Fake ChatGPT case citations

Grounded in the Mata v. Avianca record. Updated July 18, 2026.

The modern era of AI-hallucination sanctions began on May 4, 2023, when Judge P. Kevin Castel of the Southern District of New York issued an order to show cause in Mata v. Avianca, Inc. Attorney Steven Schwartz had used ChatGPT to research airline-liability precedent and pasted the chatbot’s output — including six invented case citations — into an opposition brief without opening a single underlying opinion. On June 22, 2023, the court imposed $5,000 in joint-and-several Rule 11 sanctions against Schwartz, his colleague Peter LoDuca, and their firm, and ordered notice letters sent to each of the real federal judges whose names ChatGPT had falsely attached to the nonexistent opinions.

The six ChatGPT fakes from Mata v. Avianca

• Varghese v. China Southern Airlines Co., Ltd.

  Cited as: 925 F.3d 1339 (11th Cir. 2019)

• Shaboon v. Egyptair

  Cited as: 2013 IL App (1st) 111279-U

• Petersen v. Iran Air

• Martinez v. Delta Airlines

• Estate of Durden v. KLM Royal Dutch Airlines

• Miller v. United Airlines

Why ChatGPT produces convincing fakes

Large language models are trained to produce text that pattern-matches to the surface texture of real legal writing. A citation like “925 F.3d 1339 (11th Cir. 2019)” is not a fact the model has to look up — it is a shape the model has learned to produce. Faced with a research question, the model will generate a plausible-sounding case name, a real-looking reporter volume-and-page, a plausible court-and-year parenthetical, and, if asked to elaborate, will produce fabricated excerpts of the nonexistent opinion that read like real appellate prose. Every step of that generation is optimizing for surface plausibility, not for truth.

This is why post-Mata sanctions rulings uniformly reject the “the AI told me it was real” defense: verifying a citation by asking the same model that generated it will produce another confident confirmation of a nonexistent case. The only reliable check is a primary source.

The pattern past Mata

The Charlotin dataset (linked below) now catalogs over 200 court rulings identifying AI-generated citations as fabricated, spanning federal district and appellate courts, multiple state supreme courts, the Supreme Court of India, and Commonwealth jurisdictions. ChatGPT accounts for a plurality of the identified fabrications where the AI system was named; Gemini, Copilot, and Claude also appear in the record, as do generic “AI-assisted research” disclosures where the specific tool was not identified.

Verify a citation now

Related resources

• How to verify a case citation is real
• List of AI-hallucinated cases
• AI sanctions tracker
• Full sanctions database

Written by the Citation Safe Research Desk · Reviewed by Andy Gaber, Founder