What happened
Roberto Mata sued the airline Avianca: he said that on a flight from El Salvador to New York’s Kennedy Airport, a metal cart hit his left knee. Avianca asked the court to dismiss the case because the claim was filed after the deadline under the Montreal Convention.
To save the claim, the plaintiff needed case law holding that an airline’s bankruptcy pauses this deadline. Schwartz had no experience with either the Montreal Convention or bankruptcy, and at the hearing, according to the court, he admitted he had thought “F.3d” stood for “federal district, third department”.
Steven Schwartz prepared the case documents, but he was not admitted to practice in federal court. So the opposition was signed by Peter LoDuca, who did not check the references and relied on a colleague he had worked with for more than 25 years.
The firm Levidow, Levidow & Oberman had only Fastcase, with limited access to federal case law. Schwartz asked ChatGPT about federal court decisions in which the limitation period was paused because of an airline’s bankruptcy, and got excerpts from cases that do not exist.
On 1 March 2023, an opposition with these references went to the court. Avianca answered that it could not find most of the cases, and that the ones it found did not support the plaintiff’s conclusions.
On 11 and 12 April, the court ordered copies of nine decisions. On 25 April, the lawyers filed excerpts from a database they did not name, and they began to reveal the truth only on 25 May.
Schwartz told the court he had thought of ChatGPT as a kind of search engine and did not imagine it could make up cases. Judge Castel called the legal analysis in one of the made-up decisions, Varghese, gibberish.
A detail worth remembering: the made-up Varghese also contained references to real cases. But those cases did not support the conclusions given, so a check by a number or a name alone would have shown nothing.
The court’s ruling
On 22 June 2023, Judge P. Kevin Castel (Southern District of New York) fined Schwartz, LoDuca and their firm $5,000. The ground: a filing with references to six nonexistent decisions that ChatGPT produced.
The court found bad faith on the part of both lawyers. It pointed to conscious avoidance of checking and to false statements to the court made after the questions about the references.
Besides the fine, the lawyers had to send letters to each judge whom ChatGPT named as the author of a made-up decision. Each letter came with a copy of Castel’s ruling, the made-up decision attributed to that judge, and the affidavit of 25 April in which the lawyers kept citing these cases.
The judge wrote separately that if the lawyers had admitted the error right after Avianca’s objection, the case would have looked different. The sanction came not from ChatGPT itself but from the fact that the made-up cases were still defended.
When choosing the sanction, the court took into account the publicity and the lawyers’ remorse, their clean disciplinary record and the low chance of a repeat. The firm said it would comply with the ruling but disagreed with the finding of bad faith and saw what happened as a good-faith mistake.
Mata’s claim was dismissed as time-barred. The client lost the case, and the firm became an example that courts cite in later rulings on made-up references.
What changed in the courts
On 30 May 2023, before the Mata ruling, Judge Brantley Starr (Northern District of Texas) required all lawyers in his cases to file a certificate on the use of generative AI.
In the certificate, a lawyer confirms one of two things: the filing has no text prepared by generative AI, or a person checked such text against print reporters or traditional legal databases. The court strikes a filing without a certificate, and the lawyer is responsible for the content of a signed filing under Rule 11.
England and Wales answered with guidance for judges. Its October 2025 version says that representatives of the parties are responsible for the accuracy of their materials, and a judge may ask for confirmation that the lawyer personally checked references obtained with a chatbot.
The requirements did not stop the flow. As of 15 October 2026, a global database counts 2,190 cases of AI-made-up content, including 1,492 in the US, 232 in Canada, 112 in Australia and 76 in the UK.
Courts do not ban AI, though. They require checking and punish its absence, so Mata v. Avianca became a precedent on a lawyer’s duty to open each decision before signing, not a precedent about technology.
A lesson for lawyers in Russia
Russian courts reached the same conclusion three years later. In May 2026, the Arbitration Court of the West Siberian District fined a company $600 for references in a cassation appeal to nonexistent rulings created by a neural network (case No. A27–7831/2025).
The court stated that preparing an appeal with AI does not release a party from responsibility: the party to the case answers for the accuracy of the generated text (Part 2, Article 9 of the Russian Arbitration Procedure Code). A request by the new representative to disregard these references did not cancel the fine, just as late explanations did not help in Mata.
The Supreme Court of Russia added a general rule in ruling No. 301-ES21–23756. A party prepares a position made with software at its own risk and must check the result.
The Varghese trap repeated in Russia too. In a Vologda case, No. A13–14310/2025, the case numbers in the plaintiff’s references were real, but other disputes had been heard under them, so the check must be made against the text of the ruling, not the number.
What to take from Mata into practice:
1. A public chatbot is not a database of court rulings. Search case law in systems with the primary source, and treat an AI answer as a draft.
2. A person opens every reference before signing: the number, the parties, the court level, the exact quote.
3. Admit an error found after filing at once. In Mata, the sanction came from the attempt to defend made-up cases.
4. The one who checked signs. LoDuca relied on a colleague and was held responsible together with him.
AI still saves hours on search and drafts if it works on the firm’s database of rulings and shows the source document for each conclusion. How such a search works is covered on the page about AI case law search.