Incident

Incidents · Mata v. Avianca

Lawyers were sanctioned in Mata v. Avianca for ChatGPT case citations

This is the case people mean when they talk about a chatbot inventing authorities. The citations were judicial opinions, not journal articles. The sanction is real. A scholarly reference checker would have been the wrong instrument.

Paste references and check themReferences

Submitting opens RefCheck and starts a check. 3 free reference checks without an account, plus 5 more free after you sign in. The text is placed in the web address; paste into the full checker if you would rather not do that. One reference per line. For a long list, use the full checker so the text is not put in the address.

What the order says

On 22 June 2023, Judge P. Kevin Castel of the Southern District of New York filed an opinion and order on sanctions in Mata v. Avianca, Inc., No. 22-cv-1461, later reported at 678 F. Supp. 3d 443. The order says Peter LoDuca, Steven A. Schwartz, and the law firm Levidow, Levidow & Oberman P.C. submitted nonexistent judicial opinions with fake quotes and citations created by ChatGPT, and then stood by those opinions after the court asked about them.

The court found subjective bad faith and, under Rule 11 or alternatively its inherent authority, imposed a $5,000 penalty jointly and severally, to be paid into the court registry within 14 days. It also ordered the lawyers to send the opinion, the 8 June 2023 hearing transcript, and the 25 April affirmation to the plaintiff, Roberto Mata, and to each judge falsely named as the author of the opinions called Varghese, Shaboon, Petersen, Martinez, Durden, and Miller. The order says there is nothing inherently improper about using a reliable AI tool, and that attorneys still have to check what they file.

Why this page is careful about RefCheck

RefCheck searches scholarly databases: Crossref, OpenAlex, Semantic Scholar, PubMed, Europe PMC, arXiv, DBLP, DataCite, and book indexes. It does not search Westlaw, Lexis, or any court’s opinions. Pasting “Varghese v. China Southern” into it will not tell you whether that case exists. Case law has its own reporters and citators. Use those.

The transferable lesson is narrower. A language model can produce a citation that has the surface form of a real authority: a caption, a reporter, a year, a quotation. The person who files or publishes it has to resolve that citation in the database that actually holds the source type. For a journal article, that is a bibliographic database. For a case, it is not RefCheck.

What not to generalize

The order is about these lawyers and this filing. It is not a finding about every later use of ChatGPT, and it is not a finding that a particular scholarly paper is fabricated. Later incidents on this site, including the MAHA report and the Deloitte Australia report, are different documents with their own sources.

If your document cites papers rather than cases, start with how to check a citation. If it cites cases, this tool will not help, and the order itself is the thing to read.

Typical cases

A caption that sounds like a real case

The order names six opinions that were not real: Varghese, Shaboon, Petersen, Martinez, Durden, and Miller. The form of a case citation was intact. The opinions were not.

Standing by the citation after a question

The order’s bad-faith finding is about what happened after the court asked, not only about the first filing. Checking after a challenge is later than checking before you file.

Using RefCheck on the brief

It would be the wrong database. A miss in Crossref does not prove a case is fake, and a hit would not be expected for a judicial opinion.

Sources

  1. Mata v. Avianca, Inc., opinion and order, 22 June 2023 — Judge P. Kevin Castel’s sanctions order in the Southern District of New York, docket 22-cv-1461, document 54.
  2. CourtListener docket, 1:22-cv-01461 — Docket entry for the 22 June 2023 opinion and order on sanctions, including the $5,000 penalty and the six named opinions.

This page only states what these sources say. It does not say RefCheck was used on the document, and a later correction is not summarized here unless one of these sources describes it.

How it works

Three steps for this kind of reference

01Paste

One reference per line, or a numbered list. Keep the authors, title, year, and the DOI or arXiv id if you have them.

02Cross-check

Scholarly references can be checked here. Judicial opinions cannot. Mata is the second kind.

03Read the verdict

Real, real with errors, likely fabricated, or unconfirmed. A corrected citation is shown when the record disagrees, and a retraction is flagged when one is on record.

Related pages

Frequently asked questions

Who was sanctioned, and for how much?

Judge Castel’s 22 June 2023 order imposed a $5,000 penalty jointly and severally on Peter LoDuca, Steven A. Schwartz, and Levidow, Levidow & Oberman P.C.

Which fake opinions does the order name?

The order tells the lawyers to write to the judges falsely identified as the authors of Varghese, Shaboon, Petersen, Martinez, Durden, and Miller.

Can RefCheck check case citations?

No. It does not search court opinions. A case citation has to be resolved in a legal database. RefCheck is for scholarly references.

Did the court ban AI?

No. The order says using a reliable AI tool is not inherently improper, and that lawyers remain responsible for checking what they file.

What does a check cost?

3 free reference checks without an account, plus 5 more free after you sign in. Then each reference costs 1 AILatest credit. Failed lookups are refunded. Credits last 1 year and the expiry extends on every top-up. Packs are $1.99 for 200, $4.99 for 600, and $9.99 for 1,500. There is no subscription.

Cite only what's real.

3 free reference checks without an account, plus 5 more free after you sign in. Then 1 credit each.