Photo by Feyza Tuğba on <a href="https://www.pexels.com/photo/a-person-holding-a-book-12752026/" rel="nofollow">Pexels.com</a>
An Orange County immigration lawyer is now barred from the immigration courts, the Board of Immigration Appeals and DHS until mid-December. The cause was a Ninth Circuit brief citing cases that never existed. In Matter of Sethi, 30 I&N Dec. 112 (BIA 2026), the Board extended the Ninth Circuit’s six-month suspension across all of immigration practice.
The case is a warning about how easily AI errors reach a court and how hard they are to catch. By the lawyers’ own eventual admission, AI was the likely source of the fake cases. They got through because no lawyer checked them. The suspension came from what the lawyers said afterward.
What the briefs got wrong
In Lnu v. Blanche, an asylum appeal, Mike Singh Sethi’s briefs contained three kinds of error. The court caught them; opposing counsel did not.
| Citation | The problem |
|---|---|
| “Eduardo v. Garland, 28 F.4th 742 (9th Cir. 2022)” | Does not exist |
| “Lay v. Holder, 729 F.3d 962 (9th Cir. 2013)” | Does not exist |
| Kamalthas v. INS, 251 F.3d 1279 | Real case; quoted for words it never says |
| Avendano-Hernandez v. Lynch, 800 F.3d 1072 | Real case; quoted for words it never says |
| Bandari v. INS, 227 F.3d 1160 | Real case; quotation partly invented |
| Singh v. Gonzales, 491 F.3d 1090 | Real case about reopening deadlines, cited on credibility |
| Zahedi v. INS, 222 F.3d 1157 | Real case, described as a holding on affidavits it never discusses |
It was not a one-off. The court found five more invented cases in two other clients’ briefs. “Eduardo,” with the same reporter number, appeared in both Lnu and Contreras Pelayo v. Bondi.
The cause was the workflow. An unlicensed law graduate wrote the arguments and chose the authorities. Sethi signed without checking them; he told the court the firm did not “normally vet citations” from its brief writers.
The cover-up did the real damage
The court said plainly that an early admission might have earned lesser sanctions. Instead, the lawyers missed every exit.
- The “typo” fix. Sethi’s Motion to Correct called the fake cases typographical errors and swapped in look-alikes, Udo v. Garland and Lai v. Holder. Names, volumes, pages and even a year differed, and neither substitute supported the brief.
- The denials. At oral argument, colleague William Rounds said three times that AI “was not used.” Minutes later he conceded it was “possible.”
- The quiet errata. After being ordered to explain himself, Sethi filed a correction in another case that swapped fake citations for real ones without saying they were fabricated.
The Ninth Circuit found these statements knowingly or recklessly false. That finding, more than the bad citations, produced the suspension. The BIA then refused to revisit it, holding that the rules apply whatever tool produced the errors.
Why AI is a real risk in legal work
The record never names the tool used, but the errors fall into two types the court described. Fabrications, such as invented case names and reporter numbers, are typical of general-purpose chatbots. Inaccuracies, real cases cited for things they do not hold, are subtler and pass any check that only confirms a case exists. The Lnu briefs had both.
Tools built for lawyers do not remove the risk. A peer-reviewed Stanford and Yale study, cited by the Ninth Circuit, tested Lexis+ AI and Thomson Reuters’ Westlaw and Practical Law tools in 2024. Vendors had marketed them as reducing or eliminating hallucinations. They invented fewer cases than GPT-4, but still gave wrong or unsupported answers on 17 to 33 percent of queries, between one in six and one in three.
Their typical mistakes were misread holdings and misplaced authority, the kind found in the Lnu reply brief. The court warned that such errors may prove more dangerous to the profession than outright fakes, because they look right. Catching them takes the same work as research without AI: pulling and reading every case. In high-volume practices like immigration, the pressure to skip that step is where the danger lies.
There is an irony here. The Ninth Circuit’s order says Westlaw erred 17 percent of the time and Lexis 33 percent. The study’s own summary appears to say the reverse: Lexis+ AI performed best, and Westlaw hallucinated nearly twice as often. A ruling about citation accuracy may have swapped its own figures.
The takeaway
The Ninth Circuit said it was not punishing AI use as such; its rules apply at the signature. A signature attests that the lawyer personally read every authority cited. AI makes skipping that duty tempting and the resulting errors convincing. That combination is what cost Sethi six months of immigration practice, and the court now requires his firm to disclose any AI use under penalty of perjury for two years.
Three practical lessons follow:
- Read every case you cite. Confirming a case exists is not enough; three of the bad citations here were real.
- Supervise the drafters. Delegating research does not delegate responsibility, and “Eduardo” shows how one error spreads through templates.
- Disclose fast and fully. Name the error and its source. Every denial in Lnu made the penalty heavier.
For the clients, the errors cost nothing: the court granted their petition on other grounds. For their lawyer, the cost was his practice before every immigration forum for six months.
Sources
- Lnu v. Blanche, 177 F.4th 1014 (9th Cir. June 3, 2026)
- Matter of Sethi, 30 I&N Dec. 112 (BIA 2026), Interim Decision #4254
- Ninth Circuit order to show cause and oral argument excerpts, via Volokh Conspiracy (Nov. 10, 2025)
- Magesh et al., “Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools,” J. Empirical Legal Stud. (2025)
About the Author
Discover more from Immigration Analytics
Subscribe to get the latest posts sent to your email.
