- In Mata v. Avianca Inc., No. 22-cv-1461 (S.D.N.Y. 2023), attorney Steven Schwartz submitted a legal brief containing six AI-generated case citations that did not exist. Judge Castel imposed USD 5,000 in sanctions.
- The case establishes a clear principle: when a professional uses an AI tool to produce output and submits that output under their professional authority, they are responsible for its accuracy. The tool does not absorb the responsibility.
- The same principle applies outside law to any operator whose AI-assisted outputs are relied on by others: accountants, compliance officers, engineers, medical professionals, and any business whose customer-facing AI agents make factual representations.
- Professional indemnity policies are the relevant coverage instrument but are increasingly including AI-specific exclusions or disclosure requirements. Operators should review their current policy and ensure AI use is disclosed.
- Hallucination is not a temporary problem pending better models. It is a structural property of large language models that requires verification processes, not trust. Building verification into AI-assisted workflows is both an operational necessity and an insurance requirement.
What happened
Roberto Mata sued Avianca Airlines in the Southern District of New York following a personal injury incident on a 2019 flight. The case, filed in 2022, was handled by attorney Steven Schwartz of Levidow, Levidow and Oberman. When preparing a brief opposing Avianca's motion to dismiss, Schwartz used ChatGPT to assist with legal research.
ChatGPT produced several case citations in support of the brief's arguments. Schwartz did not verify these citations against Westlaw, Lexis, or any other legal database. He filed the brief. Among the cases cited were: Varghese v. China Southern Airlines Co. Ltd.; Shaboon v. Egyptair; Petersen v. Iran Air; Martinez v. Delta Air Lines; Estate of Durden v. KLM Royal Dutch Airlines; and Zicherman v. Korean Air Lines Co. None of these cases, as cited with the docket numbers and holdings that ChatGPT provided, existed.
Avianca's counsel flagged that the cited cases could not be located. Judge P. Kevin Castel directed Schwartz to provide copies of the cited opinions. Schwartz asked ChatGPT to confirm that the cases were real. ChatGPT confirmed they were real. Schwartz continued to believe them. He filed an affidavit explaining the situation, which included his account of asking ChatGPT to confirm the cases and its responses.
Judge Castel's June 2023 opinion was precise about what had gone wrong. Schwartz and the firm had "abandoned their responsibilities" by failing to verify the AI's output against any actual legal source. The court found no evidence of bad faith in the sense of intentional fabrication, but found that the conduct fell below the standard required of a licensed attorney. The sanction was USD 5,000 paid jointly by Schwartz and the firm, plus a requirement to send copies of the sanction order to the clients whose matter had been affected.
Why the financial sanction understates the damage
USD 5,000 is a modest sanction. It does not capture the full cost of the incident. The firm's name was attached to a published federal court opinion describing professional misconduct. Every opposing counsel who searches the partners' names will find it. Every client who asks will need an explanation. The State Bar received a referral. The attorneys spent months of time on the response to the court's inquiries and the remediation process. The reputational and professional consequences of the incident extend well beyond the financial one.
For any professional who thinks "I will use AI and just pay the fine if something goes wrong," Mata v. Avianca provides the corrective. The fine is not the cost. The professional record is.
The principle the case establishes
The case is often described as a story about AI going wrong. That is not quite right. It is a story about professional responsibility. ChatGPT did exactly what it was designed to do: it produced fluent, plausible text about legal cases. It did not know the cases did not exist, because it does not have access to a live legal database and does not represent its output as verified fact. It is not the responsible party.
The responsible party is the attorney who submitted the output under their professional authority without verifying it. The principle Judge Castel applied is not specific to AI. It is the principle that a professional is responsible for the accuracy of work submitted under their name. A lawyer who relies on a paralegal's research and submits it without checking is equally responsible if the research is wrong. A doctor who copies a junior's prescription note without verifying the dosage is equally responsible if it is wrong. An accountant who files a client's numbers without checking them is equally responsible if they are incorrect.
AI changes the speed at which plausible but wrong content can be produced. It does not change the principle that the professional who submits it owns the responsibility.
The operator implications
The Mata v. Avianca principle extends far beyond legal practice. Any operator whose AI-assisted outputs are relied on by others faces the same structural risk. Consider several scenarios that are directly analogous.
A compliance officer uses an AI tool to produce a regulatory filing. The AI misstates a provision of the relevant regulation. The officer submits the filing without checking the provision. The regulator raises the error in an enforcement action. The question of professional responsibility for the error lands on the officer and the organisation, not on the AI vendor.
A customer service AI agent tells a customer that a specific warranty provision applies to their product, citing a policy that the company discontinued two years ago. The customer acts on this information, suffers a loss, and brings a claim. The Moffatt v. Air Canada tribunal decision from February 2024, discussed in our Air Canada case article, confirms that the business is responsible for what its AI agent tells customers. The Mata principle adds that the professional within the business who deployed the agent without adequate verification bears professional responsibility alongside the entity.
A software engineer uses an AI code generator to produce a cryptographic library function. The function has a security flaw. The engineer ships it without reviewing the security properties. A breach occurs. The engineer and their employer are responsible for the defective code. The AI tool is not.
In each scenario, the variable is not whether the operator used AI. It is whether the operator applied their professional judgment to verify the AI's output before relying on it.
What verification looks like in practice
Verification is not the same as asking the AI to confirm its own output. This is one of the clearest lessons from Mata: Schwartz asked ChatGPT whether the cases were real, ChatGPT said yes, and he continued to believe it. A hallucinating model asked to verify its own hallucinations will confirm them confidently. Verification requires checking AI output against an authoritative source that is independent of the AI.
For factual claims: locate the primary source the AI is apparently citing and check that it says what the AI claims. Do not rely on what the AI says about the source. Retrieve the source itself.
For citations: retrieve the cited document and confirm it exists, has the correct title, author, date, and holding or content the AI attributes to it. In legal contexts this means checking a verified legal database. In scientific contexts this means checking PubMed or the journal's own archive. In regulatory contexts this means checking the Official Journal or the relevant authority's published legislation.
For numerical data: cross-check the figure against the original data source, whether that is a financial statement, a regulatory filing, a survey dataset, or a government publication. AI models frequently interpolate plausible numbers that are not in any source.
For regulatory or legal statements: verify against the current text of the relevant instrument, including any amendments that have entered into force since the AI's training cutoff. EU AI Act Articles, for example, are being interpreted and supplemented by AI Office guidance that is more recent than most models' training data.
Building verification into the workflow means treating AI output as a draft, not a source. The AI compresses the time to a first draft. The human completes the work by verifying it. This is not a limitation of AI tools. It is the correct division of labour between a tool and a professional.
The insurance implications
Professional indemnity insurance covers losses arising from professional errors, including errors in AI-assisted work product. The relevant question for any professional or business using AI for substantive outputs is whether their current professional indemnity policy responds to a Mata-style claim.
Many professional indemnity insurers are revising their policy wordings in response to the growth of AI-assisted professional work. Some are adding exclusions for losses arising from AI-generated outputs that were not adequately verified. Others are adding disclosure requirements, requiring policyholders to notify the insurer when AI tools are used for professional outputs covered by the policy. Failure to disclose may affect coverage. Operators should review their current professional indemnity policy for these provisions and ensure their AI use is disclosed as required.
Where a professional indemnity policy includes an AI exclusion or does not cover AI-assisted outputs, a specialist AI liability product may be needed. The market for specialist AI professional liability coverage is early but forming. For the current state of the market and where coverage is and is not available, see the three-step coverage pathway.
For the EU regulatory layer that applies to businesses deploying AI agents commercially, including the EU AI Act operator obligations and the revised Product Liability Directive exposure that arrives in December 2026, see Agent Liability EU's operator obligations guide.
A note on the case's limits
Mata v. Avianca is a federal district court sanction order, not a court of appeals decision or a Supreme Court ruling. It does not create binding precedent outside the Southern District of New York. It has been widely cited, discussed, and referenced by bar associations, legal educators, and professional bodies across the United States and in several European jurisdictions, but it is persuasive authority rather than binding law in most of those contexts.
That said, the principle it applies, that the professional is responsible for the accuracy of work submitted under their authority, is not novel. It is the application of a longstanding professional responsibility principle to a new tool. Courts in other jurisdictions reasoning about AI-assisted professional errors will find the same principle in their own professional responsibility frameworks, even without the Mata decision as a direct precedent.
For operators in Europe, the relevant professional responsibility frameworks are national, and the enforcement trajectory is behind the US by a year or two. That gap is narrowing. The Why It Matters section of this site provides the current view of where European case law and regulatory enforcement is heading.