Court sanctions lawyers after fabricated ChatGPT citations

On June 22, 2023, Judge P. Kevin Castel of the Southern District of New York sanctioned attorneys Steven A. Schwartz and Peter LoDuca and their firm, Levidow, Levidow & Oberman P.C., in Mata v. Avianca. ChatGPT had generated nonexistent judicial opinions, complete with invented quotations and citations. The lawyers submitted them and continued to defend them after opposing counsel and the court questioned their existence. Castel imposed a joint $5,000 penalty and notification requirements, finding bad faith in each lawyer’s conduct.^[1]

How invented cases entered the record

Roberto Mata alleged that an airline employee injured his knee with a serving cart during an August 2019 flight from El Salvador to New York. Avianca sought dismissal because his February 2022 action was filed beyond the Montreal Convention’s two-year deadline. The dispute included whether Avianca’s bankruptcy had paused that period.^[2]

Schwartz did the substantive research and writing; LoDuca was the attorney of record because Schwartz was not admitted to that federal district. According to the sanctions order, Schwartz turned to ChatGPT after finding the firm’s research subscription insufficient for federal cases. He asked for decisions supporting the bankruptcy argument and received generated summaries and excerpts, without obtaining full authentic opinions. LoDuca signed and filed the opposition after reviewing its style, without checking the cited authorities (Findings 2, 5–6, 37–40).^[1]

The warnings and subsequent submissions were central to the court’s decision:

Date in 2023	What happened
March 1	LoDuca filed the opposition containing the fabricated authorities.
March 15	Avianca’s reply identified purported decisions its lawyers could not locate. The opposition was not withdrawn.
April 11–12	The court ordered copies of cited decisions after its own searches failed.
April 25	LoDuca filed an affidavit prepared by Schwartz, attaching supposed opinion excerpts and referring to an unnamed online database.
May 4 and May 25	The court ordered LoDuca to explain why he should not be sanctioned; the May 25 response first acknowledged ChatGPT’s fabricated cases.
June 8 and June 22	The court heard sworn testimony, then issued its findings and sanctions.

This sequence is documented in Findings 5–23 and 42–50.^[1]

The fabrications extended beyond incorrect case names. The supposed Varghese opinion shifted between incompatible accounts of the underlying dispute and bankruptcy. Its docket number and reporter citation belonged to unrelated proceedings. The supposed Petersen opinion cited itself as precedent. The lawyers ultimately acknowledged that six submitted opinions did not exist (Findings 24–36).^[1]

What the court found about responsibility

Schwartz testified, “I just never thought it could be made up.” The order also reproduces chatbot assurances that the cases were real. But the judge found Schwartz had already encountered evidence that the authorities could not be located, and found his changing explanations about checking ChatGPT and merely supplementing other research misleading. The chatbot’s reassurance was part of the failed verification process, rather than independent confirmation (Findings 11–12, 41, 43–47).^[1]

Castel assessed the lawyers separately. LoDuca signed without checking the authorities or the factual basis of the later affidavit, ignored warnings and falsely claimed that he himself was on vacation to obtain an extension. Schwartz concealed failed searches and continued supplying the suspect material. The court found subjective bad faith through conscious avoidance and false or misleading statements, beyond simply careless research. The firm was jointly responsible under Rule 11, the federal rule governing representations in court filings (Conclusions 21–27).^[1]

Reuters reported that the firm disputed the bad-faith finding, describing its conduct as a good-faith mistake about what the technology could invent. That was the firm’s response to the ruling, rather than a reversal of the court’s findings.^[3]

Penalty, notification and later filings

The $5,000 was one penalty payable to the court, jointly and severally owed by both lawyers and the firm—not $5,000 each or compensation to Mata. Within 14 days, they also had to mail Mata and each judge falsely named as an opinion’s author the sanctions order, hearing transcript and relevant April 25 materials, then file copies of the letters. The court did not compel an apology or award Avianca attorney fees; it considered the firm’s arranged training and the lawyers’ remorse when selecting deterrent measures (Conclusions 30–34 and final order).^[1]

On July 5, their counsel Tyler Maulsby filed copies of notification letters and reported that the required payment had been made. The filing includes a letter apologizing to Mata. It documents the respondents’ reported compliance; delivery and later training effectiveness are separate questions.^[4]

Mata’s injury action was dismissed in a separate June 22 opinion because the court found it untimely under the Montreal Convention. Dismissal was not the sanction for using ChatGPT.^[2] The sanctions order expressly allowed for reliable AI assistance while retaining attorneys’ responsibility for accurate filings.^[1] Hallucination describes the generated-content problem; the incidents guide considers the wider evidence question. Here the documented failure joins invented output to human decisions that let it survive repeated challenges to its authenticity.