The hallucinated citation and the duty of competence

In 2023, a New York lawyer named Steven Schwartz filed a federal court brief citing six cases. None of them existed. He had asked ChatGPT to research the law, and the chatbot invented plausible-sounding decisions with fake quotes and fake internal citations. When opposing counsel could not find the cases, Schwartz went back to ChatGPT and asked if they were real. The tool said yes. They were not.

The case was *Mata v. Avianca*, and it became the cautionary tale every legal AI training now opens with. Judge P. Kevin Castel sanctioned the lawyers and their firm. The lesson was not "AI is dangerous." The lesson was narrower and sharper: the tool does not sign the brief. You do.

What actually happened in Mata v. Avianca

The underlying dispute was mundane: a passenger sued the airline Avianca over a knee injury from a serving cart. The airline moved to dismiss on statute-of-limitations grounds. Schwartz, whose firm lacked federal-court access to the airline's preferred database, turned to ChatGPT to find cases supporting his client's position.

ChatGPT produced them. Cases like "Varghese v. China Southern Airlines" with full citations and reasoning. All fabricated.

This is called a hallucinationhallucinationA hallucination is when an AI model generates output that is fluent and confident but factually wrong, fabricated, or unsupported by its source data.View full definition →: when a generative AI model produces confident, fluent output that is factually false. Large language models predict the next likely word, not the true fact. A citation that "looks right" statistically can be entirely invented.

Judge Castel imposed a $5,000 sanction on the lawyers and required them to notify each judge falsely named as the author of a fake opinion. You can read the actual sanctions order, which is short and readable, here on CourtListener.

The court was blunt: the problem was not using a new tool. The problem was submitting fake material and then defending it without checking.

The duty of competence: your foundation

Every US lawyer is bound by rules of professional conduct, modeled on the American Bar Association (ABA) Model Rules. These are the ethical standards adopted, with variation, by each state bar.

Rule 1.1 (Competence) requires a lawyer to provide competent representation, meaning the legal knowledge, skill, thoroughness, and preparation reasonably necessary.

In 2012 the ABA added Comment 8 to Rule 1.1. It says competence includes keeping abreast of "the benefits and risks associated with relevant technology." This is often called the duty of technology competence.

Read that carefully. It cuts both ways.

• You cannot ignore useful technology out of ignorance.
• You cannot deploy technology whose risks you do not understand.

A lawyer using a generative AI tool has an affirmative duty to understand that the tool can fabricate citations. "I did not know it could do that" is not a defense. It is the violation.

You can read the ABA Model Rules of Professional Conduct here.

Why the partner is answerable, not the tool

Here is where junior lawyers and non-lawyers often get confused. The AI made the mistake. Why is the human on the hook?

Because a tool has no license to lose. Responsibility flows through the person who signed.

Rule 11 and the signature

In federal court, Federal Rule of Civil Procedure 11 governs signatures on filings. When a lawyer signs a document, they certify that the legal contentions are warranted by existing law and that they conducted a reasonable inquiry.

A "reasonable inquiry" means you checked. Citing six cases you never read, because a chatbot listed them, is the opposite of reasonable inquiry. The signature is a personal warranty. AI cannot make it for you.

Rules 5.1 and 5.3: supervision

Two more rules pin responsibility to senior lawyers.

Rule 5.1 (Supervision of Lawyers) makes partners and managers responsible for ensuring their firm and their subordinate lawyers follow the ethics rules.

Rule 5.3 (Responsibilities Regarding Nonlawyer Assistance) originally covered paralegals and secretaries. Bar authorities now widely treat AI tools as "nonlawyer assistance." That means the supervising lawyer is responsible for the AI's work product the same way they are responsible for a paralegal's.

Put those together. If an associate uses AI, the associate is responsible under Rule 1.1. The supervising partner is responsible under Rule 5.1 for the associate. And under Rule 5.3, the firm is responsible for the tool itself.

There is no seat at the table where the AI takes the blame. The blame has nowhere to go but up the human chain.

The concrete workflow that prevents a Mata

Competence in 2026 is not "avoid AI." Courts and bar associations now generally accept AI-assisted work. The standard is verification. Here is a practical protocol firms are adopting.

1. Never cite what you have not read. Every case a tool surfaces gets pulled and read in a primary source: Westlaw, Lexis, or the free Google Scholar case law search. If the case does not exist there, it does not exist.

2. Check the citation, the holding, and the quote separately. A real case can be cited for a proposition it never stood for. Verify the case exists, then verify it actually says what you claim.

3. Prefer closed, grounded tools for legal research. General chatbots invent freely. Purpose-built legal research tools that use retrieval-augmented generation (RAG, meaning the AI answers only from a verified database of real cases) hallucinate far less. But even these are not perfect. Independent testing has shown even legal-specific tools produce errors, so verification still applies.

4. Log your prompts and outputs. If a filing is challenged, you want a record of what you asked and what the tool returned.

Here is what a verification checklist looks like as simple pseudocode. This is a mental model, not software you must run.

for each citation in brief:
    case = pull_from_primary_source(citation)   # Westlaw, Lexis, Google Scholar
    if case is None:
        REJECT("citation may be hallucinated")
    if not case.supports(claimed_proposition):
        REJECT("case exists but does not stand for this")
    if quoted_text not in case.full_text:
        REJECT("quote fabricated or altered")
    mark_verified(citation)

If any citation fails, it does not go in the brief. Full stop.

What courts expect now

Since *Mata*, dozens of judges have issued standing orders (rules a specific judge applies to all cases before them) requiring lawyers to disclose AI use or to certify that all citations were human-verified.

The trend is not toward banning AI. It is toward demanding accountability. Some courts require a specific certification. Others simply enforce Rule 11 more aggressively, knowing hallucinated citations are now a known risk that competent counsel must screen for.

This changes the ignorance calculation. In 2023, a judge might have viewed AI hallucinationAI hallucinationA hallucination is when an AI model generates output that is fluent and confident but factually wrong, fabricated, or unsupported by its source data.View full definition → as a novel surprise. In 2026, it is a documented, publicized, foreseeable risk. A lawyer who files a fabricated citation today has a weaker excuse than Schwartz did, because the risk is now common knowledge in the profession.

The client dimension

Two more rules deserve a mention.

Rule 1.6 (Confidentiality) matters because typing client facts into a public chatbot may expose confidential information to a third party. Many general AI tools use inputs to train future models. Feeding privileged client details into such a tool can itself be an ethics breach, separate from any hallucinationhallucinationA hallucination is when an AI model generates output that is fluent and confident but factually wrong, fabricated, or unsupported by its source data.View full definition →.

Rule 1.5 (Fees) is emerging as a live question. If AI makes research faster, can a lawyer bill the hours it would have taken manually? The general answer is no: you bill for value and time actually spent, not for time saved by automation.

Key Takeaways

• The tool never signs the brief. Under Rule 11 and ABA Rules 1.1, 5.1, and 5.3, responsibility for every citation flows to the human who signed and to the partner who supervised. AI cannot absorb blame.

• Hallucination is a foreseeable risk, not a surprise. After *Mata v. Avianca*, "I did not know AI could invent cases" is an admission of incompetence under the duty of technology competence, not a defense.

• Verify in a primary source, always. Pull every AI-surfaced case from Westlaw, Lexis, or Google Scholar. Confirm the case exists, that it stands for your proposition, and that any quote is real.

• Protect confidentiality and bill honestly. Do not feed privileged client facts into public chatbots (Rule 1.6), and bill for time actually spent, not time AI saved you (Rule 1.5).

• Competence in 2026 means using AI well, not avoiding it. The professional standard is disciplined verification and clear supervision, not fear.