Most people who reference Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. 2023), know the headline version: a lawyer used ChatGPT, it made up cases, sanctions followed. The actual order Judge P. Kevin Castel issued is more specific than the headline version, and the specifics matter for anyone trying to draw the right lesson from it rather than the easy one.
What was actually filed
Attorney Steven Schwartz, researching an aviation personal-injury matter, used ChatGPT to find supporting precedent and submitted a brief citing several cases — including one called Varghese v. China Southern Airlines — that included fabricated case names, docket numbers, and quoted excerpts of purported holdings. Peter LoDuca, the attorney of record, signed the filing. Neither had independently verified the citations against a primary legal database before filing.
The part most retellings skip: what happened after opposing counsel raised the issue
The sanctions order is unusually detailed about what happened between the fabrication being flagged and the hearing. When Avianca’s counsel could not locate the cited cases and raised the issue with the court, Schwartz did not immediately disclose that ChatGPT had generated the citations. Instead, he asked ChatGPT to confirm the cases were real, received a false reassurance, and submitted that reassurance to the court as supporting evidence — effectively compounding the original error with a second AI-generated, unverified representation. Judge Castel’s order specifically identifies this sequence, not merely the original fabrication, as the basis for finding bad faith.
Why this sequencing detail changes the lesson
Judge Castel stated directly that he might have handled the matter differently — potentially without monetary sanctions — had the attorneys come forward candidly once the problem was identified, rather than doubling down. This is the detail most retellings of the case compress out, and it matters enormously for how lawyers should think about their own exposure: the sanctions in this specific case were driven as much by the response to being caught as by the underlying fabrication itself.
The monetary and non-monetary components
The court imposed a $5,000 monetary sanction against the attorneys and their firm, Levidow, Levidow & Oberman. Beyond the fine, the order required the attorneys to notify their client of the sanctions and to notify each judge falsely identified in the fabricated citations (the fake cases had been attributed to specific, real judges) — a notably personal and reputationally significant remedy beyond the dollar figure.
What the order does not say
The order does not hold that using ChatGPT for legal research is itself sanctionable conduct. It holds that filing a brief with fabricated content, and then failing to correct the record candidly once challenged, is sanctionable — a standard that predates generative AI by decades and traces to Rule 11’s existing reasonable-inquiry requirement. This distinction matters because it means the fix was never “stop using AI,” it has always been “verify before filing, and correct immediately and candidly if you find an error after the fact.”
Why this case is still the reference point more than two years later
Mata was not the first instance of an AI-fabricated citation reaching a court, but the specificity of Judge Castel’s order — walking through exactly how the fabrication was discovered, exactly what Schwartz did once challenged, and exactly why that response (not just the initial error) drove the sanction — makes it unusually useful as a training document, not just a cautionary headline. Subsequent cases in the public record (now well past 1,750 documented per Damien Charlotin’s database) largely follow the same underlying pattern the order first laid out in detail.
The hearing itself: what the transcript reveals
Beyond the written order, the sanctions hearing transcript in Mata v. Avianca is itself instructive reading for anyone who wants to understand how a judge actually walks through this kind of failure in real time. Judge Castel questioned Schwartz directly and methodically about each fabricated citation, and the record shows Schwartz struggling to explain, in real time, basic details about cases he had cited as controlling authority — a strong signal to the court that no genuine research had occurred, beyond the citations' non-existence alone.
The role of opposing counsel in surfacing the issue
It is easy to forget, given how much attention the case has received, that the fabrication was not caught by the court on its own initiative — it was caught by Avianca's counsel, who could not locate the cited cases while preparing a response and flagged the discrepancy to the court. This detail reinforces a theme that recurs across nearly every documented case: opposing counsel's ordinary research process, not some specialized AI-detection effort, is very often what surfaces the fabrication in the first place.
How the legal press coverage shaped the case's lasting influence
The scale of mainstream press coverage — well beyond typical legal-press attention for a sanctions order — is itself part of why this case still functions as the reference point more than two years later. Reuters, the New York Times, and numerous legal trade publications all covered the case in detail, which meant the specific facts (rather than a generic summary) became widely known across the profession, giving Judge Castel's detailed reasoning far more reach and influence than a typical district court sanctions order would ordinarily receive.
The firm's own internal response after the sanction
Levidow, Levidow & Oberman, the firm involved, reportedly implemented new internal review procedures for any AI-assisted research following the sanction, according to subsequent coverage of the matter. This is a useful, if less-discussed, coda to the case: the sanction did not just resolve the immediate dispute, it drove a specific institutional response at the firm level, which is precisely the kind of concrete process change the broader profession has been urged to adopt in response to the case's wider publicity.
Why some commentators think the sanction was too lenient, and others too harsh
The $5,000 figure has drawn commentary from both directions: some legal ethics commentators argued the sanction was too modest given the severity of the underlying conduct (fabricating case law and then compounding it with a false reassurance to the court), while others argued Judge Castel's approach was appropriately calibrated to a genuinely novel situation at the time, given that ChatGPT had only been publicly available for a matter of months when the underlying brief was filed. This disagreement itself is instructive: it shows the case sits at a genuine inflection point in how the profession was still calibrating appropriate consequences for a new category of error.
A common question
If I find a hallucinated citation in something I already filed, what does Mata suggest I should do?
Based on Judge Castel’s own stated reasoning, the specific lesson is: disclose promptly and candidly to the court and opposing counsel rather than attempting to defend or re-verify the citation through the same tool that produced it. The order strongly suggests candor at the point of discovery is treated very differently from candor extracted only after being caught.