LLM09:2025Misinformation
Slide 10 of 27
Part 2 · TypesSlide 10
Slide 10 · Fabricated Citations
The case that made lawyers fear chatbots.
Mata v. Avianca, June 2023.
Confirmed Incident · June 2023 · U.S. District Court, S.D.N.Y.
Mata v. Avianca: ChatGPT Fabricates Six Case Citations
No CVE · Sanctions Order by Judge P. Kevin Castel · June 22, 2023

The setup: Attorney Steven Schwartz used ChatGPT to research case law for an aviation injury suit. ChatGPT returned six cases — complete with names, courts, docket numbers, and legal summaries.

What was fabricated: Every case was invented: Varghese v. China Southern Airlines, Martinez v. Delta Airlines, Zicherman v. Korean Air, Shaboon v. EgyptAir, Petersen v. Iran Air, Miller v. United Airlines. None existed in any court record.

The overreliance loop: Opposing counsel could not locate any of them. The judge demanded copies. The attorney asked ChatGPT to confirm the cases were real. ChatGPT reaffirmed them and produced more fabricated supporting detail. The attorney filed that too.

The consequence: The firm was sanctioned $5,000. Judges whose names had been falsely attributed to opinions they never wrote had to be personally notified by the attorneys.

Why it matters for LLM09: The model did not hedge. When asked to verify, it doubled down with more fabrication. This is the overreliance loop in its purest form: model fabricates → user asks model to confirm → model confirms fabrication → user acts → harm.
The Defense That Would Have Stopped This

One citation check: searching the case names in Westlaw, LexisNexis, CourtListener, or PACER before filing. The URLs would not have resolved. Zero automation required — just the verification step the attorney skipped.

← BackNext → Package Hallucination
📄 View sources for this lesson