You remember the case. Every attorney does.
In the summer of 2023, two New York lawyers filed a brief in federal court citing six decisions that did not exist. The brief had been drafted with the help of ChatGPT. The cases were fabrications — plausibly named, plausibly reasoned, and entirely imaginary. The judge noticed. The attorneys were sanctioned. The legal press spent a full news cycle treating it as a parable about the dangers of AI in law practice.
Three years later, most attorneys' mental model of AI in legal research still lives inside that news cycle. Mata v. Avianca is the story you heard when you were deciding whether to trust AI with your bar license, and it is the story that told you not to.
That instinct was correct at the time. It is worth revisiting now, because the lesson of Mata v. Avianca was never "AI cannot be trusted in legal research." The lesson was much more specific than that — and the specificity is the part that matters when you are deciding, in 2026, whether to let an AI tool anywhere near the work you sign your name to.
What actually happened
The attorneys in Mata v. Avianca did something specific: they asked a general-purpose conversational AI — ChatGPT, built for broad text generation — to tell them whether certain case law existed that supported their position. The AI, having been trained to produce fluent text, produced fluent text. It generated case names. It generated citations. It generated holding summaries. None of it was grounded in an actual database of real opinions, because the AI did not have one; it had been trained on a snapshot of internet text and was predicting what a persuasive legal brief might say, not retrieving what an actual legal brief had said.
The attorneys then asked the AI whether the cases were real. The AI said yes. It was again doing what it had been trained to do — generate plausible text — and "yes" is the kind of answer that fits conversational patterns about existence. There was no separate verification mechanism because the AI had no ability to verify anything; it was a language model, not a research tool, and it had been used as something it was never designed to be.
The attorneys did not verify the citations independently. That is the moment the case became a disciplinary matter. Had either of them opened Westlaw, or PACER, or the actual federal reporter, the fabrication would have surfaced in ninety seconds. They did not. They trusted the AI's confidence as a substitute for verification, and the court treated that failure as the sanctionable conduct.
What the case actually taught
Read carefully, Mata v. Avianca is not a case about AI. It is a case about the use of a general-purpose language model as if it were a research tool, combined with a failure to verify.
The distinction matters because there are now two categorically different kinds of AI tools attorneys are being asked to evaluate, and conflating them is how good attorneys talk themselves out of helpful technology.
Generative AI. Tools like ChatGPT, Claude (as a general assistant), and Gemini produce text by predicting what a plausible response would look like based on patterns in their training data. They do not have a database of opinions to retrieve from. When they "cite" cases, they are producing text that looks like a citation, which may or may not correspond to a real case. This is the category the Mata v. Avianca attorneys used, and the category that caused the problem.
Retrieval-augmented AI. Tools built specifically for legal research operate differently. They begin with a database of actual opinions — ingested, indexed, and searchable. When a user asks a question, the tool retrieves the relevant real opinions first, then uses AI to synthesize an answer grounded in the retrieved text. The AI is not generating cases; it is summarizing cases the system has already pulled from a verified corpus. The citations in the answer correspond to real documents the tool can link to.
The difference is not academic. A retrieval-augmented system that is working correctly cannot fabricate a citation, because its citations come from the corpus, not from the model. If the corpus contains the opinion, the tool can cite it; if the corpus does not, the tool has nothing to cite. The Mata v. Avianca failure mode is not available in a properly-constructed retrieval system.
The questions that matter in 2026
If you are evaluating an AI legal research tool today — ours, a competitor's, or one that has not launched yet — the lesson of Mata v. Avianca is not "walk away." It is "ask better questions."
Where do the citations come from? If the answer is "the model generates them," walk away. If the answer is "the system retrieves them from a database of actual opinions," keep going.
Can I open every cited case? The citation should link to a real source — a court website, a PDF, a verifiable document. If you cannot click through to the actual opinion, you cannot verify, and if you cannot verify, the tool has recreated the Mata v. Avianca failure mode.
What is the corpus? A tool that claims to cover "all U.S. law" is either impossibly ambitious or meaningfully shallow in any given jurisdiction. A tool that covers a specific, defined corpus — every West Virginia Supreme Court of Appeals opinion since 1991, every Intermediate Court of Appeals opinion since 2022, the full West Virginia Code — can make precise claims about what it knows and, just as importantly, what it does not.
What does the tool do when it does not know? A trustworthy tool says "there is no controlling WV authority on this question" when that is the truth. An untrustworthy tool fabricates authority. The Mata v. Avianca attorneys were destroyed by a tool that could not say "I don't know," and any AI research tool you adopt should be one that can.
Can I see the reasoning? A black-box answer — here is the conclusion, trust us — is a liability. A transparent answer shows the retrieved opinions, shows how they apply, and shows where the synthesis is drawing its conclusions. The attorney reviews the reasoning the same way she would review a first-year associate's memo: critically, with the authority to disagree.
The verification mandate is unchanged
The final lesson of Mata v. Avianca is the one that should be most reassuring, because it is the one that was already baked into the practice of law long before AI existed.
Every attorney is responsible for every citation in every document she signs. That obligation existed before ChatGPT. It existed before Westlaw. It existed when Shepard's was a physical set of books. An attorney who files a brief with an incorrect citation — whether that citation came from a junior associate, a treatise, a secondary source, or an AI — is the attorney whose name is on the filing, and the filing is her responsibility.
AI does not change that obligation. It is a tool, and like every tool in the practice of law, it is subject to the attorney's verification. The Mata v. Avianca attorneys did not get sanctioned because they used AI. They got sanctioned because they failed to verify. Had they used AI responsibly — retrieved, read, and confirmed the cases before filing — the case would not be in our memory, because there would be no case.
A well-built AI research tool makes verification easier, not harder. It gives the attorney the real opinion, links to the real PDF, shows the real reasoning, and then steps out of the way so the attorney can do what attorneys do: read carefully, think critically, and decide whether to rely.
What to do with this
If the Mata v. Avianca story has been the reason you have not yet evaluated any AI research tool, the reason is no longer current. The category of tool that caused the problem is not the only category of tool that exists. The category of tool that is useful to you as a West Virginia attorney — retrieval-augmented, corpus-specific, citation-linked, verifiable — is a different category entirely, and it has been maturing quickly.
The right posture is not trust-without-verification, which is what got the Mata v. Avianca attorneys sanctioned. The right posture is also not avoidance-without-evaluation, which is the posture most attorneys have inherited from that story without examining it.
The right posture is the one that has always governed good legal practice: ask what the tool does, ask how it does it, verify the output before you rely on it, and keep your judgment at the center of the work. Attorneys who adopt that posture toward a well-built AI research tool get the benefit of the technology — less time hunting, more time thinking — without the liability the Mata v. Avianca attorneys incurred.
The tool does not replace the lawyer. It never will. But the 2023 story about what went wrong when a general-purpose chatbot was misused is not the 2026 story about what is possible when the right tool is used correctly. The distinction is worth your attention.
West Virginia Case Search answers legal questions in plain English, grounded in every WV Supreme Court of Appeals opinion since 1991, every WV Intermediate Court of Appeals opinion since 2022, and the full West Virginia Code. Every citation links to the source PDF on the court's website. Every answer is graded by precedent strength. You verify before you rely.
Start a 7-day free trial — no credit card required.