Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y.), is the federal court case in which attorneys Steven A. Schwartz and Peter LoDuca of the New York firm Levidow, Levidow & Oberman were sanctioned $5,000 by Judge P. Kevin Castel for submitting a legal brief containing citations to six cases that did not exist. The fabricated citations were generated by ChatGPT, which attorney Schwartz had used to conduct legal research without verifying the existence of the cited authorities.
The sanctions order, issued on June 22, 2023, became a reference point for courts, bar associations, and legal educators addressing the professional responsibility implications of generative AI use in legal practice. The case is not notable for its underlying facts — a personal injury action against Avianca, Inc. — but for establishing, in a concrete judicial record, what happens when an attorney submits AI-generated citations without verification.
The term "hallucination" in the context of large language models refers to the generation of plausible-sounding but factually false content — including case citations that appear legitimate but do not exist in any court's record. Mata v. Avianca is the most widely cited judicial example of this failure mode causing professional consequences.
The core failure. On March 1, 2023, attorney Schwartz submitted an opposition brief in Mata v. Avianca, Inc. that cited six cases in support of arguments about the application of the Montreal Convention to international aviation claims. When Avianca's counsel filed a reply noting that several of the cited cases could not be located, Judge Castel ordered counsel to provide copies of the cited cases. Schwartz and LoDuca submitted documents — generated by ChatGPT when asked to produce the cases — that were not actual judicial opinions. When the court investigated further, it emerged that the cases did not exist in any court's records.
The six fabricated cases. The brief cited: Varghese v. China Southern Airlines Co. Ltd.; Shaboon v. Egyptair; Petersen v. Iran Air; Estate of Durden v. KLM Royal Dutch Airlines; Zicherman v. Korean Air Lines Co. (not the actual Zicherman litigation that does exist); and Matter of Korean Air Lines Disaster of September 1, 1983 (a distorted version of real litigation). These citations were not errors about the holdings of real cases — the cases themselves did not exist in the form cited. ChatGPT had fabricated citations with plausible-sounding party names, airline defendants, and citation formats.
The sanctions and their basis. Judge Castel's June 22, 2023 sanctions order imposed $5,000 in monetary sanctions on both attorneys and the law firm under Federal Rule of Civil Procedure 11(b)(3), which requires that factual contentions in filings have evidentiary support or, if specifically identified, will likely have such support after a reasonable opportunity for further investigation. The court found that submitting citations to non-existent cases violated this obligation. The court additionally required the sanctioned attorneys to provide copies of the sanctions order to any judge before whom they appeared in the following year and to complete continuing legal education on AI use in legal practice.
Scale of the problem. Mata v. Avianca was the most prominent but not the only documented instance of AI-hallucinated citations reaching courts. Bloomberg Law's tracking of similar incidents documented more than 27 cases in US federal and state courts between June 2023 and December 2025 in which counsel submitted AI-generated citations that could not be verified, resulting in sanctions, reprimand, or required remedial steps. These cases span multiple jurisdictions, practice areas, and firm sizes — from solo practitioners to large regional firms.
Professional responsibility dimensions. The American Bar Association issued Formal Opinion 512 in July 2024, addressing the competence and diligence implications of generative AI use. The Opinion stated that Model Rule 1.1 (Competence) requires lawyers to understand the benefits and risks of the technology they use in their practice, including the tendency of large language models to generate inaccurate information. Model Rule 1.3 (Diligence) requires lawyers to act with commitment and dedication to the interests of clients, which the ABA interpreted to include taking reasonable steps to verify AI outputs before relying on them. The ABA did not issue a blanket prohibition on AI use, but the Opinion makes clear that submitting unverified AI-generated citations to a tribunal is a competence failure under the Rules.
Court-specific disclosure requirements. Following Mata v. Avianca, at least twelve US federal courts and several state courts had adopted specific local rules or standing orders addressing AI use in filings by the end of 2025. These requirements vary: some require disclosure when AI tools were used to assist in drafting; others require certification that AI-generated citations were verified; a small number require disclosure of which AI tools were used. Lawyers practicing in multiple jurisdictions must track the specific requirements of each court.
How It Works (Technical)
Why language models hallucinate citations. Large language models generate text by predicting the most probable next token given the preceding context. They are trained on large corpora of text that include legal documents, law review articles, and judicial opinions. During training, the model develops statistical associations between citation patterns — the structure of case citations, common party types, frequently cited courts — without developing a reliable memory of specific case holdings or a mechanism for verifying whether a cited case exists.
When asked to research a legal question, a language model generates a plausible response, including citations that fit the expected pattern of legal authority. The model does not search a database of actual cases; it generates text that resembles a research memorandum. This distinction is fundamental. A hallucinated citation is not a mischaracterisation of a real case — it is a fabricated string of text that resembles a case citation but does not correspond to any actual judicial decision.
The probability of hallucination increases with: the obscurity of the legal question (less training data means more extrapolation); the specificity of the citation requested (requesting a case on point for a narrow proposition increases fabrication risk); and prompting that implicitly pressures the model to produce an answer (asking "what cases support this argument?" invites fabrication more than "does this argument have case support?").
General-purpose vs. law-specific AI tools. General-purpose language models like ChatGPT — the tool used in Mata v. Avianca — are not connected to legal databases. They generate citations from training data, not from real-time searches of Westlaw, Lexis, or PACER. This is the fundamental distinction between general-purpose AI tools and legal AI research platforms: the latter query verified databases of actual cases before generating research outputs.
The rubber duck problem. Schwartz's conduct included asking ChatGPT whether the cases it had generated were real. ChatGPT responded that they were. This illustrates a critical point: a language model cannot reliably verify its own outputs. Asking ChatGPT "is this case real?" does not produce an independent verification — it produces another token prediction based on the pattern of the prompt. Self-verification through a general-purpose language model is not verification.
How hallucination differs from research error. Traditional legal research errors involve misreading, mischaracterising, or incorrectly applying a real case. Hallucination is categorically different: the source itself does not exist. This matters for professional responsibility analysis because the degree of culpability is different — a lawyer who misreads a case may have exercised judgment that turned out to be wrong; a lawyer who submits a non-existent citation to a court has submitted a representation that is straightforwardly false.
How Legal AI Vendors Address It
Westlaw Precision AI (Thomson Reuters) integrates citation verification directly into its research workflow. Research outputs include citations that are pulled from the Westlaw case law database, not generated by the language model from training data. Before any citation appears in a research output, it is verified against Westlaw's database of actual cases. The Precision AI interface also includes inline Shepard's-equivalent citation analysis, alerting researchers to cases that have been overruled or criticised. Limitation: Westlaw Precision AI's citation verification covers the cases it surfaces; a lawyer who takes a Westlaw output and then supplements it with general-purpose AI research may introduce unverified citations into a workflow that otherwise had appropriate safeguards.
Lexis+ AI (LexisNexis) incorporates Shepard's Citation Service verification into its AI research outputs. Citations generated through Lexis+ AI's research tools are drawn from the LexisNexis case law database and are accompanied by Shepard's signals indicating citation history and validity. Limitation: Lexis+ AI's citation verification applies to the AI-assisted research workflow within the platform. Lawyers who export research summaries and further edit them in word processors may inadvertently alter citations in ways that introduce errors.
CoCounsel (Thomson Reuters) provides a dedicated citation check workflow that allows lawyers to input a list of citations and receive verification of their existence and current status against the Westlaw database. This workflow is directly responsive to the Mata v. Avianca failure mode: it provides an explicit verification step for citations before filing. Limitation: CoCounsel's citation check verifies existence and validity but does not verify the accuracy of how the lawyer has characterised a case's holding — a citation can be real and valid while still being misapplied.
What general-purpose AI tools cannot provide. ChatGPT, Claude, Gemini, and similar general-purpose language models — even in their most recent 2025–2026 versions with web browsing capabilities — are not reliable citation verification tools. Web browsing capabilities improve the chance that a model will locate an actual case, but they do not provide the systematic, database-verified citation checking that legal research platforms offer. Using a general-purpose AI tool with web access to verify citations is materially better than not verifying at all, but it is not equivalent to Westlaw or Lexis citation verification.
How Lawyers Should Verify / Apply It
-
Establish a zero-tolerance policy for unverified citations in court filings. Every citation in every court filing must be verified against a primary legal database — Westlaw, Lexis, or an equivalent — before filing. This policy must apply regardless of the source of the citation: case citations generated by AI tools, citations found in secondary sources, and citations provided by colleagues all require independent verification. The verification step must be documented in the matter file.
-
Separate AI-assisted drafting from citation verification. Use AI tools for drafting arguments, summarising case law, and identifying relevant legal principles. Use a verified legal database — not the AI tool itself — for citation verification. Never ask the AI tool that generated citations to verify its own outputs. Self-verification of AI-generated citations is not verification.
-
Check applicable court AI disclosure rules before filing. As of 2026, dozens of federal and state courts have adopted local rules or standing orders addressing AI use. Before filing in any court, check whether that court requires: disclosure of AI assistance in drafting; certification that AI-generated citations were independently verified; or disclosure of which AI tools were used. The requirements vary and are updated frequently. Maintain a current list for each jurisdiction in which you practice.
-
Implement a citation verification checklist for junior associates and paralegals. In matters where AI tools are used for research or drafting, implement a mandatory checklist requiring that each citation appearing in a court filing be verified against a primary database. The checklist should require the verifier to record: the citation verified, the database used for verification, the date of verification, and the verification outcome (exists; overruled; distinguished). This checklist becomes part of the matter file.
-
Treat a missing citation as a filing blocker. If a citation cannot be verified — it does not appear in Westlaw or Lexis, or the case that appears does not stand for the proposition cited — the citation must be removed or corrected before filing. There is no circumstance in which submitting an unverified or unverifiable citation to a court is professionally acceptable. The professional obligation runs to the tribunal (Model Rule 3.3 — Candor Toward the Tribunal), not merely to the client.