case-law

Mata v. Avianca: When AI Hallucinations Meet Rule 11 — Sanctions, Bad Faith, and the Attorney's Duty of Competence

Decision & Law Editorial Team

March 5, 2026

14 min read

Mata v. Avianca: When AI Hallucinations Meet Rule 11 — Sanctions, Bad Faith, and the Attorney's Duty of Competence

The landmark Mata v. Avianca, Inc. decision, issued by the United States District Court for the Southern District of New York on June 22, 2023, has become one of the defining cases of the generative AI era in legal practice. What began as a straightforward personal injury claim governed by the Montreal Convention's two-year limitations period evolved into a cautionary tale of unchecked AI-assisted legal research: plaintiff's counsel submitted an affirmation in opposition containing no fewer than six entirely fabricated judicial opinions, complete with fake citations, invented quotes, and non-existent docket numbers — all generated by ChatGPT. The court, presided over by Judge P. Kevin Castel, imposed $5,000 in sanctions jointly and severally on attorneys Peter LoDuca and Steven A. Schwartz and their firm, Levidow, Levidow & Oberman P.C., under Federal Rule of Civil Procedure 11. The case raises fundamental questions about the attorney's gatekeeping duty, the limits of permissible reliance on AI tools, and what "subjective bad faith" means in the age of large language models.

The Underlying Claim: A Straightforward Limitations Bar

Before turning to the sanctions episode, it is necessary to understand the legal framework that rendered the hallucinated citations both reckless and ultimately futile. Plaintiff Roberto Mata alleged that on or about August 27–28, 2019, an Avianca flight attendant struck his left knee with a metal serving cart during a flight from El Salvador to John F. Kennedy Airport, causing serious injury.

Avianca moved to dismiss under Federal Rule of Civil Procedure 12(b)(6), invoking Article 35 of the Montreal Convention for the Unification of Certain Rules Relating to International Carriage by Air (Done at Montreal, Canada, on 28 May 1999, S. Treaty Doc. 106-45, 1999 WL 33292734) (the "Montreal Convention"), which extinguishes the right to damages "if an action is not brought within a period of two years, reckoned from the date of arrival at the destination." Mata's flight arrived at JFK on or about August 28, 2019; his complaint was not filed until February 2, 2022 — more than six months beyond the two-year hard stop.

Plaintiff advanced two arguments to escape the bar. First, he urged that the three-year negligence limitations period under New York Civil Practice Law and Rules section 214(5) should govern. The court rejected this squarely, consistent with Cohen v. American Airlines, Inc., 13 F.4th 240 (2d Cir. 2021), which reaffirmed that the Montreal Convention preempts state tort law and provides the exclusive avenue for damages claims within its scope.

Second, and more consequentially for the sanctions proceeding, plaintiff argued that the two-year period was tolled while Avianca was subject to the automatic bankruptcy stay under 11 U.S.C. § 362(a). The Second Circuit addressed the substantively identical provision of the predecessor Warsaw Convention in Fishman by Fishman v. Delta Air Lines, Inc., 132 F.3d 138 (2d Cir. 1998), characterizing the time limitation as a condition precedent to suit rather than an ordinary statute of limitations, and therefore not subject to equitable tolling principles. Following Fishman, district courts had consistently held that Article 35 creates a "hard stop" — a characterization emphatically confirmed in Ireland v. AMR Corp., 20 F. Supp. 3d 341 (E.D.N.Y. 2014). The court dismissed the complaint.

The substantive outcome was thus legally unremarkable. What followed was not.

The Anatomy of a Hallucination: From ChatGPT Prompt to Fabricated Precedent

Attorney Steven A. Schwartz, not admitted to practice in the Southern District of New York, performed all substantive legal work while his colleague Peter LoDuca — admitted to the district — served as attorney of record. When Schwartz began researching the Montreal Convention's limitations period and its interaction with the bankruptcy stay, the firm's Fastcase subscription offered limited access to federal cases. Rather than seeking alternative verified research tools, Schwartz turned to ChatGPT, a generative AI tool he had learned of through press reports and family conversations and had never previously used for legal research.

The failure unfolded in stages, each compounding the one before.

Stage one — the prompts. Schwartz queried ChatGPT to "argue that the statute of limitations is tolled by bankruptcy of defendant pursuant to montreal convention," then progressively narrowed his requests: "provide case law in support," "show me specific holdings in federal cases," "show me more cases." Directed to produce case law, the model complied by fabricating it. ChatGPT generated summaries and excerpts — never full opinions — of six cases purportedly supporting the tolling argument: Varghese v. China Southern Airlines Co., Ltd., 925 F.3d 1339 (11th Cir. 2019); Shaboon v. Egyptair, 2013 IL App (1st) 111279-U; Peterson v. Iran Air, 905 F. Supp. 2d 121 (D.D.C. 2012); Martinez v. Delta Airlines, Inc., 2019 WL 4639462 (Tex. App.); Estate of Durden v. KLM Royal Dutch Airlines, 2017 WL 2418825 (Ga. Ct. App.); and Miller v. United Airlines, Inc., 174 F.3d 366 (2d Cir. 1999). None existed.

Stage two — the affirmation. On March 1, 2023, LoDuca filed an "Affirmation in Opposition" — a document he signed under penalty of perjury without reading a single cited case or making any inquiry of Schwartz as to the nature or reliability of his research. The affirmation cited and quoted from the six fabricated opinions as if they were binding or persuasive authority.

Stage three — the red flags ignored. On March 15, 2023, Avianca's reply memorandum explicitly flagged that counsel had been "unable to locate most of the case law cited in Plaintiff's Affirmation in Opposition." No respondent sought to withdraw the affirmation or investigate. The court independently searched for the cases and could not locate multiple cited authorities. On April 11 and 12, 2023, Judge Castel issued orders directing LoDuca to file an affidavit annexing actual copies of the cited decisions, warning that failure to comply would result in dismissal under Rule 41(b). LoDuca requested an extension, falsely claiming he was "out of the office on vacation" — when in truth it was Schwartz who was traveling.

Stage four — the April 25 Affidavit. The affidavit filed on April 25, 2023, which Schwartz drafted and LoDuca signed without meaningful review, annexed purported excerpts of the fake opinions — documents generated or compiled from ChatGPT output that had no basis in actual judicial records. The "Varghese" excerpt alone exhibited multiple internal contradictions, inconsistent plaintiff names, unpaired quotation marks, an abrupt ending without conclusion, and internal citations to cases that themselves did not exist. One cited case, Zicherman v. Korean Air Lines Co., Ltd., 516 F.3d 1237 (11th Cir. 2008), could not be located — because the Federal Reporter citation at that coordinate belongs to Miccosukee Tribe v. United States, 516 F.3d 1235 (11th Cir. 2008), while a real Supreme Court decision named Zicherman exists at an entirely different citation. No explanation was offered.

Stage five — the slow disclosure. Only after the court issued an Order to Show Cause on May 4, 2023, did Schwartz begin to acknowledge in a May 25 affidavit that the cases were ChatGPT-generated fabrications. Even then, his characterization of ChatGPT as a mere "supplement" to his research was demonstrably false: under direct examination at the June 8 sanctions hearing, he conceded that ChatGPT was not supplementing his research — it was his research.

The Gatekeeping Duty and What Rule 11 Actually Requires

The court's legal analysis begins with Federal Rule of Civil Procedure 11(b)(2), which provides that by presenting a paper to the court, an attorney certifies that "the claims, defenses, and other legal contentions are warranted by existing law or by a nonfrivolous argument for extending, modifying, or reversing existing law or for establishing new law." This certification is not assessed solely at the moment of filing; the 1993 amendments to Rule 11 extended the obligation to the ongoing advocacy of filed contentions.

The court identified three key doctrinal propositions that together framed the sanctions analysis.

First, a fake opinion is not "existing law." By definition, a fabricated case citation cannot satisfy Rule 11(b)(2)'s requirement that legal contentions be warranted by existing law or by a non-frivolous argument to change it. Reliance on hallucinated precedent is not merely a weak legal argument — it is the submission of a false statement of the law to a tribunal, a separate and distinct violation from making an unwarranted legal contention about real authority.

Second, the filing of papers "without taking the necessary care in their preparation" is sanctionable as an abuse of the judicial system. Cooter & Gell v. Hartmarx Corp., 496 U.S. 384, 398 (1990). Rule 11 imposes an affirmative duty on each attorney to conduct a reasonable inquiry into the viability of a pleading before signing it. In the present case, LoDuca's failure to read even one cited case before swearing to the truth of an affirmation — particularly given his knowledge that Schwartz lacked familiarity with federal law, the Montreal Convention, and bankruptcy stays — constituted an absence of inquiry rather than merely an inadequate one.

Third, because the court imposed sanctions sua sponte rather than on motion of an adversary, the applicable standard required a finding of subjective bad faith — a heightened mens rea standard distinct from the objective unreasonableness standard applicable when an adversary initiates Rule 11 proceedings. Muhammad v. Walmart Stores East, L.P., 732 F.3d 104, 108 (2d Cir. 2013). Subjective bad faith encompasses the knowing and intentional submission of a false statement of fact, as well as conduct meeting the conscious avoidance standard: a person acts with the equivalent of knowledge when they are "aware of a high probability" of a fact and consciously avoid confirming it. United States v. Finkelstein, 229 F.3d 90, 95 (2d Cir. 2000).

Separate Findings of Bad Faith: LoDuca and Schwartz

The court insisted on separately assessing the conduct of each respondent, consistent with the principle that "the knowledge and conduct of each respondent lawyer must be separately assessed and principles of imputation of knowledge do not apply." Weddington v. Sentry Indus., Inc., 2020 WL 264431, at *7 (S.D.N.Y. Jan. 17, 2020).

Peter LoDuca was found to have acted with subjective bad faith on three independent grounds. First, he signed and filed the March 1 affirmation without reading a single cited case or making any inquiry — not a deficient inquiry, but a total absence of inquiry — despite knowing of Schwartz's unfamiliarity with the relevant areas of federal law. Second, he signed the April 25 affidavit without having participated in its preparation, without having reviewed it in draft, and without making a single inquiry of Schwartz as to whether the excerpts annexed were genuine — a cursory review would have revealed that "Varghese" was internally inconsistent and nonsensical, that "Zicherman" could not be found, and that several submissions were mere excerpts rather than full opinions. Third, he knowingly made a false statement to the court — claiming he was "out of the office on vacation" when seeking the extension of time for the April 25 affidavit, when in truth it was Schwartz who was traveling. This lie, the court found, was designed to conceal Schwartz's authorship of both affidavits and LoDuca's own lack of meaningful involvement.

Steven A. Schwartz was found to have acted with subjective bad faith through conscious avoidance. Most damning was the exchange over Varghese: Schwartz admitted at the sanctions hearing that he had looked up the case citation and could not find it — yet cited it in the brief anyway, operating under what he described as a "false assumption" that ChatGPT could not fabricate cases. The court rejected this explanation. By the time of the April 25 Affidavit, Schwartz had multiple independent signals that the cases did not exist: Avianca's reply brief had named them as unlocatable; the court's own orders showed it had searched and failed to find them; and Schwartz himself could not locate "Varghese" or "Zicherman." Rather than disclosing this, he submitted excerpts of fabricated opinions as if they were authentic. His shifting and contradictory accounts — asserting first that ChatGPT merely "supplemented" his research, then conceding it was his only source; asserting first that ChatGPT's assurances convinced him the cases were real, then asserting in a later declaration that he had grown suspicious the chatbot was fabricating — further undermined his credibility and supported the bad faith finding.

The Harms of AI Hallucinations in Litigation: A Systemic Lens

The court did not treat the sanctions as a purely individual matter. Judge Castel enumerated the systemic harms flowing from the submission of fake opinions with a clarity that has made this section of the opinion widely cited beyond the immediate case:

The opposing party is forced to expend time and resources exposing a deception that should never have occurred. The court's time is diverted from legitimate matters. The client may be deprived of arguments based on authentic precedent — the very precedent that might, in some cases, actually support their position. Judges and courts whose names are falsely attributed to fabricated holdings suffer reputational harm. The legal profession's credibility is eroded in the eyes of the public. And perhaps most insidiously, as the court noted, a future litigant might be tempted to challenge an authentic ruling by feigning doubt about its authenticity — a form of epistemic contamination that reaches beyond any individual case.

These harms, the court observed, are not mitigated by the fact that the fabrications did not achieve their intended purpose — the claim was untimely regardless of the tolling argument — or that no financial gain was sought. The wrong lies in the submission itself and in the failure to correct it once the deception became apparent.

The Question the Court Left Open: Competent AI Use

There is a passage at the outset of Judge Castel's sanctions opinion that has attracted as much attention as the sanctions themselves: "Technological advances are commonplace and there is nothing inherently improper about using a reliable artificial intelligence tool for assistance." This sentence is carefully hedged — the qualifier "reliable" carries the entire weight of the proposition — but it signals an important institutional posture. Courts are not categorically hostile to AI-assisted legal research. What they require is that the attorney's gatekeeping function remain intact.

The opinion thus raises, without fully answering, the question of what competent AI use in legal practice looks like. From the facts of Mata and the court's reasoning, several propositions can be derived. First, AI output must be verified against primary sources before being cited to a tribunal; the generation of a case name and citation by an AI tool is not sufficient verification of that case's existence. Second, AI tools that are known to "hallucinate" — to generate plausible-sounding but factually false outputs — cannot be treated as equivalent to verified legal databases such as Westlaw or LexisNexis; the attorney must understand the tool's limitations before relying on it for legal authority. Third, the obligation runs to each attorney who signs or files a document, regardless of who drafted it; the mere delegation of research to a more senior or junior colleague does not discharge the Rule 11 duty. Fourth, when an adversary raises questions about the existence of cited authorities, the attorney's obligation extends to investigating those questions and correcting the record — continuing to advocate for fabricated citations after receiving notice of their non-existence is itself a separate violation.

What the opinion does not address — because it was not before the court — is whether specific protocols (disclosure of AI use, mandatory verification steps, supervision requirements) might suffice to render AI-assisted research compliant with professional responsibility standards. That question is now being actively addressed by bar associations, courts adopting AI disclosure requirements, and legislative bodies in multiple jurisdictions.

Sanctions Calibration: Deterrence Without Disproportionality

Rule 11(c)(4) requires that sanctions be "limited to what suffices to deter repetition of the conduct or comparable conduct by others similarly situated." The court, conscious that it was acting simultaneously as accuser, fact-finder, and sentencing judge, exercised the restraint that this dual role demands.

Several mitigating factors bore on the sanctions calculation. The Levidow Firm had already arranged for mandatory continuing legal education on technological competence, AI tools, and notarization practices, conducted by outside counsel. Avianca had not sought fee reimbursement. The fake cases were not submitted for financial gain or out of personal animus toward any party. The individual respondents had expressed sincere remorse, had no history of disciplinary violations, and presented a low likelihood of recurrence. Additional mandatory education would have been redundant.

Balancing these considerations against the need for both specific and general deterrence — particularly given the significant publicity the case had already generated — the court imposed a penalty of $5,000 payable into the Registry of the Court, jointly and severally by LoDuca, Schwartz, and the Levidow Firm. The court also required respondents to notify their client and each judge whose name had been falsely attributed to a fabricated opinion, attaching the sanctions opinion, the hearing transcript, and the fake opinions submitted.

The decision not to impose the full measure of available sanctions — fee shifting, referral to disciplinary authorities, heavier fines — reflects a judgment that the public exposure itself, and the precedential weight of the published opinion, would serve the general deterrence function that a higher monetary sanction might otherwise serve. Whether subsequent conduct by the legal profession has vindicated that judgment remains an open empirical question.

Conclusions and Practical Takeaways

Mata v. Avianca is not primarily a case about artificial intelligence. It is a case about professional responsibility, the gatekeeping function of counsel, and what it means to sign a document under penalty of perjury. AI happens to be the instrument through which those duties were breached — which is precisely what makes the case instructive for practitioners navigating an environment in which generative AI tools are increasingly embedded in legal workflows.

The following propositions emerge from the court's analysis:

AI-generated case citations are not verified legal authority. Before citing any case in a court submission, counsel must independently verify its existence and the accuracy of quoted language in a primary, authenticated source. Generating a citation is not finding a case.
The duty of reasonable inquiry is non-delegable. Every attorney who signs or files a document bears personal responsibility for its contents under Rule 11. Reliance on a trusted colleague's work does not discharge the obligation to conduct at least a basic verification.
Continuing to advocate fabricated authority after receiving notice of its non-existence constitutes a separate violation. Rule 11 compliance is assessed not only at the moment of filing but throughout the pendency of the motion.
Subjective bad faith can be proven through conscious avoidance. An attorney who is aware of a high probability that cited cases do not exist and deliberately refrains from confirming that fact is treated as having acted with knowledge for sanctions purposes.
The harm of AI hallucinations is systemic, not merely individual. Courts will assess the broader consequences for the integrity of the adversary system, the reputation of the judiciary, and the credibility of the profession — not merely the prejudice to the immediate opposing party.

¿Necesitas un análisis personalizado sobre el uso de IA en tu práctica jurídica o el cumplimiento normativo de sistemas de IA? Contacta aquí

Caso analizado: Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023) (Castel, J.), decided June 22, 2023. The underlying motion to dismiss opinion is reported at Case No. 22-cv-1461 (PKC), ECF No. 55.

Back to News