The hallucination problem and Mata v. Avianca

The single most important thing to understand about legal AI is that it fabricates citations, and the foundational cautionary case is Mata v. Avianca. Getting its facts exactly right matters, because it defines the risk and the lesson.

What happened in Mata v. Avianca. In a personal-injury case in the U.S. District Court for the Southern District of New York, plaintiff's counsel opposed a motion to dismiss with a brief citing several non-existent judicial decisions — with fabricated quotations and internal citations. The cases (such as a fictitious "Varghese v. China Southern Airlines") had been generated by ChatGPT. When opposing counsel and the court could not find them, attorney Steven Schwartz admitted he had used ChatGPT for the research — and had even asked ChatGPT whether the cases were real, and it falsely assured him they were.

The sanction — state this precisely. In June 2023, Judge P. Kevin Castel imposed a $5,000 penalty, jointly and severally, on the two attorneys (Steven Schwartz and Peter LoDuca) and their firm, under Rule 11. He also required them to notify their client and each real judge whose name had been falsely invoked in the fabricated opinions. A crucial nuance: the court did not hold that using AI is inherently sanctionable — the violation was the failure to verify the citations and then defending the fabricated cases once questioned. (Common error to avoid: it was $5,000 total, joint-and-several — not $5,000 each, and not a larger figure. Later cases carry far bigger sanctions; don't blur them into Mata.)

Why it's the foundational lesson:

• AI fabricates confidently. The fake cases had realistic names, citations, and quotes. Fluency is not accuracy.
• The tool will "confirm" its own fabrications. Asking the AI "are these real?" is not verification — it will lie reassuringly.
• The duty to verify is the lawyer's, always. The sanction was for filing unverified fabrications, not for using a tool. Verification against primary sources is non-negotiable.
• Defending the error made it worse. Candor to the court (Module 3) means correcting mistakes promptly, not doubling down.

What every lawyer using AI must do:

• Verify every citation, quote, and holding against a primary source (the actual reporter, Westlaw/Lexis, the real opinion) before filing or relying on it.
• Never treat the AI's self-confirmation as verification — check the source yourself.
• Assume any AI-provided citation might be fabricated until you've confirmed it exists and says what the AI claims.
• If you discover an error in a filing, correct it promptly with the court.

The mindset: Mata v. Avianca is the case every lawyer using AI must know — generative AI fabricates realistic-looking citations, quotes, and holdings, and the tool will even falsely confirm they're real. The $5,000 Rule 11 sanction was for failing to verify and then defending the fabrications, not for using AI. The enduring lesson is absolute: verify every AI-provided citation and proposition against a primary source before relying on it, never mistake the AI's reassurance for verification, and correct errors candidly. Master this, and you avoid the defining failure of legal AI. Ignore it, and you join the growing list of sanctioned lawyers (Module 3).