On 22 June 2023, a federal judge in New York imposed sanctions on two lawyers who had submitted a legal brief containing six entirely fictional case citations — all generated by ChatGPT. The case, Mata v. Avianca, became the first major judicial reckoning with AI hallucinations in legal practice. What it established still governs the standard of professional conduct for every lawyer who uses generative AI anywhere in the world.
THE CASE IN BRIEF
Roberto Mata sued Avianca Airlines in the United States District Court for the Southern District of New York, alleging that he had been injured by a metal serving cart on a flight. The personal injury claim was, in itself, unremarkable. What happened next was not.
In March 2023, Mata’s attorney Peter LoDuca filed an opposition to Avianca’s motion to dismiss. The brief cited, among other cases, Varghese v. China Southern Airlines Co., Shaboon v. Egyptair, Petersen v. Iran Air, Miller v. United Airlines, Estate of Durden v. KLM Royal Dutch Airlines, and Zicherman v. Korean Air Lines Co. None of the first five cases existed. The sixth existed but the cited passage did not appear in it.
When Avianca’s counsel searched for these cases, they found nothing. They notified the court. Judge P. Kevin Castel ordered LoDuca to produce copies of the cited decisions. LoDuca turned to his colleague Steven Schwartz, who had actually prepared the brief using ChatGPT. Schwartz submitted an affidavit acknowledging that he had used ChatGPT and that he had relied on the AI’s representations without independently verifying them. He had, in fact, asked ChatGPT directly whether the cases were real, and ChatGPT had confirmed that they were. He accepted this confirmation without checking Westlaw, Lexis, or any other legal database.
WHAT CHATGPT ACTUALLY SAID
One of the most striking aspects of the Mata record is the ChatGPT conversation that Schwartz submitted to the court as an exhibit. When Schwartz asked the AI whether the cases it had generated were real and could be found in legal databases, ChatGPT responded:
“Yes, the cases I provided are real and can be found in reputable legal databases such as LexisNexis and Westlaw.”
This exchange is critical to understanding what the court was confronted with. Schwartz had not simply copied output without thinking. He had specifically asked the AI to verify its own output — and the AI had confirmed false information with complete confidence. The problem, the court found, was not that Schwartz had used AI. The problem was that he had treated the AI’s self-verification as sufficient, without any independent check against an actual legal database.
Judge Castel’s opinion dissected this failure in terms that have since become the standard framework for evaluating AI-assisted legal work. The court found that Schwartz’s conduct was not excused by ignorance of the technology’s limitations, because those limitations were both inherent in the technology and widely reported. An attorney who relies on a source he has not verified against authoritative databases is not exercising professional judgment — he is delegating judgment to a machine incapable of exercising it.
THE COURT’S FINDINGS AND SANCTIONS
Judge Castel’s 34-page opinion, dated 22 June 2023, made the following core findings.
First, the court found that Schwartz’s conduct constituted a violation of Federal Rule of Civil Procedure 11(b)(2), which requires an attorney certifying a filing to represent that the legal contentions therein are warranted by existing law. By submitting citations to non-existent cases, Schwartz had violated this certification.
Second, the court found that LoDuca, who had signed the original brief, bore independent responsibility because he had submitted the brief without verifying the citations that a colleague had prepared.
Third, and most instructively, the court examined the conduct after the false citations were flagged. When Avianca raised the issue, the response from the firm was not to acknowledge error but to double down: a subsequent submission still failed to clarify the position clearly, and the affidavit submitted to the court was itself found to contain inaccuracies about what had occurred.
The sanction imposed was $5,000 — modest in dollar terms, but accompanied by something far more consequential. The court ordered that copies of the sanctions order be sent to the judges whose names had appeared in the fictitious case citations. The reputational dimension of this was significant: judicial names had been attached, without authorisation, to opinions that did not exist.
Additionally, the court required both Schwartz and LoDuca to complete continuing legal education on the ethical use of artificial intelligence, and Levidow, Levidow & Oberman was also named as a sanctioned party.
THE PRINCIPLE MATA ESTABLISHED
The legal principle that emerges from Mata can be stated plainly: an attorney who uses generative AI to identify legal authority retains personal, non-delegable responsibility for verifying that authority against a reliable database before submitting it to any court or tribunal. The AI’s own confirmation that a case exists is not verification. It is the source of the problem.
Judge Castel articulated this with precision. Generative AI systems, including ChatGPT, are large language models that predict statistically probable text. They do not retrieve cases from a database. They generate plausible-sounding text based on patterns in their training data. A case name, a court, a year, a judge, a holding — all of these are things a language model can generate convincingly, because it has seen thousands of real examples. That does not make the output true. The model does not know whether the cases it generates exist. It produces what is probable in the statistical sense, not what is accurate in the legal sense.
This distinction — between statistical probability and legal accuracy — is the foundation of every subsequent judicial decision on AI hallucinations. Mata did not merely sanction two lawyers. It defined the verification obligation that governs AI-assisted legal research globally.
THE AFTERMATH: COURT ORDERS ACROSS THE US
Within months of the Mata decision, federal courts across the United States began issuing standing orders on the use of AI in legal filings. The Northern District of Texas, the Eastern District of Louisiana, and several other districts issued orders requiring attorneys to certify that any AI-generated content in filings had been reviewed and verified by a licensed attorney. Some courts went further, requiring disclosure of AI use in filings as a categorical matter.
The Mata decision was also the trigger for a broader professional responsibility conversation in the United States. State bars and bar associations began issuing guidance on the ethical obligations of attorneys using AI. The consensus that emerged — and that has since been reinforced in subsequent decisions — is that the duty of competence under professional responsibility rules extends to understanding the tools an attorney uses, including their limitations.
The trajectory from Mata has been one of escalation. In 2024 and 2025, US courts imposed progressively larger sanctions on attorneys who repeated the same error. By 2025, the Oregon district court had imposed sanctions of $110,000 on a firm for AI hallucination — more than twenty times the Mata penalty, reflecting the court’s view that the profession had been warned and had not responded adequately. That case is examined separately on this blog.
THE INDIA PARALLEL: A FASTER TRAJECTORY
When Mata was decided, the immediate reaction in India was one of distant observation — this was an American problem, concerning American courts, American bar rules, and American attorneys. That perception has since been overtaken by events.
Between late 2024 and early 2026, Indian courts and tribunals confronted AI hallucinations at every level of the judicial hierarchy. The Income Tax Appellate Tribunal at Bangalore withdrew an order that cited non-existent authority (Buckeye Trust v. PCIT). The Bombay High Court quashed an income tax assessment of Rs. 27.91 crore because the Assessing Officer had cited three fake judgments (KMG Wires v. ITO). The Delhi High Court set aside a GST Show Cause Notice because it relied on fabricated citations, and the court physically retrieved books from its library to confirm the citations did not exist (Jeetmal Choraria v. UoI). The Bombay High Court imposed adverse costs on a pro se litigant who cited a non-existent case (Deepak Bahry v. Heart & Soul). And most significantly, the Supreme Court of India declared, in Gummadi Usha Rani v. Sure MR, that AI hallucination by a judicial officer constitutes misconduct — and issued notice to the Attorney General, Solicitor General, and Bar Council of India to develop national guidelines.
India’s trajectory has been compressed relative to the US experience. What took the United States from 2023 to 2025 — from the first significant sanctions case to the development of a professional consensus — India has largely traversed in approximately eighteen months. And crucially, India’s response has the potential to be more systematic: where the US response has been court-by-court and order-by-order, India’s Supreme Court intervention in Gummadi creates the conditions for a binding national framework applicable to all courts, tribunals, and quasi-judicial authorities simultaneously.
WHAT THE MATA STANDARD MEANS FOR INDIAN PRACTITIONERS
The Mata standard — that an attorney bears personal, non-delegable responsibility for verifying AI-generated citations — is now the operative standard in India as well, by virtue of the Indian decisions that have followed the same analytical path. The specific rules differ (Federal Rule 11 does not apply in Indian proceedings), but the underlying professional obligation is identical: a practitioner who submits a citation to any court, tribunal, or quasi-judicial authority is certifying, implicitly, that the citation is accurate. Generative AI does not discharge that obligation. It creates it.
The practical consequences for Indian practitioners are specific and immediate. Any citation in any submission — whether in a writ petition before a High Court, a memo of appeal before the ITAT, a reply to a GST Show Cause Notice, or written submissions in an arbitration — must be independently verified against SCC Online, Manupatra, Indian Kanoon, or an equivalent authoritative database before submission. This is not a good practice recommendation. It is, following KMG Wires, Jeetmal Choraria, and Gummadi, a professional obligation with disciplinary and costs consequences for non-compliance.
The additional lesson from Mata that Indian practitioners should internalise is the danger of the AI’s own self-verification. Schwartz’s error was not simply to use ChatGPT to find cases. His error was to ask ChatGPT whether the cases were real and to accept the affirmative answer. This is a failure mode that any practitioner using AI research tools must guard against. A generative AI system that produces a hallucinated citation will, if asked to confirm it, typically confirm it — for the same statistical reasons that caused the hallucination in the first place. The AI has no access to reality to check against. The only verification that counts is verification against an authoritative source that actually contains the case.
THE LEGACY OF MATA
Mata v. Avianca is not primarily a case about personal injury or airline liability. It is a case about the limits of delegated professional judgment in the age of generative AI. Its legacy is threefold.
First, it established that the introduction of a false citation, regardless of the reason for its introduction, is a violation of a lawyer’s duty to the court. The fact that AI was the source of the error does not create a new category of excusable mistake. It is misconduct regardless of the technology that produced it.
Second, it established that the duty to verify cannot be satisfied by asking the AI itself. Verification means checking the citation against a source that is independent of the AI and that actually contains the cited material.
Third, and most enduringly, it established that ignorance of AI limitations is not a defence. As of June 2023, every lawyer in every jurisdiction who uses generative AI for legal research operates with notice of the risk. From that date forward, reliance on unverified AI output is not a mistake. It is a choice to take a risk that a court has explicitly declared unacceptable.
In India, where the Supreme Court has now issued notice in Gummadi and where national guidelines are anticipated, the Mata standard is not merely persuasive. It is the foundation on which the emerging Indian framework is being built. Every Indian practitioner who wants to understand where AI law is going in India needs to understand where it started in the United States. It started with six fake cases, a confident AI, and a federal judge who was not prepared to let the profession treat hallucinations as harmless errors.
CASE DETAILS
| Case | Mata v. Avianca, Inc. |
| Citation | No. 1:22-cv-01461 (PKC), 2023 WL 4114965 |
| Court | United States District Court, S.D.N.Y. |
| Judge | Judge P. Kevin Castel |
| Decision Date | 22 June 2023 |
| AI Tool Used | ChatGPT (OpenAI) |
| Number of Fake Cases | 6 (five wholly fictitious; one existing but misquoted) |
| Sanction | $5,000 + notice to cited judges + mandatory CLE |
| Category | AI Hallucinations & Sanctions |
| Related India Cases | KMG Wires v. ITO (Bombay HC 2025); Jeetmal Choraria v. UoI (Delhi HC 2025); Gummadi Usha Rani v. Sure MR (SC 2026) |