What did the court actually hold in Mata v. Avianca, Inc., and what was the sanction?
In Mata v. Avianca, Inc., No. 22-cv-1461 (PKC) (S.D.N.Y. 22 June 2023), Judge P. Kevin Castel held that attorneys Steven A. Schwartz and Peter LoDuca had violated Federal Rule of Civil Procedure 11(b)(2) by submitting an affirmation citing and quoting judicial opinions that did not exist, and that they had acted in subjective bad faith. The opinion states that the “Respondents abandoned their responsibilities when they submitted non-existent judicial opinions with fake quotes and citations.”
The sanction was a $5,000 penalty payable to the court registry, imposed jointly and severally on Mr. Schwartz, Mr. LoDuca and their firm, Levidow, Levidow & Oberman, P.C., within fourteen days. The court also ordered letters to the plaintiff, Roberto Mata, and to each of the six real judges falsely identified as the author of a fabricated opinion, attaching the opinion, the hearing transcript and the fake decisions.
Two things the court declined to do are quoted less often and matter as much. It refused to order an apology, reasoning that a compelled apology is not a sincere apology, and it did not refer the matter to disciplinary authorities in that opinion. The folklore has the case ending careers. The order ended in a fine and a stack of letters.
What categories of conduct have courts sanctioned since Mata v. Avianca, Inc.?
Four recur across the reported decisions, and only the first is what most people picture. The first is citation to authority that does not exist: in Wadsworth v. Walmart Inc. (D. Wyo. 24 February 2025), Judge Kelly H. Rankin sanctioned three lawyers over motions in limine citing eight non-existent cases, revoking Rudwin Ayala’s pro hac vice admission and fining him $3,000 with $1,000 each for T. Michael Morgan and Taly Goody. The ground is the one Mata v. Avianca, Inc. had already stated: a fake opinion is not “existing law,” and “citation to a fake opinion does not provide a non-frivolous ground for extending, modifying, or reversing existing law, or for establishing new law.”
The second is real authority cited for something it does not say. In a second Coomer v. Lindell order, reported on 9 May 2026, Judge Nina Y. Wang imposed a $5,000 Rule 11 sanction on Christopher I. Kachouroff and his firm where Capital Solutions, LLC v. Konica Minolta Business Solutions U.S.A., Inc., a district court decision reported at 695 F. Supp. 2d 1149, had been presented as binding Tenth Circuit precedent and cited for a Reexamination Clause proposition the decision had expressly disclaimed. That is the category most firms are not looking for, because a citation checker confirms the case is real.
The third is failure to verify by a lawyer who did not generate the text. In Johnson v. Dunn, 792 F. Supp. 3d 1241 (N.D. Ala. 23 July 2025), two of the three sanctioned attorneys did not know the citations had come from ChatGPT; the court found their disregard for the truth of motions bearing their names in the signature block was recklessness tantamount to bad faith. The fourth is the denial or concealment of AI use, discussed below.
Does it matter how a lawyer responds after fabricated citations are discovered?
In several published decisions it has mattered more than the citations themselves. Judge Castel’s opinion in Mata v. Avianca, Inc. turned on this: after Avianca’s brief of 15 March 2023 questioned whether the cases existed, and after court orders in April demanded that the decisions be produced, the respondents “continued to stand by the fake opinions,” and the opinion records that they “doubled down and did not begin to dribble out the truth until May 25.”
The same pattern shows up later. In Tercero v. Sacramento Logistics, LLC (E.D. Cal. 9 September 2025) the attorney denied using AI, was fined $1,500 and was ordered to serve the order on her client, with the clerk directed to serve it on the State Bar of California. In Coomer v. Lindell (D. Colo. 7 July 2025), Judge Wang rejected the explanation that a prior draft had been filed by human error, pointing to contradictory statements, the absence of corroborating evidence and similar conduct in another case, and imposed $3,000 each on Christopher I. Kachouroff and Jennifer T. DeMaster under Rule 11.
The Nebraska matter is the clearest version, and it has to be stated as reporting rather than as a read of an order. No primary Nebraska Supreme Court order was located for this article; what follows is what WOWT in Omaha reported on 8 and 15 April 2026. Omaha attorney Greg Lake was asked at a February 2026 hearing before the Nebraska Supreme Court whether he had used artificial intelligence and answered that he had not. The court struck his appellate brief in March 2026 over 57 defective citations out of 63, and found that his assertion that he had simply filed the wrong draft “lacks credibility.” Two days before he was temporarily suspended, he filed an affidavit admitting the AI use and calling it a grave error of judgment. The chief justice ordered him temporarily suspended from practice on 15 April 2026 until further notice from the court, with a referee to be appointed and the length of any suspension to follow a full disciplinary hearing. That pattern is not a rule in any jurisdiction, but it is the most consistent feature of the reported record. The incident response area covers the first hour.
What do standing orders on AI use in filings actually require?
They are not uniform, and the obligation usually lives with the individual judge rather than the district. The first was Judge Brantley Starr’s “Mandatory Certification Regarding Generative Artificial Intelligence” in the Northern District of Texas, issued 30 May 2023, which requires a certificate either that no portion of the filing was drafted by generative AI or that any AI-drafted language was checked for accuracy by a human being using print reporters or traditional legal databases.
What followed spans the range. Some orders require disclosure plus a certification of independent verification, as under Judge Vernon S. Broderick’s rules in the Southern District of New York; Judge John P. Cronan’s individual rules in the same district, as amended 23 October 2025, require a certification describing in detail the steps taken to verify AI-generated content. A minority prohibit the use of generative AI in filings, including Judge Michael J. Newman’s standing order in the Southern District of Ohio dated 27 August 2025. The Northern District of Texas took it district-wide in Local Civil Rule 7.2(f), which requires disclosure on the first page of a brief under its own heading. No authoritative count of federal judges with such orders exists: the Duke Law project Responsible AI in Legal Services compiled a tracker of court orders, local rules and guidelines on AI use, but states that it is no longer being updated and that its data runs through May 2025, and the totals in circulation trace to vendor pages rather than to a maintained list.
Circuit-level uniformity was proposed and refused. The United States Court of Appeals for the Fifth Circuit published a proposed amendment to Fifth Circuit Rule 32.3 and Form 6 on 22 November 2023 that would have required counsel to certify either that no generative AI was used or that all generated text, including citations and legal analysis, had been reviewed for accuracy and approved by a human. On 12 June 2024 the court declined to adopt a special artificial intelligence rule at that time, pointing to the accuracy obligations the existing rules already impose and stating that “I used AI” will not excuse an otherwise sanctionable filing. The practical consequence is a checklist item rather than a policy: the assigned judge’s standing order has to be read before the filing goes out.
Which rule are courts sanctioning under — Rule 11 or the inherent power?
Mostly Federal Rule of Civil Procedure 11, but not always, and the gap is instructive. Wadsworth v. Walmart Inc. and both Coomer v. Lindell orders proceeded under Rule 11, on the ground that a fabricated or misdescribed decision cannot make a legal contention warranted by existing law under Rule 11(b)(2). Mata v. Avianca, Inc. found Rule 11(b)(2) violations but hedged the source of the remedy, ordering its sanctions “pursuant to Rule 11, or, alternatively, its inherent authority.”
Johnson v. Dunn is the exception worth knowing. Because the offending filings were discovery motions under Rules 30 and 37, Judge Manasco held that Rule 11 “does not apply” by force of Rule 11(d), described the result as an apparent unintended anomaly in the rule, and observed that Rules 26 and 37 furnished no basis to address false statements of law in such motions. She also declined to extend Alabama Rule of Professional Conduct 3.3 to conduct that was reckless rather than knowing. That series of gaps left the court with its inherent power, which in the Eleventh Circuit requires bad faith or conduct tantamount to bad faith.
The sanction she chose under that power was not monetary. The court publicly reprimanded Matthew B. Reeves, William J. Cranford III and William R. Lunsford, ordered the order published to effect the reprimand, disqualified them from the case, and referred the matter to the Alabama State Bar and other licensing authorities, while releasing Butler Snow LLP and two other attorneys without sanction. The opinion says plainly that fabricating legal authority “demands substantially greater accountability than the reprimands and modest fines that have become common.”
Is anyone keeping a running count of these decisions?
Yes. Damien Charlotin, a research fellow at HEC Paris, maintains the AI Hallucination Cases database, which as of 14 September 2026 listed 2,041 decisions worldwide, of which 1,395 were in the United States. Its bar for inclusion is narrow by design: it tracks cases where the court or tribunal “has explicitly found (or implied) that a party relied on hallucinated content or material,” rather than mere allegations — though Charlotin states that at his discretion the list also holds some decisions where AI use was alleged but never confirmed.
Two cautions about using that number. It counts decisions rather than lawyers, and a single matter can generate several; and it is worldwide, so a United States figure has to be read off the country breakdown rather than the headline. What the series shows is direction rather than exposure, and the direction is steep.
Everything above is a description of what particular courts did on particular dates. It is not legal advice, it is not ethics advice, and it does not tell any firm what its jurisdiction requires; rules of professional conduct are adopted state by state and diverge. The parallel bar record is in bar guidance, state by state, the process that keeps this upstream of a filing is in citation verification, and the underlying error rate is in how often legal AI still gets it wrong.