The Verification Duty: What the Mata Sanctions Cases Actually Require

Three years after a federal judge in Manhattan sanctioned two lawyers for filing a brief stuffed with cases that did not exist, courts are still issuing hallucination sanctions at a steady clip. That is a professional embarrassment, but it is no longer a doctrinal mystery. The line of cases that began with Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023), has matured into something a young lawyer can actually work with: a verification duty with reasonably clear content. This piece distills it.

The baseline: Mata’s two moves

Judge Castel’s Mata opinion made two moves that everything since has built on. First, it declined to treat the technology as the problem. There is nothing inherently improper about a lawyer using an AI tool for research assistance; the opinion said so expressly. Second, it located the duty in rules that predate the technology by decades. Rule 11(b)(2) of the Federal Rules of Civil Procedure makes every signature a certification that the legal contentions are warranted by existing law, and the court described lawyers as playing a gatekeeping role in ensuring the accuracy of what they file.

The detail most worth remembering from Mata is not the fabricated citations. It is how the lawyers responded when opposing counsel and the court flagged them. Rather than checking, one of the lawyers asked the chatbot itself whether the cases were real — and accepted its reassurance. The court’s bad-faith finding rested substantially on the doubling down, not the initial error. That allocation of blame has held up across the later cases: the initial hallucination is negligence; standing behind it after challenge is what converts a correctable mistake into sanctionable misconduct and a candor problem under Model Rule 3.3.

The line hardens

The appellate courts joined quickly. In Park v. Kim, 91 F.4th 610 (2d Cir. 2024), the Second Circuit confronted a brief citing a nonexistent case generated by ChatGPT and put the point at its simplest: Rule 11 requires, at a minimum, that an attorney read and verify the authorities cited in a filing. The court referred counsel to its grievance panel — a signal that this conduct is a disciplinary matter, not merely a costs question.

The duty is not limited to lawyers or to federal court. In Kruse v. Karlen, 692 S.W.3d 43 (Mo. Ct. App. 2024), a Missouri appellate court dismissed a pro se litigant’s appeal after his brief cited fabricated authority, holding self-represented parties to the same obligation of candor. Federal district courts have added remedial teeth: in Gauthier v. Goodyear Tire & Rubber Co. (E.D. Tex. 2024), the court paired a monetary sanction with mandatory education on generative AI in legal practice, and in Wadsworth v. Walmart Inc. (D. Wyo. 2025), the court fined plaintiffs’ counsel at a prominent national firm and revoked a pro hac vice admission after AI-fabricated citations appeared in motions in limine. Courts have also grown willing to sanction firms, not just the individuals who pressed the button — a development supervising attorneys should read alongside Model Rules 5.1 and 5.3.

The content of the duty

Read together, the cases specify the verification duty with more precision than most commentary acknowledges:

1. Retrieve and read every cited authority. Not the model’s summary of it — the authority itself, from an official or established source. Park states this floor explicitly.
2. Verification is personal and non-delegable to the tool. Asking a generative system to confirm its own output is not verification; it is the Mata error in a second iteration. The same logic applies to checking one model’s citations with another model.
3. The duty covers more than existence. A real case cited for a fabricated holding, or adorned with a fabricated quotation, violates the same certification. Several of the recent sanctions orders involve exactly this subtler failure — real citations, invented substance.
4. Candor obligations attach the moment a problem surfaces. Prompt investigation and correction has repeatedly mitigated sanctions; bluffing has repeatedly aggravated them.

The ethics infrastructure now says the same thing. Comment 8 to Model Rule 1.1 has required technological competence since 2012, and ABA Formal Opinion 512 (2024) applied it to generative AI directly: lawyers must understand the tools well enough to know that outputs require independent verification, and competence cannot be outsourced to the software.

Why disclosure mandates are the wrong lesson

After Mata, dozens of federal judges issued standing orders requiring disclosure or certification of AI use, beginning with Judge Brantley Starr in the Northern District of Texas. The trend is understandable and, I think, mostly misguided. The sanctions cases demonstrate that existing rules — Rule 11, Rule 3.3, Rule 1.1 — already reach the misconduct, and they reach it whether the fabrication came from a chatbot, a careless associate, or a lawyer’s imagination. Disclosure regimes tax the compliant, do nothing to deter the careless (who by definition do not know their citations are fake), and imply that AI-assisted work is presumptively suspect when the actual rule is that all work is subject to the same certification.

The better frame for a young practitioner is this: generative AI has not changed your obligations at all. It has changed the failure mode. The tools produce fluent, confident, format-perfect error, which defeats the informal heuristics lawyers previously used to spot weak research. The verification duty is the same one your predecessors owed; it simply can no longer be discharged by skimming. Pull the case. Read the case. Then sign.