Mata v. Avianca: The $5,000 AI Sanction That Changed How Courts View ChatGPT
On June 22, 2023, Judge P. Kevin Castel of the Southern District of New York sanctioned attorneys Steven A. Schwartz, Peter LoDuca, and their firm Levidow, Levidow & Oberman, P.C. a total of $5,000 for submitting a brief containing six fabricated case citations generated by ChatGPT. The citations looked entirely real — complete with case names, reporter volumes, jurisdiction designators, and page numbers. Not a single one existed. The ruling became the founding document of the legal AI compliance era and triggered AI disclosure requirements in dozens of federal courts across the country.
⚖️ Mata v. Avianca, Inc. — S.D.N.Y., June 22, 2023
| Citation | No. 22-cv-1461 (PKC), 678 F. Supp. 3d 443 (S.D.N.Y. 2023) |
| Court | United States District Court for the Southern District of New York |
| Judge | Hon. P. Kevin Castel, United States District Judge |
| Sanction Date | June 22, 2023 |
| Total Sanction | $5,000 (against Schwartz, LoDuca, and the firm collectively) |
| Attorney | Steven A. Schwartz, admitted to NY bar 1992, Levidow, Levidow & Oberman, P.C. |
| Also Sanctioned | Peter LoDuca (filing partner) and the firm |
| AI Tool | ChatGPT (public consumer version) |
| Core Violation | Fed. R. Civ. P. 11(b)(2) — certification that legal contentions are warranted by existing law |
Mata v. Avianca began as an unremarkable personal injury lawsuit — a passenger alleging injuries sustained on a flight operated by Avianca. What transformed it into a landmark of legal AI history was the research process Attorney Steven Schwartz used to find supporting case law. Instead of using Westlaw, Lexis, or any established legal research database, Schwartz turned to ChatGPT. The model returned a list of apparently on-point aviation injury decisions. Schwartz, by his own subsequent admission, did not verify a single citation. He submitted the brief. Six of the cited cases did not exist anywhere in any legal database on earth.
Timeline: How the Fake Cases Were Discovered
OpenAI releases ChatGPT to broad public access. Legal professionals begin experimenting with it for research, drafting, and summarization without formal guidance or bar ethics opinions addressing AI use.
Attorney Steven A. Schwartz, handling a personal injury claim against Avianca, queries ChatGPT for relevant case law. The model returns six plausible-sounding citations with detailed case names, courts, dates, and page numbers. Schwartz does not verify them against any legal database.
Partner Peter LoDuca files the brief in the Southern District of New York. The brief cites Varghese v. China Southern Airlines Co., Shaboon v. Egyptair, Petersen v. Iran Air, Martinez v. Delta Air Lines, Estate of Durden v. KLM, and Zicherman v. Korean Air Lines — none of which appear in Westlaw, Lexis, or any court database.
Opposing counsel at Avianca conducts routine citation checks and discovers that none of the six cases can be located in any legal research database. Avianca's attorneys notify Judge Castel and request copies of the decisions.
Rather than checking authoritative databases, Schwartz returns to ChatGPT and asks it to confirm the cases are real. ChatGPT generates fake "screenshots" purporting to show the decisions. Schwartz submits these to the court, compounding the original error.
Judge Castel orders the attorneys to explain why sanctions should not be imposed. In their response, Schwartz admits he was unaware that ChatGPT could fabricate citations, and that he believed the AI's output to be reliable without independent verification.
Judge Castel issues a 46-page sanctions opinion imposing $5,000 in sanctions jointly against Schwartz, LoDuca, and the firm. The opinion articulates a framework for attorney responsibility in the AI era that state bars across the country would spend the next two years encoding into formal ethics opinions.
Judge Castel's Key Statements on Attorney Responsibility
Judge Castel's opinion is notable not merely for imposing sanctions but for the clarity and force of its articulation of professional responsibility principles in the context of AI tools. The following passages from the sanctions order have been cited repeatedly by courts, bar associations, and legal commentators:
On the harm of fake citations: "Many harms flow from the submission of fake opinions: the opposing party wastes time and money researching non-existent cases; the Court's time is taken from real cases; and there is the potential that a judge may be misled by a citation that cannot be found and verified."
This passage established the framework through which subsequent courts have analyzed AI-hallucinated citations. The harm is not merely abstract — it is specific, measurable, and falls on identifiable parties: the opposing counsel who spent hours trying to locate non-existent decisions, the court staff who attempted to pull cases from databases, and the judicial system's credibility when fabricated authority is submitted as if real.
On the nondelegable duty to verify: Judge Castel held that attorneys have a professional obligation to verify every citation they submit to a court. This duty cannot be delegated to any tool — artificial or human — and cannot be discharged by reliance on an AI's representation that its output is accurate. The obligation runs directly from the attorney to the court under Rule 11.
On Schwartz's claim that he was unaware ChatGPT could fabricate cases, Judge Castel was direct: the lack of awareness did not constitute a defense. Attorneys have an affirmative duty to understand the tools they use in practice, a duty codified in ABA Model Rule 1.1's competence requirement. Using a tool without understanding its fundamental limitations is not a mitigating circumstance — it is itself a competence failure.
When Schwartz asked ChatGPT to verify that the cases existed — and the model generated fake "confirmation" — he was relying on the same unreliable system to validate its own output. This is a fundamental methodological error: using a tool that can confabulate to check whether its confabulations are accurate. The court found this made the conduct worse, not better.
LoDuca's sanctioning established a separate but equally important principle: a supervising attorney who signs a brief is responsible for its contents, even if they did not conduct or review the underlying research. Signing a filing is a certification to the court under Rule 11. That certification cannot be made in good faith when the signatory has not reviewed the authority cited in the brief.
Why ChatGPT Hallucinates Legal Citations
The narrative in the popular press framed the Mata v. Avianca disaster as an AI "hallucination" — as though ChatGPT had experienced some kind of glitch. It had not. The fabrication of legal citations is an entirely predictable consequence of how large language models work, and understanding that mechanism is essential for any attorney deploying AI tools in legal practice.
How LLMs Generate Text — and Why It Produces Fake Citations
Large language models like GPT-4 are not retrieval systems. They do not query databases. They do not look up cases. They are statistical text generators: given an input sequence of tokens, they calculate the probability distribution over possible next tokens and select from that distribution. The model has seen millions of legal documents in training, including case reporters, briefs, motions, and law review articles. It has learned the patterns of what legal citations look like — the structure of case names, reporter abbreviations, volume numbers, court designators, and year formats.
When asked for "aviation injury cases involving delayed flights," the model does not search a database. It generates text that looks statistically like a list of aviation injury cases involving delayed flights. It fills in the case name, the court, the year, the reporter volume, and the page number by selecting tokens that have high probability of appearing in that context, given what it learned about legal citation patterns. The model has no mechanism to check whether the combination it generated corresponds to a real decision — and, crucially, it has no mechanism to know that it does not know.
Consumer ChatGPT Legal Research
- No connection to Westlaw, Lexis, or any legal database
- Generates citation text statistically, not from records
- Cannot distinguish real cases from fabrications
- Presents invented citations with same confidence as real ones
- Will "confirm" its own fabrications if asked
- No audit trail of what was searched or returned
- No real-time legal database access of any kind
- Training data cutoff — unaware of recent decisions
- Cannot signal when it lacks sufficient information
- Rule 11 violation risk with every filing
Claire's Citation-Verified Legal Research
- Integrated verification against authoritative legal databases
- Every citation confirmed before delivery to attorney
- Unverifiable citations flagged explicitly, not omitted silently
- Audit trail of every search, query, and output
- Signals confidence level for each research result
- Cannot generate and present unverified citations as real
- Attorney retains final verification responsibility
- Workflow integrates with Westlaw / Lexis verification step
- Session logs available for malpractice defense
- Rule 11-aware research workflow documentation
Training Data, Pattern Completion, and the Absence of Grounding
Three specific architectural characteristics of public ChatGPT create the hallucination risk in legal contexts. First, the model was trained on a static dataset with a knowledge cutoff — it has no awareness of cases decided after that cutoff, and it cannot distinguish between cases it "knows" from training and cases it invents to fill the pattern. Second, the model is optimized to be helpful and to produce fluent, complete responses — which means it will attempt to answer even questions it cannot reliably answer, rather than declining to respond. Third, there is no retrieval grounding: the model does not query a live legal database before generating its response, and there is no feedback loop that would allow it to detect when a generated citation fails to match any real record.
Retrieval-Augmented Generation (RAG) architectures, purpose-built legal research models with database access, and citation-verification layers can substantially reduce hallucination rates. Public ChatGPT, as Schwartz used it, had none of these safeguards. The architectural conditions that produced the Mata v. Avianca disaster are still present in the consumer product today.
Court AI Disclosure Requirements That Followed
The Mata v. Avianca ruling catalyzed a rapid and broad response from federal courts. Within eighteen months of Judge Castel's opinion, more than two dozen federal courts had issued standing orders or local rules requiring attorneys to disclose the use of AI tools in court filings. The following table identifies ten of the most consequential orders:
12-Item AI Citation Verification Checklist for Law Firms
AI Citation Verification Checklist — Post-Mata v. Avianca
Adopt a written firm policy: no citation sourced from any AI tool may be submitted to any court without independent verification through Westlaw, Lexis, Fastcase, or CourtListener. Make this a condition of filing approval.
Before reading the holding, analyzing the facts, or using the citation in any argument, confirm that the case exists in a primary legal database. A citation that cannot be found cannot be used, regardless of how well it appears to support your argument.
LLMs often generate plausible but incorrect reporter volumes and page numbers even when the case itself is real. Verify that the specific reporter citation — not just the case name — pulls the correct decision. Bluebook accuracy is legally material.
AI tools frequently mischaracterize holdings, extract quotes out of context, or conflate the majority opinion with a dissent. Every cited case must be read by the attorney who will certify the filing — not summarized by AI and taken on faith.
A case that exists may have been subsequently overruled, distinguished, or limited. Run KeyCite or Shepard's on every case before citing it. AI tools do not perform this check and have no awareness of post-training case developments.
The Schwartz compounding error: asking ChatGPT to confirm citations it generated. An AI system cannot reliably check its own confabulations. Verification must always use a source that is architecturally independent of the system that generated the output.
Maintain a log of each AI-assisted research session: date, AI tool used, queries submitted, citations returned, and verification steps taken. This documentation is your malpractice defense and your response to a court's show-cause order.
The filing attorney must personally review every citation in any brief touched by AI research. Delegation to associates or paralegals without personal review exposes the supervising attorney to Rule 5.1/5.3 liability, as LoDuca discovered.
Determine whether the court in which you are filing has issued an AI disclosure standing order. More than 20 federal courts now require affirmative disclosure of AI use in filings. Failure to comply with a standing order is an independent sanctionable violation.
Every attorney and paralegal who may use AI for legal research must receive specific training on LLM hallucination, the absence of legal database access in consumer AI tools, and the non-delegable verification obligation under Rule 11.
Per ABA Formal Opinion 512 (2023), attorneys should disclose AI use in engagement letters. Include a description of the AI tools used and the verification protocols your firm employs to ensure accuracy of AI-assisted work product.
The most reliable protection against Mata-style sanctions is deploying an AI research tool that is architecturally incapable of presenting unverified citations as real. Enterprise legal AI with integrated database verification eliminates the hallucination risk at the source rather than relying solely on attorney review post-hoc.
How Claire Prevents Hallucination in Legal Contexts
Claire's Citation-Verified Legal Research Architecture
The Mata v. Avianca failure was an architectural problem. Consumer ChatGPT was not designed for legal citation accuracy — it was designed to produce helpful, fluent text. Claire's legal research capabilities were built from the ground up to address the specific failure modes that Judge Castel's opinion identified.
No Unverified Citation Delivery
Claire does not deliver legal citations to attorneys without verification against authoritative sources. When a legal research query returns case citations, the system queries integrated legal databases to confirm existence before including them in research output. A citation that cannot be verified is flagged explicitly — not silently omitted and not presented as real.
Grounded Retrieval, Not Pattern Completion
Unlike consumer ChatGPT, which generates citation text statistically from training patterns, Claire's legal research workflow uses retrieval-augmented generation: the model queries a legal database and grounds its response in retrieved documents. The difference is architectural — the system has access to actual case records, not just statistical patterns of what citations look like.
Audit Trail for Every Research Session
Every legal research session in Claire is logged with full detail: timestamp, attorney ID, matter number, queries submitted, sources retrieved, and output delivered. This audit trail is stored in your practice management system under your control — not in Claire's systems. It constitutes the documentation required by courts issuing post-Mata AI disclosure orders.
Explicit Confidence Signaling
Claire communicates uncertainty explicitly rather than presenting all output with equal confidence. When the system has lower confidence in a research result, it tells the attorney — with specific reasons — rather than presenting uncertain output with the same prose fluency as high-confidence output. Attorneys receive calibrated information, not false certainty.
Court-Specific Standing Order Compliance
Claire's research workflow integrates awareness of the specific AI disclosure requirements that apply in the courts where your firm practices. The system can generate the affirmation language required by standing orders in the 5th Circuit, N.D. Cal., E.D.N.Y., and other courts that have issued AI disclosure requirements, ensuring that your filings satisfy procedural requirements as well as substantive research standards.
Rule 11 Workflow Integration
Claire's legal research module integrates with your filing workflow to flag AI-assisted research sections before submission. The system reminds attorneys of their independent verification obligations and generates documentation of the verification steps taken — creating a defensible record that satisfies the "reasonable inquiry" standard of Rule 11.
Consumer ChatGPT vs. Claire: Citation Research Comparison
The pattern from Mata v. Avianca through the cascade of subsequent cases is unambiguous: courts are not treating AI hallucinations as a novel mitigating circumstance, they are not accepting lack of familiarity with AI limitations as a defense, and they are increasingly imposing significant procedural and financial consequences for failures to verify AI-generated legal research. The verification obligation is not aspirational guidance — it is enforceable professional responsibility backed by Rule 11 and a growing body of sanctions authority.
For law firms, the choice is not whether to use AI for legal research — the efficiency benefits are too substantial to ignore. The choice is which AI architecture to use. Consumer ChatGPT, as the Mata v. Avianca record makes clear, was not designed for the verification requirements of legal practice. Purpose-built enterprise legal AI with integrated citation verification, audit trails, and court-specific compliance features represents the only responsible path forward for firms that intend to benefit from AI efficiency without replicating Schwartz's mistakes.
For a deeper analysis of how AI use can waive attorney-client privilege — the next dimension of AI risk in legal practice — see United States v. Heppner: The Federal Ruling That Redrew Attorney-Client Privilege for the AI Era. For an overview of ABA Model Rules compliance obligations for AI use, see ABA Model Rules 1.1, 1.6, 5.3 and AI: The Legal Ethics Framework Every Law Firm Needs.
