Three Years After Mata, the Lesson Is Not About ChatGPT - Issue #9
"It is about whether law firm leadership has decided how AI-assisted work gets reviewed, supervised, and governed before it carries the firm's name."
- Rex Anderson
What Mata Actually Taught
In June 2023, a federal court in New York sanctioned two attorneys after a brief filed in Mata v. Avianca cited six nonexistent cases generated through ChatGPT. The cases were not merely misquoted. They were fabricated: names, docket numbers, parties, quotations, and legal analysis. But the sanction order is often flattened into the wrong lesson. The court did not impose monetary sanctions simply because an AI tool produced false citations.
The turning point came after the problem was identified. Opposing counsel questioned the existence of the cited cases. The court ordered production of the authorities. At that point, the attorneys had an opportunity to disclose what had happened. Instead, they submitted materials purporting to support the nonexistent cases and did not begin revealing the truth until after the court issued an order to show cause. In Mata, Judge Castel made the distinction explicit: if the attorneys had come clean earlier, “the record now would look quite different.” What drew the financial sanction was not only the original mistake, but the decision to double down after the mistake had been exposed.
That distinction matters. Mata did not hold that lawyers may not use AI. Judge Castel wrote that “there is nothing inherently improper about using a reliable artificial intelligence tool for assistance.” The professional failure was unverified AI-assisted legal work submitted under the firm’s name, followed by evasive conduct when the problem surfaced. The first lesson is verification. The second is candor. Together, they make Mata an ethics case, not a chatbot case.
Two other details are often blurred in casual discussion of Mata. First, the client did not lose the underlying case because of ChatGPT. The claim was dismissed on a separate timing issue under the Montreal Convention. The AI incident produced a sanctions proceeding, not the merits result. Second, the court did not impose mandatory technology training as a sanction. The firm had already arranged its own technology competence training, and the court declined to add that requirement.
Those details are not trivia. They are the difference between using Mata as a scare story and understanding it as a professional responsibility case. The lesson was not that AI is forbidden. The lesson was that AI-assisted work still requires verification, supervision, and accountability before it leaves the firm.
What Has Changed Since Mata
In 2023, Mata could still be treated as an early warning from the first public wave of generative AI. Three years later, that posture is no longer available. AI-generated false authority is no longer a novelty. It is a foreseeable professional responsibility problem.
The important shift is not numerical. It is institutional. A lawyer who submits fabricated authority can no longer plausibly say the risk was too new to understand. The question now is whether the lawyer, the supervising attorneys, and the firm had a method for verifying AI-assisted work before it became filed work, client work, or professional advice carrying the firm’s name.
The Ninth Circuit’s decision in Lnu v. Blanche (June 2026) marks that shift. The court treated AI-tainted filings as professional misconduct with consequences beyond embarrassment. Among other sanctions, the court ordered every attorney at their firm, for two years, to disclose in future filings whether generative AI was used, identify the tool, and certify under penalty of perjury that the signing attorney personally reviewed the filing and confirmed that all citations and quotations referred to existing authority. That matters because the consequence was not only personal discipline. It became an operating burden.
In my home state of Arizona, the Ninth Circuit’s opinion is now part of the professional landscape for lawyers. The Arizona State Bar issued its own guidance in 2024 covering competence, confidentiality, supervision, and billing, with an explicit warning that relying on AI without independent verification could result in discipline under existing rules. Lawyers and firms can no longer treat AI verification as an academic topic, a technology-policy side issue, or a problem limited to someone else’s careless filing.
The movement is not toward forbidding AI. It is toward ordinary professional accountability. AI-assisted work still has to be competent, confidential, supervised, candid, and reasonable. The tools are new. The professional questions are not.
The Real Gap Is Not Technical
The deeper problem is not that AI sometimes produces false authority. Lawyers already know legal work must be checked before it is filed, sent, or relied upon. The deeper problem is that AI use can spread inside a firm before leadership has decided who may use it, for what kinds of work, under what supervision, and with what verification before the work carries the firm’s name.
That is how isolated mistakes become institutional exposure. A lawyer who uses an AI tool badly may create the immediate problem. But a firm that has never clarified what tools may be used, what information may be entered, what output must be verified, who must review it, and how lawyers and staff should handle AI-assisted work has created the operating condition in which the problem can occur.
The issue is not whether a managing partner understands transformers, tokens, model weights, or prompt engineering. The issue is whether firm leadership has made clear decisions about professional work product in an environment where AI tools can produce fluent error with extraordinary confidence. When those decisions have not been made, the firm has not avoided the issue. It has simply delegated the decision to individual lawyers and staff, one prompt at a time.
Verification Is Still a Lawyer's Work
One reason Mata continues to matter is that it forces a very old question into a new workflow: who verified this before it carried the firm's name? Verification is not a vague instruction to "check the AI." It means confirming that each cited authority exists. It means reading the actual source rather than relying on a generated summary. It means confirming that quoted language appears in the source. It means confirming that the proposition for which the authority is cited is actually supported by the case, rule, statute, or other source. It means checking whether the authority has been reversed, limited, distinguished, superseded, or otherwise weakened.
Those are not new duties. They are the same review habits a supervising lawyer would expect when reviewing a junior associate's research memo. AI changes the risk profile because it can produce a polished answer that appears complete while failing every one of those checks. It can provide names, dates, quotations, and analytical confidence without legal reality underneath them.
That is the danger. Not that AI output is always wrong. Not that lawyers should avoid AI tools. The danger is that AI output can be wrong in a way that looks finished. That makes verification a workflow issue, not just an individual virtue. A firm cannot rely on every lawyer and staff member to improvise a verification method every time AI enters a task. If AI-assisted research, drafting, summarization, contract review, correspondence, discovery work, or client communication is permitted, the firm needs a clear answer to what review is required before that work becomes firm work product.
The Question for Managing Partners
For a managing partner, the most important question after Mata is not whether AI hallucination exists. The profession has had three years to learn that lesson. The more important question is whether the firm has decided what AI-assisted work requires before it leaves the firm. That decision has four basic components.
First, which AI tools are permitted, restricted, or prohibited in specific categories of work? A tool used to summarize public information is not the same as a tool used to draft a client memorandum, analyze privileged documents, or prepare language for a court filing. The policy question is not "AI, yes or no." It is which AI, for which work, with what limits.
Second, what information may be entered into those tools? Confidential information, privileged material, client documents, personal data, litigation strategy, and internal firm information do not all present the same issues. If lawyers and staff have no clear guidance, they will make individual judgments under pressure.
Third, what verification is required before AI-assisted output is used? The answer may differ by workflow. A generated internal outline may require one level of review. A legal research citation, expert summary, client-facing advice, or court filing requires another. The point is not to overbuild bureaucracy. The point is to make the review expectation explicit before a mistake forces the conversation.
Fourth, who is responsible for that review? AI does not hold a license, sign a pleading, owe duties to clients, or stand before a court. Responsibility remains with lawyers, supervisors, and ultimately firm leadership. If the firm permits AI-assisted work but has not clarified responsibility, the firm has created ambiguity at precisely the point where professional duties require clarity.
These are not technology questions. They are supervision questions. They are quality-control questions. They are professional judgment questions. In a mid-size law firm, they are leadership questions.
Responsible Adoption Requires a Firm Position
The firm that has not addressed these issues has not avoided AI risk. It has simply allowed AI decisions to be made by default at the individual level. One lawyer experiments with a drafting tool. A staff member uses an AI summarizer. A vendor adds AI features to software the firm already uses. A partner assumes the firm has a policy because the issue was mentioned in a meeting. Over time, practice forms without leadership having made a firm decision. That is adoption by default.
Responsible adoption is different. It does not require panic, prohibition, or a massive technology initiative. It requires a firm position. Leadership should be able to explain what AI use the firm allows, what it restricts, what it prohibits, what must be reviewed, and why those decisions fit the firm's professional duties, client obligations, risk tolerance, and standards of work.
That position does not need to answer every future question. AI tools will continue to change. Courts, clients, carriers, and bar authorities will continue to refine expectations. But a firm that has made deliberate decisions is in a very different posture from a firm that discovers its AI practices only after an error appears in a filing, a client asks a direct question, or a court imposes a disclosure obligation.
Mata was the first famous warning. Three years later, the profession is no longer at the stage of surprise. Courts and ethics authorities are moving toward accountability for AI-assisted legal work. The question is whether law firm leadership is moving with the same seriousness.
Responsible adoption begins when firm leadership makes the review, supervision, and governance decisions before unsupervised AI-assisted work becomes ordinary practice.
------------------------------
For managing partners who want to know whether their firm’s AI posture is intentional or developing by default, I have prepared a short self-assessment for law firm leadership. No email required. Click here for the Managing Partner Self-Assessment on AI Adoption
Rex C. Anderson
The retired attorney law firms call about AI Adoption
Forwarded to you?
Subscribe for direct delivery of future articles.