Mata v. Avianca: The Landmark AI Hallucination Case That Changed Legal Practice | Advocate Prakhar

Case at a Glance

Full Citation	Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), 2023 WL 4114965 (S.D.N.Y. June 22, 2023)
Court	U.S. District Court, Southern District of New York, Judge P. Kevin Castel
Date	June 22, 2023
Category	AI Hallucination & Rule 11 Sanctions
Jurisdiction	USA — New York (Federal)
AI Tool Used	ChatGPT (OpenAI)
Outcome	$5,000 sanctions; referral to grievance committee; public reprimand of three attorneys; attorneys ordered to send copy of order to clients

Background

Roberto Mata sued Avianca, Inc. for personal injury, claiming he was injured by a metal serving cart on a flight. When Avianca moved to dismiss based on a statute of limitations defence, attorney Steven Schwartz of Levidow, Levidow & Oberman filed an opposition brief. Schwartz had asked ChatGPT to help him research cases supporting the opposition, and ChatGPT generated six case citations — Varghese v. China Southern Airlines, Shaboon v. Egyptair, Petersen v. Iran Air, Zicherman v. Korean Air Lines, Martínez v. Delta Airlines, and Estate of Durden v. KLM Royal Dutch Airlines. None of these cases existed.

The AI Issue

When Avianca’s counsel pointed out they could not locate the cited cases, Schwartz doubled down. He filed letters purporting to attach copies of the cases — which were themselves fabricated documents. Only when Judge Castel directly ordered production of authenticated copies did the full truth emerge. Schwartz admitted he had used ChatGPT and had trusted its output without independent verification. He stated he was unaware ChatGPT could fabricate citations. The court found this explanation inadequate — ChatGPT had itself told Schwartz the cases were real when he asked it to confirm, demonstrating the circular unreliability of asking an AI to verify its own hallucinations.

What the Court Decided

• Filing non-existent cases violates Federal Rule of Civil Procedure 11, which requires attorneys to certify that citations are supported by existing law [Rule 11 certification obligation].
• The violation was aggravated by the attorneys’ failure to acknowledge the error when first challenged and by the filing of fraudulent confirmation letters [candour to tribunal — MRPC Rule 3.3].
• $5,000 in sanctions were imposed, payable to the court, along with a requirement to notify the client of the sanctions order [monetary sanction].
• Three attorneys — Schwartz, co-counsel LoDuca, and the firm principal — were sanctioned because all had signed Rule 11 certifications [supervisory attorney responsibility].
• The court referred the matter to the grievance committee for potential bar discipline [professional discipline referral].

“Many courts have grappled with the challenge of how to deal with lawyers who submit fictitious citations. The Court hopes that it will not have to revisit this difficult situation… It is the duty of every attorney to cite accurately and honestly to the Court.”

— Judge P. Kevin Castel, S.D.N.Y., June 22, 2023

The India Angle

Indian Law Equivalent

India does not have a direct equivalent of FRCP Rule 11, but equivalent obligations exist under Order VI Rule 15 and 15A CPC (verification of pleadings), the Advocates Act, 1961, and the court’s inherent powers under Section 151 CPC. The Supreme Court of India has consistently held that advocate officers of the court have a solemn duty not to mislead — a standard captured in cases like R.D. Saxena v. Balram Prasad Sharma (2000) 7 SCC 264. Contempt powers under the Contempt of Courts Act, 1971 would cover deliberate submission of fabricated citations.

Bar Council Rules

Under Bar Council of India Rules, Part VI, Chapter II, Section II, Rules 1 and 2, advocates must act with utmost good faith and not suppress or misrepresent the law to the court. Filing fabricated ChatGPT-generated citations as genuine case law would constitute a direct violation of these rules, potentially attracting proceedings before the State Bar Council Disciplinary Committee under Section 35 of the Advocates Act, 1961, with outcomes ranging from reprimand to removal from the rolls.

Practical Advice for Indian Advocates

• Treat Mata v. Avianca as the foundational warning: ChatGPT and similar AI tools fabricate case citations with confident-sounding authority — never file any AI-generated citation without cross-checking it on SCC Online, Manupatra, or Indian Kanoon.
• If you discover a citation error after filing, correct it immediately through a formal corrigendum or note to the court — delaying acknowledgment or attempting to cover the error with additional AI-generated content, as happened in Mata, dramatically increases sanction exposure.
• Build a firm-wide protocol requiring a senior advocate to personally verify every case cited in an AI-assisted brief before the document is signed and filed, maintaining a verification log as evidence of due diligence.

Quick Takeaways

• ChatGPT fabricated six complete case citations, including fake parties, reporters, and holdings — and then confirmed they were real when asked.
• Covering up an AI citation error by filing fabricated confirmation documents vastly multiplies sanctions exposure.
• Mata v. Avianca remains the global gold standard for understanding why AI-generated legal research must always be independently verified.

---

Deep Dive: How Mata v. Avianca Reshaped Global AI-in-Law Policy

Mata v. Avianca did not merely produce a $5,000 sanction — it triggered a global cascade of judicial, bar association, and legislative responses that continues to shape AI governance in legal practice three years later. Within weeks of Judge Castel’s June 2023 order, dozens of US district courts issued standing orders or local rules requiring disclosure of AI tool use in filed documents. The New York City Bar Association, the American Bar Association, and numerous state bars issued formal ethics guidance on AI use. International bar bodies from the UK’s Law Society to India’s Bar Council of India received formal questions about whether similar guidance was needed.

The case’s most lasting contribution may be its exposure of the “circular hallucination trap.” When Schwartz asked ChatGPT whether the cases he had found were real, the model confirmed they were — because it was predicting what a plausible answer to that question would look like, not actually checking any database. This demonstrated that AI models cannot verify their own hallucinations, establishing the foundational principle that human verification against authoritative legal databases is not optional but structurally necessary. Courts in the UK (Rodney, 2026), Australia (Edmonds, 2026), Canada (Marks v. Acuitas, 2026), and Ireland (IECA cases, 2026) have all cited or paralleled this principle in their AI-hallucination rulings.

The aggravating factor in Mata — the filing of fraudulent “confirmation” documents — is what elevated it from a cautionary tale to a landmark. Had Schwartz immediately acknowledged the error when first challenged, the outcome might have been a private reprimand or warning. Instead, the compounding of the original hallucination with deliberate deception created a case study in how not to respond to AI citation failures. This lesson has been absorbed by later courts globally: judges in the 2026 wave of AI-hallucination cases consistently credit immediate acknowledgment and apology as mitigating factors (as in Rodney), while treating evasion and cover-up as aggravating ones.

For Indian advocates, Mata v. Avianca should be understood as both a warning and a resource. The full judgment is freely available and serves as the clearest exposition in any common-law jurisdiction of why AI hallucinations occur, why they are dangerous in legal filings, and what a court expects when an attorney discovers such an error. India’s emerging AI governance framework — including MEITY’s advisory on responsible AI (2023) and the National Strategy for Artificial Intelligence — does not yet address legal practice AI governance specifically, creating a regulatory gap that the Bar Council of India should close with urgency before the first major Indian AI hallucination scandal forces a reactive rather than proactive response.