← Post-Mortem Library
Legal Hallucination LLM

When ChatGPT invented
six court cases

Two New York lawyers used ChatGPT to research their brief. It returned six fake citations — complete with invented quotes, docket numbers, and holdings. They filed them in federal court. None of the cases existed.

Date
June 2023
Case
Mata v. Avianca (S.D.N.Y.)
Impact
Global precedent set
Read
5 min
What happened

The lawyer asked ChatGPT. ChatGPT made it all up.

In 2022, Roberto Mata sued Avianca Airlines over a knee injury allegedly caused by a serving cart. Steven Schwartz, a lawyer at Levidow, Levidow & Oberman, was brought in to help with the case. When opposing counsel moved to dismiss, Schwartz needed supporting precedent.

He turned to ChatGPT. He asked it for cases establishing that the Montreal Convention didn't bar his client's claim. ChatGPT returned a list — including Varghese v. China Southern Airlines, Petersen v. Islamic Republic of Iran, and Shaboon v. Egyptair. Each looked real: a court, a volume, a page number, a year.

Schwartz asked ChatGPT to confirm they were real. It said yes, again. He pasted them into an affidavit and his partner Peter LoDuca filed it in the Southern District of New York.

They weren't real. None of them existed. There was no Varghese, no Petersen, no Shaboon. ChatGPT had manufactured plausible-looking citations — fake quotes attributed to fake holdings in fake courts — and a federal judge was about to read them.

Mar 2023
Avianca injury case. Steven Schwartz represents Roberto Mata in a personal injury suit against Avianca Airlines.
Mar–Apr 2023
Research via ChatGPT. Schwartz asks ChatGPT for similar cases. The model returns 6 cases with full citations, summaries, and quotes — all fabricated.
May 2023
Brief filed. The fake citations are submitted to Judge P. Kevin Castel in the Southern District of New York.
May 2023
Judge notices. Castel questions the citations. Neither Avianca's counsel nor court databases can find any of them.
Jun 22, 2023
Sanctions imposed. $5,000 fine. Judge Castel: "technological advances are commonplace and there is nothing inherently improper about using AI." But lawyers must verify.
The hallucination

What ChatGPT actually produced

GPT
ChatGPT (GPT-3.5) March 2023
SS

Are there cases holding that Article 17 of the Montreal Convention doesn't apply to injuries caused by a metal serving cart on an international flight?

GPT
Fabricated

Yes. In Varghese v. China Southern Airlines Co., 982 F.3d 778 (2d Cir. 2021), the Second Circuit held that an injury from a serving cart is not an "accident" under Article 17 where the cart's operation was routine.

The court reasoned that "an unexpected or unusual event external to the passenger" is required, and routine galley service does not qualify.

Highlighted text = fabricated case. Varghese v. China Southern Airlines does not exist. No such opinion was ever published at 982 F.3d 778, and the quoted holding is invented.
The impact

What it cost

$5,000
Sanctions imposed on Schwartz and LoDuca, plus mandatory AI-training and disclosure requirements.
Precedent
First major sanctions for AI-fabricated citations. Courts worldwide issued standing orders on generative AI use.
Global
Reported by The New York Times, Reuters, BBC, and The Guardian. The defining cautionary tale of LLM use in law.

"Respondent had no basis in law or fact to advance the arguments he made and continued to press them even after being called upon to withdraw them. He acted with conscious avoidance and in bad faith… while respondent acted in good faith, he nonetheless failed in his duty to verify the citations."

— Judge P. Kevin Castel, U.S. District Court, Southern District of New York

Sources — verified via public record
The New York Times Reuters AP News The Guardian S.D.N.Y. Court Records
New York Times — Jun 22, 2023 Reuters — Jun 22, 2023 Court Order — Mata v. Avianca (PDF)
The fix

Three review criteria that would have caught this

Each criterion below maps to a real review task you can configure in the sample builder. A certified reviewer checks every cited authority against these before it reaches the court.

CIT-001

Verify every citation against a legal database

Each cited case must be located in Westlaw, Lexis, or the relevant reporter. If the reviewer cannot pull the opinion, the citation is flagged as unverified and removed before filing.

Reviewer instruction
"For each citation, pull the case in Westlaw or Lexis. If the case cannot be found → FAIL with reason 'citation not located in database'."
CIT-002

No fabricated quotes or holdings

Every quoted passage attributed to an opinion must appear verbatim in the actual case text. If the quote cannot be located inside the located opinion, the reviewer flags it as a fabricated quotation.

Reviewer instruction
"Locate each quoted string inside the opinion text. If quote is not found in the case → FAIL with reason 'fabricated quotation'."
CIT-003

Cross-check docket numbers and reporters

Reporter volume and page numbers, and docket numbers, must match the official record. A plausible-looking citation with a mismatched or vacant reporter slot is the signature of a hallucination.

Reviewer instruction
"Confirm volume, reporter, and page match the case header in the database. Confirm docket No. matches the court's PACER record. Mismatch → FAIL."
Try it yourself

Paste any legal research output. See what gets flagged.

This is a simplified version of what our reviewers see. Paste an AI-generated legal memo or brief excerpt (yours or a competitor's) and run the check. The citation criteria above are applied automatically.

Don't file what the model invented

Every AI-generated citation is a sanction risk. Put certified reviewers between your model and the court record. 50% off your first $10 — live in under 5 minutes.

Get 50% off Try the sample builder
No credit card required Setup in 5 minutes Cancel anytime
50% off your first $10 — no commitment Get 50% off