Lawyers, ChatGPT and the cases that never existed

On 15 March 2023 the lawyers for Avianca told a federal court in New York that they could not find most of the decisions the other side had cited (Mata v. Avianca, Inc., 2023, findings ¶ 7). The first admission that ChatGPT had invented them came on 25 May, 71 days later (my arithmetic from two dates in the opinion) (Mata v. Avianca, Inc., 2023, findings ¶ 44). If you sign papers that someone else or an AI tool has drafted, those 71 days are the part of Mata v. Avianca most worth your time.

What did the court say about using ChatGPT to prepare a filing?

That it is not wrong in itself, provided the tool is reliable, but that existing rules give lawyers a gatekeeping role over the accuracy of what they file (Mata v. Avianca, Inc., 2023, p. 1).

The two lawyers, one who signed the filing and one who drafted it, and their firm, Levidow, Levidow & Oberman P.C., submitted non-existent judicial opinions with fake quotes and citations created by ChatGPT, then stood by them after judicial orders called their existence into question (Mata v. Avianca, Inc., 2023, p. 1).

What was filed, and how were the fakes found?

On 1 March 2023 the signing lawyer filed an opposition to Avianca's motion to dismiss in a suit by Roberto Mata, who said a metal serving cart had struck his knee on a flight to JFK airport. It cited and quoted decisions said to be published in the Federal Reporter, the Federal Supplement and Westlaw (Mata v. Avianca, Inc., 2023, findings ¶ 1; Mata v. Avianca, Inc., 2023, findings ¶ 5). He signed it but did not write it. The drafting lawyer researched and wrote it, and the signing lawyer read it "looking for a flow", without reading any of the cases (Mata v. Avianca, Inc., 2023, findings ¶ 6). Once Avianca said it could not find them, the court searched too, could not locate several of them, ordered copies of nine decisions and received only excerpts (Mata v. Avianca, Inc., 2023, findings ¶ 9; Mata v. Avianca, Inc., 2023, findings ¶ 13; Mata v. Avianca, Inc., 2023, findings ¶ 14; Mata v. Avianca, Inc., 2023, findings ¶ 22).

The Clerk of the Eleventh Circuit confirmed that "Varghese" was not a ruling of that court, and the docket number printed on it belonged to a different case (Mata v. Avianca, Inc., 2023, findings ¶ 26; Mata v. Avianca, Inc., 2023, findings ¶ 28). Reading only its opening passages would have shown it to be internally inconsistent and nonsensical (Mata v. Avianca, Inc., 2023, conclusions ¶ 23).

What did ChatGPT say when it was asked?

When asked for case law, it supplied it by inventing it, a case of AI hallucination. The log of prompts attached to the record shows the drafting lawyer moving from a general question to requests such as "provide case law in support…" and "show me more cases"; the court wrote that the chatbot "complied by making them up" (Mata v. Avianca, Inc., 2023, findings ¶ 39). He also asked it directly "Is Varghese a real case" and "Are the other cases you provided fake", and was told they were real authorities that could be found through Westlaw, LexisNexis and the Federal Reporter (Mata v. Avianca, Inc., 2023, findings ¶ 45).

This is a record of ChatGPT in 2023; I say nothing about how any tool answers today. The screenshots of the "is it real" questions carry no date either: the drafting lawyer gave two accounts, one implying he asked before the filing, the other that he asked after the order of 4 May (Mata v. Avianca, Inc., 2023, findings ¶ 46; Mata v. Avianca, Inc., 2023, findings ¶ 47). The respondents later acknowledged that the cases did not exist (Mata v. Avianca, Inc., 2023, findings ¶ 36).

Was the finding of bad faith about the AI, or about what came after?

About what came after. The court found bad faith on the part of the two lawyers, based on acts of conscious avoidance and false and misleading statements to the court (Mata v. Avianca, Inc., 2023, p. 2). It wrote that had the respondents come clean after Avianca's brief of 15 March, or after the orders of 11 and 12 April, "the record now would look quite different"; instead they "doubled down" and began to tell the truth only from 25 May (Mata v. Avianca, Inc., 2023, p. 2).

For the drafting lawyer the court cited several things. He could not find "Varghese", cited it anyway, and did not say so in the affidavit of 25 April (Mata v. Avianca, Inc., 2023, findings ¶ 12; Mata v. Avianca, Inc., 2023, conclusions ¶ 24). He had called ChatGPT a "supplement" to his research, but at the hearing admitted that apart from an unsuccessful try on Fastcase, his firm's research service, he had done nothing else (Mata v. Avianca, Inc., 2023, findings ¶ 37; Mata v. Avianca, Inc., 2023, findings ¶ 41). "Poor and sloppy research would merely have been objectively unreasonable", the court wrote; bad faith lay in being aware of facts signalling a high probability that "Varghese" and "Zicherman" did not exist and consciously avoiding confirming it (Mata v. Avianca, Inc., 2023, conclusions ¶ 24).

For the signing lawyer the court cited his failure to read Avianca's reply before forwarding it, signing an affidavit he had never seen in draft, and seeking an extension on the ground that he was on holiday, which he knew was false (Mata v. Avianca, Inc., 2023, findings ¶ 10; Mata v. Avianca, Inc., 2023, findings ¶ 17; Mata v. Avianca, Inc., 2023, findings ¶ 21). It drew a line: an inadequate inquiry may merely be unreasonable, but no inquiry at all supports a finding of bad faith (Mata v. Avianca, Inc., 2023, conclusions ¶ 23).

What was the penalty, and what else did the order contain?

A penalty of USD 5,000, imposed jointly and severally on the two lawyers and the firm, payable into the court's registry within 14 days (Mata v. Avianca, Inc., 2023, p. 34). In the same 14 days they had to write to Mr Mata, and to each judge falsely named as the author of one of six fake opinions, enclosing the opinion, the hearing transcript and a copy of the 25 April filing, and file copies of the letters with the court (Mata v. Avianca, Inc., 2023, p. 34). The court framed the penalty as deterrence, not punishment (Mata v. Avianca, Inc., 2023, conclusions ¶ 34).

It also recorded what it chose not to do: it required no apology, because a compelled apology is not a sincere one (Mata v. Avianca, Inc., 2023, conclusions ¶ 33).

What does "checking" mean for the person who signs?

It means opening the original where it is published by an authoritative source and reading the passage you rely on; asking the tool again is not checking. The middle column below is what the opinion records; the right-hand column is my own inference for a company's documents.

The table sits on two rungs of the five-rung verification ladder, facts and human approval; that ladder is this site's grouping, not something the court set out.

On whose name is on the paper: signer and drafter were different people. The drafting lawyer was not admitted to practise in that district, so the signing lawyer entered the appearance while the drafting lawyer did all the substantive work, as the signing lawyer explained it (Mata v. Avianca, Inc., 2023, findings ¶ 2).

What should you do when you find an error in your own document?

Say so early, withdraw the faulty version and tell the recipient exactly what was wrong. That is my inference for business, since the court spoke only of lawyers.

Counsel for the drafting lawyer and the firm, filing their response to the show-cause orders, caught and corrected on their own the fact that three fake cases had gone into the table of authorities of the first version; the court called it an innocent mistake (Mata v. Avianca, Inc., 2023, p. 1, n. 1). For the filing of 1 March, no respondent ever wrote to withdraw it or to say the court should no longer rely on it (Mata v. Avianca, Inc., 2023, findings ¶ 51). A lawyer's duty runs past the moment of filing to "later advocating" the position, and any warning received is relevant to bad faith (Mata v. Avianca, Inc., 2023, conclusions ¶ 10; Mata v. Avianca, Inc., 2023, conclusions ¶ 16).

A hypothetical, with no real company or person. A sales director at an exporter sends a bank a market report containing a statistic an AI assistant drafted. The bank asks for the source, and the director cannot trace it. One route is a correction the same day; the other is a vague reply and a wait. In this opinion, silence and misleading statements after being challenged are what the court used to find bad faith.

Going deeper: what is this order, and what does it not say?

I read the "Opinion and Order on Sanctions" of 22 June 2023 from the US District Court for the Southern District of New York, No. 22-cv-1461, issued after a hearing on 8 June 2023 (Mata v. Avianca, Inc., 2023, p. 3); the library lists it as 678 F. Supp. 3d 443. I read the main text, pages 1–34. I did not read Appendices A and B (pages 35–43; my text copy has no text on those pages), the underlying filings the opinion refers to, or the hearing transcript; the lawyers' words I quote are quoted within the opinion. "Findings ¶ N" and "conclusions ¶ N" are the numbered paragraphs in each part, and "p. N" is the printed page number.

First limit: this is a sanctions opinion, not a ruling on AI in general. The court sanctioned under Rule 11 or, alternatively, its inherent power (Mata v. Avianca, Inc., 2023, conclusions ¶ 27), and the text I read does not decide the merits of Mr Mata's claim.

Second, the legal standard. Because the court acted on its own initiative, it needed a finding of subjective bad faith; where an adversary starts the process under Rule 11(c)(2), the lawyer has a 21-day safe harbour to correct or withdraw, and sanctions can still follow if the statement was objectively unreasonable (Mata v. Avianca, Inc., 2023, conclusions ¶ 14).

Third, the facts are the court's, including where it did not believe a witness (Mata v. Avianca, Inc., 2023, findings ¶ 47; Mata v. Avianca, Inc., 2023, p. 9, n. 6); the lawyers' own accounts may differ. The checklist and the hypothetical are mine, not the court's.