ChatGPT Fake Cases and Lawyer Sanctions

Mata v. Avianca: how the template was set

In June 2023, Judge P. Kevin Castel of the U.S. District Court for the Southern District of New York issued a sanctions order that became a landmark in legal AI ethics. Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023). The underlying case was a personal injury claim. The problem was the brief.

Plaintiff’s counsel had used ChatGPT to research supporting case law. The brief cited roughly half a dozen cases — with plausible-sounding party names, court designations, reporters, and quoted language — that did not exist. The cases were fabrications generated by the model.

When opposing counsel could not locate the cases, and the court ordered copies to be produced, the attorneys went back to ChatGPT. The model again confirmed the cases were real. The attorneys submitted those confirmations. The court found this constituted subjective bad faith — not just careless use of a new technology, but doubling down on fabrications after being put on notice.

The sanctions: Judge Castel imposed $5,000 separately against each attorney and $5,000 against the firm — $15,000 in total payable to the court’s registry, plus mandatory apology letters to each judge falsely named as the author of the invented opinions. The court identified six fabricated decisions: the purported Varghese, Shaboon, Petersen, Martinez, Durden, and Miller opinions.

Why AI hallucinations keep producing fake citations

Understanding why this keeps happening is more useful than simply knowing that it does. Large language models like ChatGPT are trained to predict the next most-likely piece of text given what has come before. They are not search engines. They do not query Westlaw, LexisNexis, or any legal database when you ask them a legal question.

When a model generates a legal argument, it knows from its training data that legal arguments are supported by citations. It knows roughly what a citation looks like: a party name, a court abbreviation, a year, a volume and reporter, a page number. It generates text that fits that pattern. Whether the text describes a real case is not a question the model has the architecture to answer. It generates plausible legal text; it does not retrieve verified legal records.

This is not a bug that future versions will necessarily fix. It is a structural feature of how generative models work. Retrieval- augmented generation (RAG) — where the model’s output is grounded in documents pulled from a verified database — can reduce hallucination rates significantly for citation tasks, but it requires deliberate engineering and does not eliminate the problem entirely. Consumer ChatGPT does not do this by default for legal research.

As Judge Castel wrote in the sanctions order: “Technological advances are commonplace, and there is nothing inherently improper about using a reliable artificial intelligence tool for assistance. But existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings.”

The pattern after Mata: subsequent sanctions

Mata was not an isolated incident. The same pattern — AI-generated brief, unverified citations, fabricated authority — has recurred in documented cases since June 2023.

The compounding-error pattern

In most sanctioned cases, the initial error — submitting AI-generated citations without verification — was serious but might have been treated more leniently if corrected promptly. What drove the bad-faith findings was the attorneys’ response after they were put on notice.

Asking ChatGPT to verify that the cases are real — and then submitting ChatGPT’s confirmation as evidence — is not verification. The model will confirm its own hallucinations. It has no way to know they are wrong. Returning to the tool that produced the fabrication for confirmation of its accuracy is not a reasonable verification step under any professional standard.

Judge Castel noted in Mata that had the attorneys acknowledged the error when first questioned, “the record now would look quite different.” The $15,000 in sanctions and the professional consequences that followed were driven by the attempt to cover the error rather than the underlying error itself.

The verification workflow that prevents this

For solo practitioners doing legal research with any AI tool, this workflow closes the gap:

This process takes roughly two to four minutes per citation. It is not burdensome relative to the risk. It is also exactly what the duty of competence and the duty of candor already required before AI entered the picture — AI tools simply make the consequence of skipping this step more likely and more visible.

What this means for confidentiality — the other risk

Hallucinated citations get the most press because they produce visible, documented sanctions. But for solo and small-firm attorneys, the confidentiality exposure from AI tools may be the larger practical risk. If you are using ChatGPT Free or Plus to research client matters and inputting client-specific facts, you are operating under data-handling terms that do not meet the standard most ethics guidance requires.

Both risks — hallucination and confidentiality — are addressed by understanding how these tools actually work, not by avoiding them entirely. If you use the right tier and verify every output before relying on it, the risk profile changes substantially.

Frequently asked questions