. If you find a fabrication, correct it with full disclosure. The Contreras Pelayo errata shows that a silent swap can be treated as a new candor violation.

Days 26 to 30: train and test. Walk every drafter through the difference between fabrications and inaccuracies, using the Lnu examples. Then test the process by pulling one recent brief and checking whether the log, the sources and the AI use note are all there. Repeat that spot check monthly. If you want an outside view of how your firm’s AI governance compares with frameworks such as ISO/IEC 42001, that is the work Dynamic Comply does.

This article is provided for general informational purposes and reflects the state of the law as of October 2026. It is not legal advice. Regulations in this area are changing quickly. Confirm current requirements and consult qualified counsel before making decisions for your organization.