ChatGPT’s Fake Cases Went to Court

A citation leads to somebody else’s case

The case name was Varghese v. China Southern Airlines Co., Ltd.^C67 Its citation supplied a volume, a page, a court and a year: 925 F.3d 1339, Eleventh Circuit, 2019.^C67 Peter LoDuca put it before a federal court in New York on March 1, 2023, in a filing researched and written by his colleague Steven A. Schwartz.^C03C06C67

Follow the publication reference and there is something to read. Page 1339 in that volume falls within J.D. v. Azar, a decision from a different federal appeals court.^C87 The address exists; the case named on the filing does not.^C11C26C67C87 Judge P. Kevin Castel would find that ChatGPT had fabricated the authorities Schwartz supplied.^C11C32

A citation is a set of directions to a source. Writing plausible directions and locating what they name are different tasks. Here, the reference reached a real page without reaching the promised decision.^C67C87

What had the lawyers actually checked before those citations reached the judge?

Schwartz needs an answer to a deadline

The client was Roberto Mata, who said a serving cart injured his knee during an overnight Avianca flight from El Salvador to New York in August 2019.^C01C02 He refiled his lawsuit in February 2022, and Avianca moved it into federal court.^C01C03 The airline then asked for dismissal: the Montreal Convention, the treaty governing the claim, allowed two years to sue.^C05C36

Mata’s side argued that the clock had paused during Avianca’s bankruptcy, begun in May 2020, while lawsuits against the airline were on hold.^C86 Schwartz’s first instruction to ChatGPT asked it to argue for that pause; subsequent requests asked for cases and particular holdings.^C17C18

Schwartz did the legal work, but he could not practise in this federal court, so LoDuca appeared as the lawyer responsible for the case there.^C04 By Schwartz’s account, his firm’s Fastcase research account offered limited access to federal decisions when he began this research, and the firm had no Westlaw or LexisNexis subscription.^C15

He had never used ChatGPT before, he said, and had heard about it through news coverage and his family.^C16 At the later hearing, he described his mistaken picture of the tool as “like a super search engine”.^C14 Castel’s account of what happened when Schwartz requested more authorities was shorter: “the chatbot complied by making them up.”^C18

The missing decisions acquire more text

Avianca’s lawyers went looking for the authorities.^C08 Their March 15 reply, signed by Bartholomew J. Banino, reported that they were “unable to locate most of the case law cited”.^C08C42 Mata’s side did not withdraw the document or explain the missing cases, and the court’s own search also failed to locate several of them.^C10

Castel asked to see the decisions themselves, warning that failure to supply them would mean dismissal.^C19 LoDuca sought more time, saying he was on vacation—a statement Castel later found knowingly false.^C20 What eventually arrived was an affidavit with supposed decision excerpts, drafted and assembled by Schwartz.^C21

The added writing brought its own clues. The purported Petersen v. Iran Air decision cited itself as an earlier authority.^C25 Castel called the supposed Varghese opinion’s legal analysis “gibberish”.^C23 Even its court docket number, the identifier assigned to a proceeding, belonged to a different case.^C48

On May 4, Castel ordered LoDuca to explain why citing nonexistent decisions should not lead to a court penalty.^C27 The record was about to gain a different kind of exhibit: the questions Schwartz had asked on his phone.^C28

The chatbot says it checked

Schwartz’s May 25 filing attached smartphone screenshots and brought the first acknowledgment from the lawyers or their firm that ChatGPT had generated the bogus authorities.^C28 That is the filing date, not an established date for the conversation: Schwartz’s accounts differed, and the court called them “shifting and contradictory explanations”.^C29

In the screenshots, Schwartz asked: “Is varghese a real case”.^C28

ChatGPT replied that it “is a real case”.^C28

He asked: “Are the other cases you provided fake”.^C28

“No, the other cases I provided are real and can be found in reputable legal databases such as LexisNexis and Westlaw.”^C28

Another reply used the words “Upon double-checking” and said Varghese “does indeed exist”.^C66

The chatbot’s double-check produced another assurance, not a court decision.^C66C26C11 It also directed him to the services his firm did not subscribe to, according to his testimony.^C28C15 The lawyers ultimately acknowledged six invented decisions; the Eleventh Circuit’s clerk separately confirmed that Varghese was not an authentic ruling of that court.^C22C26

OpenAI’s launch announcement had said the conversational format could let ChatGPT “admit its mistakes”, alongside a warning about plausible but incorrect answers.^C64C65 Its March 2023 FAQ described ChatGPT as disconnected from the internet and capable of incorrect answers.^C58

OpenAI explains its language models as learning patterns in how words occur together, then predicting the words that follow.^C50 A citation has a recurring form; the decision it names still needs an independent existence check. The Varghese reference shows why: its publication details looked usable while pointing into somebody else’s case.^C67C87

One lawyer supplied excerpts; another reviewed the writing

At the June 8 hearing, Schwartz testified that he had entered the Varghese citation but could not locate the decision.^C13 He described imagining an unpublished opinion, an appeal or a difficulty getting access.^C12

“I just never thought it could be made up.”^C12

The court could not settle whether that failed lookup came before the March filing or the April affidavit.^C13 Either way, Castel found, Schwartz knew before a submission that the complete text could not be found and did not disclose it.^C13

The earlier preparation and review are clearer. Schwartz wrote from the chatbot’s excerpts without possessing a complete copy of any of its generated decisions.^C46 LoDuca testified that he had been “looking for a flow”; he checked the writing without reviewing any of the authorities it cited.^C07 The court recorded his reliance on a colleague of more than twenty-five years whose work he believed would be reliable.^C40

That answers what the lawyers checked before the citations reached the judge: the language had been reviewed, but the decisions behind it had not.^C46C07 In this account’s reading of the record, the review before signing was a chance to require those decisions, rather than accept the excerpts alone.^C46C07C77

The signature carried a duty of its own. Castel quoted Rule 11: the signer certifies a reasonable inquiry into legal positions, supported by current law or a serious proposal for changing it.^C77 Having a colleague prepare the document did not remove that responsibility.^C04C77

The lawyers’ conduct and the client’s case receive separate rulings

Both lawyers testified under oath and expressed remorse at the June hearing.^C31 On June 22, 2023, Castel found bad faith through “conscious avoidance and false and misleading statements to the Court”.^C32 His explanation of conscious avoidance required awareness of a high probability that something was true and a deliberate refusal to confirm it; a foolish mistake alone was insufficient.^C68

A shared financial penalty applied to the lawyers and their firm.^C33 The lawyers also had to write to their client and the judges falsely named as authors.^C34 Mata’s case was decided separately: Castel rejected the argument that bankruptcy had paused the treaty’s deadline.^C36C86

The dismissal rested on the filing deadline, not the invented decisions.^C36C86 Judgment for Avianca followed on July 7, 2023; the source review for this account could not independently establish whether a later appeal existed.^C43C37

Castel also recognized a legitimate use for reliable AI assistance while maintaining lawyers’ obligation to ensure accurate filings.^C35 His account of the broader harms included wasted court and opposing-party resources, and the possibility that a client would lose arguments grounded in authentic decisions.^C49

Later records contain more examples, but they do not establish what share of court filings contains these errors.^C53C82

Follow the reference before sending the draft

A citation beside an answer in a chatbot app on your phone invites the same check. The product has changed since Schwartz used it: OpenAI announced search without signup in supported regions in February 2025.^C62 Its help page saved on October 12, 2026 still warned about nonexistent references and advised treating ChatGPT’s work as a draft requiring checks.^C63

If you are preparing coursework, representing yourself in court or circulating a report, try this on your next draft this week. Retrieve each cited document, compare its title and identifier with the reference, and read the passage your argument depends on. Leave the document beside the draft while you do it. If you cannot find a source, resolve that gap before relying on it.

An unsuccessful search was already part of Schwartz’s account, as were the explanations he offered himself for it.^C13C12 Asking the chatbot for reassurance supplied no missing opinion.^C28C66C22 Avianca’s lawyers instead reported the authorities they could not locate.^C08C42

You can inspect the mismatch that began this story. The name attached to the reference was Varghese; the decision at volume 925, page 1339 is J.D. v. Azar.^C67C87