Mata v. Avianca: The AI Hallucination Sanctions Case, Citation by Citation
The first sanctions order to enter the federal record over an AI-generated brief is short, factual, and unsparing. The brief cited cases. The cases did not exist. The reporters named did not contain the opinions claimed. The judges named had not written them. The attorneys had not opened the reporter volumes before signing the filing.
Mata v. Avianca is the canonical example of a hallucinated citation reaching a docket. The opinion is short enough to read in twenty minutes. What it teaches about pre-filing verification is the subject of this piece.
The matter, in one paragraph
Roberto Mata sued Avianca, Inc. for an injury sustained on a flight. Avianca moved to dismiss on statute-of-limitations grounds under the Montreal Convention. Plaintiff's counsel filed a brief in opposition. The brief argued that the limitations period had been tolled by bankruptcy proceedings. To support the argument the brief cited eight cases. Six of the eight could not be located in the reporters the brief named. The court ordered counsel to produce the opinions. Counsel produced typed excerpts that did not appear in any reporter database. Counsel ultimately acknowledged that the cases had been produced by ChatGPT during drafting.
The citations, line by line
Six citations were identified in the order to show cause. Each was presented in reporter form. None resolved to an opinion in the named volume.
Varghese v. China Southern Airlines Co., Ltd., 925 F.3d 1339 (11th Cir. 2019). The volume cited contains other decisions. The opinion named does not appear among them.
Shaboon v. Egyptair, 2013 IL App (1st) 111279-U. The Illinois Appellate Court docket contains no opinion in that number for that party.
Petersen v. Iran Air, 905 F. Supp. 2d 121 (D.D.C. 2013). The reporter pages cited contain unrelated decisions.
Martinez v. Delta Airlines, Inc., 2019 WL 4639462 (Tex. App. Sept. 25, 2019). The Westlaw identifier does not resolve to the opinion described.
Estate of Durden v. KLM Royal Dutch Airlines, 2017 WL 2418825 (Ga. Ct. App. June 5, 2017). Same failure mode.
Miller v. United Airlines, Inc., 174 F.3d 366 (2d Cir. 1999). The volume contains other Second Circuit decisions from 1999. The opinion named is not one of them.
Each citation has the surface form of a real citation. Each is the kind of error a Bluebook formatter does not flag, because Bluebook formatters check typography, not existence. Each is the kind of error a research platform does not flag, because the research platform was not asked the question.
What a pre-filing scanner returns on these inputs
We ran exactly that. Each of the six citations below was checked, as it appeared in the brief, against CourtListener’s citation corpus of roughly eighteen million reporter citations. The results are not illustrative. They are the verdicts the resolution returns.
Five of the six do not resolve to any opinion at all. The sixth is the more instructive failure: an opinion is published at 905 F. Supp. 2d 121 — it is just United States v. ISS Marine Services, Inc., not Petersen v. Iran Air. The model reused a real citation address and attached a fabricated case name to it. A formatter sees a well-formed citation. A reader who never opens the volume sees a case on point. The resolution sees neither.
The output is structured. The partner reviewing the scanner result sees five unresolved citations and one that resolves to the wrong case, and does not need to read the brief to know it is not ready to file.
This is not a marketing claim. The mechanism is plain. A citation either resolves to an opinion in the reporter or it does not. A scanner that performs that resolution returns a verdict on each attempt. When the input is a hallucinated citation, the verdict is negative. That is the test Mata v. Avianca failed and that the audit layer is built to run.
What the case actually changed
The reaction inside the bar to Mata v. Avianca took two forms. One form was retreat: AI tools are dangerous, no one should use them, this is what happens. The other form was operational: AI tools will be used, and the duty of competent supervision now includes a verification step. The second view has prevailed in every formal pronouncement since.
Federal judges in multiple districts have issued standing orders requiring disclosure of AI use and certification of citation verification. State bar associations have published opinionsframing the duty of technology competence to include verification of AI-generated work. Sanctions orders since Mata have repeatedly cited the same operational failure: the cite was not checked against the reporter before filing under Federal Rule of Civil Procedure 11.
The institutional posture is settled. AI use is permitted. Verification is the affirmative obligation. The question is no longer whether to run the audit. The question is what artifact the firm produces to show the audit ran.
The Verification Certificate as the answer to the order to show cause
The order to show cause in Mata v. Avianca asked counsel to produce the opinions cited and to explain how the citations were generated. Counsel could not produce the opinions because the opinions did not exist. There was no contemporaneous record of a verification step because no verification step had occurred.
A Verification Certificate is the record that the verification step occurred. The certificate lists the citations checked, the verification status for each, the source records consulted, a SHA-256 hash of the certificate content, a public token URL, and a timestamp. If the firm in Mata v. Avianca had filed a certificate alongside the brief, two outcomes change. The first is that the brief is never filed in that form, because the certificate would have shown six unresolved citations. The second is that, if it had been filed anyway, the record would show a deliberate decision to file unverified authority rather than an absence of verification.
The artifact is the difference between a verification that happened and a verification that can be proven. The matter file carries the certificate. The malpractice carrier reads it. The court, if it asks, finds it.
A short coda on why this matters now
The fact pattern in Mata v. Avianca is not exotic. The same sequence has repeated dozens of times in federal and state courts since 2023. The reported sanctions are the surface of a larger pool of corrected filings, withdrawn briefs, and quiet settlements that did not generate published orders. The underlying mechanism is the same in each: a citation that came out of a model, was not checked against the reporter, and was filed.
The audit layer is not a clever feature. It is the operational response to a failure mode that the profession has now observed enough times to name. Veritas runs the layer. The Filing Risk Scanner extracts the citations, resolves them against reporter and public-record sources, classifies each by verification status, and produces the certificate. The partner signs the brief after reading the certificate, not before.