Mata v. Avianca: What every lawyer needs to know about AI hallucinations

In May 2023, attorney Steven A. Schwartz of Levidow, Levidow & Oberman, P.A. filed a routine personal injury brief in the Southern District of New York. His client had been injured by a metal serving cart on an Avianca Airlines flight from El Salvador to JFK in 2019. The case was straightforward — except that Schwartz cited six federal cases that did not exist.

What followed became the case that every law student, every bar association, and every state supreme court now studies as the cautionary tale of generative AI in legal practice.

This is what happened, what it means, and why it matters even more today than it did the day Judge P. Kevin Castel issued his sanctions order.

The case beneath the case

The plaintiff in Mata v. Avianca, Inc., Case No. 22-cv-1461 (PKC) (S.D.N.Y.), was Roberto Mata. The facts were unremarkable. In August 2019, during a flight from El Salvador to JFK, Mata's left knee was struck by a metal serving cart. The complaint, filed in February 2022, alleged personal injury under the Montreal Convention.

When Avianca moved to dismiss, the Levidow firm filed a response in March 2023 — a ten-page affirmation by Schwartz citing eight cases in support of plaintiff's position. Six of them were fabricated. None of those six cases existed. Not one.

Among the citations:

• Varghese v. China Southern Airlines Co., Ltd., 925 F.3d 1339 (11th Cir. 2019)
• Shaboon v. Egyptair, 2013 IL App (1st) 111279-U (Ill. App. Ct. 2013)
• Petersen v. Iran Air, 905 F. Supp. 2d 121 (D.D.C. 2012)
• Martinez v. Delta Airlines, Inc., 2019 WL 4639462 (Tex. App. Sept. 25, 2019)
• Estate of Durden v. KLM Royal Dutch Airlines, 2017 WL 2418825 (Ga. Ct. App. June 5, 2017)
• Miller v. United Airlines, Inc., 174 F.3d 366 (2d Cir. 1999)

The case numbers exist. The reporters exist. The courts exist. The cases do not.

How it happened

Schwartz had practiced law in New York for over thirty years. He had never used Westlaw. The Levidow firm subscribed to Fastcase. When the Mata case was filed in state court and then removed to federal court, Schwartz — handling the federal motion — used a generative AI tool for the research.

By his own affidavit, Schwartz asked the AI to provide cases addressing the tolling effect of bankruptcy stays on statutes of limitation in international aviation cases. The model responded with case names, citations, and detailed factual summaries. Schwartz used the responses verbatim.

He did, in fact, ask the model whether the cases were real. The model said yes. He asked for sources. The model provided fabricated quotes from the fabricated cases. He asked the model to confirm the cases were good law. The model confirmed. None of it was true.

When Avianca's counsel reviewed the brief, they could not locate the cases. They notified the court. Judge Castel ordered Schwartz and Peter LoDuca — who had signed the brief as counsel of record — to appear and explain.

The hearing

The June 8, 2023 hearing in front of Judge Castel is now part of legal AI history. Schwartz testified that he had used a generative AI tool, that he had not understood it could generate false information, and that he had attempted to verify the cases by asking the model itself to confirm them. He apologized. He said he had been "operating under the false perception that this website could not possibly be fabricating cases on its own."

LoDuca, whose name was on the brief, admitted he had not read the cited cases either.

Judge Castel was not unkind. He was, however, unambiguous. In his June 22, 2023 sanctions order, Judge Castel wrote that the attorneys had "abandoned their responsibilities when they submitted non-existent judicial opinions with fake quotes and citations attributed to genuine cases." The firm and the two attorneys were sanctioned $5,000 each. The court ordered them to notify each judge whose name had been attached to a fabricated decision. They were required to inform Mr. Mata himself of what had happened.

The sanction amount was modest. The professional consequence was not.

Why this case matters more now than then

Mata v. Avianca is now in the ABA's Generative AI Toolkit, in the California State Bar's Practical Guidance on the Use of Generative Artificial Intelligence, in the New York State Bar Association's AI Task Force Report, in the Florida Bar's Special Committee on AI Tools Report, and in the curriculum of every legal ethics CLE in the country.

But the deeper reason it matters is this: the conditions that produced Schwartz's mistake have not been eliminated. They have spread.

In the years following the sanctions order, dozens of similar cases have surfaced across federal and state courts. Lawyers in Texas, Colorado, Massachusetts, Missouri, Florida, and California have been sanctioned, fined, ordered to disclose AI use, referred to disciplinary authorities, and in some cases publicly named in court orders for filing briefs with fabricated citations generated by consumer AI tools.

The pattern is consistent. A generative AI model produces a citation that looks authoritative. The lawyer relies on it. The lawyer files. Opposing counsel discovers the fabrication. The court issues an order to show cause.

The "I didn't know it would hallucinate" defense, which Schwartz raised in good faith, no longer works. Every state bar in the country has now issued ethics guidance acknowledging that generative AI can produce false information. The duty to verify is now squarely on the lawyer.

What this means for every lawyer using AI today

There are three implications that every practicing attorney needs to understand.

First, the duty to verify cannot be delegated to the tool itself. Asking a generative AI model whether its own output is accurate is not verification. These models are trained to be helpful and confident; they will assert the truth of their own fabrications. Verification requires an independent source — Westlaw, Lexis, PACER, or any other authoritative legal research database. A model checking its own work is structurally incapable of catching its own hallucinations.

Second, the bar's expectation has changed. In 2023, "I didn't know" was a colorable defense. In 2024, ABA Formal Opinion 512 made clear it is not. The competence duty under Model Rule 1.1 now includes a duty to understand the limitations of the technology you use. The supervision duties under Rules 5.1 and 5.3 now apply to AI outputs the same way they apply to associate work product. The candor duty under Rule 3.3 means citing a fabricated case — even unknowingly — can be a serious ethics violation.

Third, defensibility is now a workflow question, not a tool question. A lawyer who can show — on demand, in real time — that every citation in a brief was validated against current law, that every fact was sourced, and that the chain of verification is recorded and reviewable, is in a fundamentally different posture from a lawyer who cannot. The first lawyer has a defense. The second lawyer has a sanctions risk.

What a defensible workflow actually looks like

In a defensible AI workflow, three things are always true.

First, the AI's output is verified against authoritative sources before it reaches the lawyer. Not after the brief is filed. Not when opposing counsel objects. Before the lawyer ever sees the citation, an independent system has confirmed the case exists, is correctly cited, and is still good law.

Second, the verification is recorded. The lawyer can produce, on demand, a record showing exactly which authorities were checked, when, and against which source. This is not a trust exercise. It is a defensibility exercise.

Third, the record is tamper-evident. A simple log file in a spreadsheet is not enough. The record needs to be cryptographically signed, timestamped, and chained so it can be authenticated if challenged. A lawyer producing a verification log that opposing counsel cannot dispute is in a different position than a lawyer producing an editable file.

These three requirements are the bar that any AI tool used in legal practice should be measured against. The generic consumer AI tools — the ones that produced Mata v. Avianca and every sanctions case after it — do not meet them. They were not built for this. They were built to write essays and answer questions. They were not built to produce evidence that survives a Daubert challenge.

Why we built DDEAS

Discover Docket exists because the gap between what generative AI can do for lawyers and what generative AI can defensibly do for lawyers is enormous, and the cost of getting it wrong is career-ending.

DDEAS — the Discover Docket Ethical and Accountability Standards — is the cryptographic backbone behind every output JILL produces. Every citation is validated against current law before it reaches your screen. Every fact is sourced. Every action is signed, hashed, and chained into an immutable audit log you can show to a judge. If the citation isn't real, you never see it. If you need to defend a brief, you have the record.

It is not a feature. It is the framework that makes generative AI defensible for litigation.

The lesson of Mata v. Avianca is not that AI is dangerous. It is that AI without defensibility infrastructure is dangerous.

Build the infrastructure. Practice with confidence.