Incident · Junior
Mata v. Avianca (2023): invented case law
Real incident. Key points: hallucinated citations and human in the loop.
Short answer
Real incident. Key points: hallucinated citations and human in the loop. In Mata v. Avianca, a lawyer used ChatGPT to research case law. The model invented court decisions, the lawyer submitted them to the court, and the court concluded the cases did not exist.
01 · CONTEXT
An ordinary case and a brief with cases that do not exist
In Mata v. Avianca, a lawyer used ChatGPT to research case law. The model invented court decisions, the lawyer submitted them to the court, and the court concluded the cases did not exist.
Roberto Mata alleged he was injured when a metal serving cart struck his left knee during a flight. Avianca moved to dismiss the case, arguing that the claims were time-barred under the Montreal Convention.
Steven Schwartz did all the substantive legal work. Peter LoDuca signed and filed the opposition brief, but was not its author, and before signing he did not review any of the cited authorities.
The brief cited and quoted supposed court decisions, said to be published in the Federal Reporter, the Federal Supplement, and Westlaw.
Schwartz researches on Fastcase, but the firm's account had limited access to federal cases.
Schwartz uses ChatGPT, which he himself says he wrongly assumed was a kind of super search engine.
When he asks for cases supporting the argument, ChatGPT complies by inventing decisions, and Schwartz did not have the full text of any of them.
The brief, drafted by Schwartz, is signed and filed by LoDuca, who did not review the cited authorities.
Avianca's defense reports that it could not locate most of the cited case law, and the court itself cannot locate several of the authorities.
At no point did those involved write to the court asking to withdraw the brief.
02 · WHAT CHATGPT PRODUCED
Decisions that look real
When Schwartz asked for cases supporting the argument, ChatGPT responded by inventing them. When preparing the brief, Schwartz did not have the full text of any of the generated decisions.
| Appearance | What the court found | |
|---|---|---|
| Origin of Varghese | Presented as a decision of a panel of judges of the Eleventh Circuit court of appeals | The court of appeals clerk's office confirmed that the decision was not authentic |
| Text quality | Features superficially consistent with real court decisions | Legal analysis described by the court as gibberish |
| Internal citations | Citations and excerpts from other decisions | Among them, decisions that also do not exist |
At the hearing, Schwartz stated that it never occurred to him that the case could be made up.
03 · HOW THE CASE ESCALATED
From the defense's warning to the sanctions
Avianca said it could not locate most of the cited cases, and the court issued orders on April 11 and 12 requiring the real cases to be produced. The first acknowledgment that the brief cited fake cases generated by ChatGPT came in a Schwartz declaration of May 25.
In his testimony, Schwartz admitted that he entered the Varghese citation to look it up and did not find it. The court concluded that he knew of facts indicating a high probability that Varghese and Zicherman did not exist and consciously avoided confirming this.
Schwartz also stated in writing that ChatGPT merely supplemented his research. The court found the statement misleading, since ChatGPT was the sole source of his substantive arguments.
The sanctions included a $5,000 fine, joint among the respondents and paid to the court, as well as letters to the plaintiff Roberto Mata and to each judge falsely named as the author of the fake decisions.
04 · SAFER DESIGN
A citation only goes in if someone checks the source
A legal research assistant can be useful, but the text it generates should be treated as a hypothesis, not a source. One option is to separate text generation from verification: each citation goes through a verifier that looks it up in an authorized case law database, and a person reads the full decision before the filing goes out. This design connects to the topic of regulatory compliance in system design, where professional responsibility becomes a design requirement.
The design below is a teaching proposal, not an account of what the firm had: the sources do not describe a verification system.
The lawyer requests the research from the assistant, which returns a draft with the citations marked as proposed, not yet verified.
Each proposed citation goes through the verifier, which looks it up by identifier in the authorized case law database.
A citation that is not found is blocked: the draft does not advance with it, and the model's own opinion about it does not count as proof.
The citations that are found go to a person, who reads the full decision and confirms or discards each one.
Each verification is logged, with who approved it, and the filing only goes out with the confirmed citations.
| Possible failure | Design control | |
|---|---|---|
| Citation with no origin | The model generates a case that looks real but does not exist | Verifier looks up each citation in an authorized database and blocks what it cannot find |
| Sham verification | The signer does not read the authorities | Mandatory human review, with the full text of the decision, before filing |
| Asking the model itself | Asking the model to confirm whether the case is real | Treat the model's reply as unverified; only the authorized database counts |
| Error found late | Nobody knows what was checked | Log each verification and who approved it |
05 · LESSONS
What to take into your design
The court listed the harms of submitting fake decisions: the opposing party wastes time and money exposing the deception, and there is potential harm to the reputation of judges and courts whose names are falsely invoked as authors.
Corvino, a member of the firm, acknowledged responsibility and pointed to corrective measures, among them an expanded Fastcase subscription and continuing education programs.
Frequently asked questions
What is Mata v. Avianca (2023): invented case law?
Real incident. Key points: hallucinated citations and human in the loop.
What does the Mata v. Avianca (2023): invented case law design cover?
It is organized in these sections: An ordinary case and a brief with cases that do not exist, Decisions that look real, From the defense's warning to the sanctions, A citation only goes in if someone checks the source, and What to take into your design.
How long is Mata v. Avianca (2023): invented case law and what level is it?
It takes about 9 minutes to read and is rated Junior level.
Now design it yourself
Run this case as an interview in the System Design school: you make the decisions, the tutor challenges them and you see how your design scores.