Living page — last updated June 2026. The list of AI-related sanctions and judicial commentary is growing. This page covers documented cases as of the last update; check the ABA Journal and your state bar's recent communications for new developments.

In 2023, a federal district court in New York sanctioned attorneys who submitted a brief citing cases that did not exist. The cases had plausible names, plausible citations, plausible holdings. None of them were real. A large language model had generated them, and no one on the legal team had verified them before the brief was filed.

The sanctions were modest by dollar terms. The reputational consequence was not. The case became the reference point in virtually every judicial instruction, bar opinion, and law school ethics class on AI since.

This is a guide to what has happened when AI-generated content went into court filings unverified, what courts have said about it, and the verification workflow that prevents it.

What hallucinations are and why they produce fake citations

Generative AI models do not retrieve information from a database. They predict the next statistically likely token given everything that came before it — in a conversation, in their training data, in the structure of the legal context you provided. When you ask a large language model to find cases supporting a legal argument, it generates text that looks like case citations because that is the pattern the context suggests.

The model has no access to Westlaw or Lexis. It has no internal ledger of real cases. It produces text that fits the form of a legal citation — and sometimes that text corresponds to a real case, and sometimes it does not. The model cannot tell the difference, and it does not signal uncertainty in a way that makes the fabricated citation look different from a real one.

The technical term is hallucination. The legal consequence is citing a non-existent authority in a court filing — a potential violation of Rule 3.3(a)(1) (candor toward the tribunal) and, depending on the facts, Rule 8.4(c) (conduct involving dishonesty, fraud, deceit, or misrepresentation).

The foundational case: Mata v. Avianca (S.D.N.Y. 2023)

The case that gave this issue its name involved an attorney who used ChatGPT to research cases for a personal injury brief in federal court in New York. The brief cited approximately six cases that did not exist. When opposing counsel could not locate the cases, the court inquired. The attorney confirmed they had used ChatGPT and had not verified the citations.

Judge P. Kevin Castel sanctioned the attorneys involved — both the filing attorney and the supervising partner — as well as the law firm. The order made findings about each attorney's conduct and the failure of any member of the team to verify that the cited cases were real before filing.

The court's opinion is worth reading in full. Its core holding on the professional responsibility question: the duty to verify legal citations is non-delegable and is not discharged by relying on any tool, including AI. The fact that an attorney did not type the fake citation themselves — that a machine generated it — did not reduce the attorney's responsibility for submitting it.

Several things made Mata particularly instructive beyond the sanctions order itself:

When the court asked the attorneys whether the cases existed, they initially maintained that they did — and submitted AI-generated summaries of those cases as confirmation. This compounded the original failure.
The attorneys had not used an AI tool they were familiar with and had evaluated; one filed a declaration indicating limited understanding of how the tool worked.
No one on the team had run the cited cases through Westlaw, Lexis, or any other citator before filing.

Warning

The attorneys in Mata did not intend to deceive the court. They had used a tool they did not adequately understand, trusted its output, and failed to build verification into their workflow. Sanctions in AI filing cases since 2023 have consistently reached attorneys who acted negligently, not just those who acted knowingly.

What courts have said about it

Several judicial statements across these cases are worth understanding as a framework:

The professional responsibility duty is non-delegable. Courts have consistently held that an attorney cannot point to a tool — of any kind — as the reason they failed to verify content they signed. The attorney's signature certifies the document; the certification obligation applies to everything in it.

Unfamiliarity with the tool is not a defense. Attorneys who indicated they did not understand how the AI tool worked, or had not used it before, have not received more lenient treatment for that reason. The competence duty runs before, during, and after tool use.

The post-discovery response matters. In Mata, the attorneys' responses after the court inquired about the cases — initially maintaining the citations were real, then submitting AI-generated confirmation — were treated as compounding the original error. When you discover that AI-generated content in a filing is wrong, transparent disclosure to the court is the required response.

This is not unique to ChatGPT or any specific tool. Courts have not drawn distinctions based on which AI tool was used. The issue is verification failure, not tool selection.

The verification workflow

The professional duty is clear: every legal citation in a court filing must be independently verified before filing. The following workflow meets that duty:

Before you draft:

Establish which research tools you will use for verification. Westlaw and Lexis are the most commonly accepted citators. You need access to at least one before you use AI for research.

During AI-assisted research:

Treat all AI-generated citations as candidate cases, not confirmed authorities. Write them down, note the holding the AI attributes to them, and flag them for verification.
Do not copy citations from AI output directly into a brief. Copy them into a verification list first.

Before any brief is filed:

Run every cited case through your citator. Confirm the case exists, was decided by the court cited, says what you attribute to it, and is still good law.
Run every statutory and regulatory citation through the current official code or register.
If any citation does not check out, remove it — do not substitute an AI-generated correction without also verifying the substitute.

At signature:

FRCP 11 and its state equivalents certify that representations to the court are based on a reasonable inquiry and are not knowingly false. Your signature means verification happened. If it did not, do not sign.

Tip

If you are using AI for legal research, use it for what it is genuinely good at — identifying the approximate shape of the relevant case law, generating the vocabulary of a legal argument, and surfacing leads to follow up in a real citator. The research is not done until the citator run is done.

What this means for your practice

The practical change is narrow but non-negotiable: add a verification step, and never skip it.

You do not need to stop using AI for legal research. AI can be a useful first-pass tool — for identifying case law themes, generating search terms you might not have thought of, finding the approximate shape of an argument before you go to Westlaw to develop it properly. What you cannot do is take the AI's citations at face value and put them in a brief.

For a solo or small firm, the verification step is the same step you have always been required to take before citing a case: confirm it exists, confirm it says what you say it says, confirm it is still good law. The change is only that AI adds a new source of candidate citations that did not exist before — candidates that require the same verification as any other source.

For a look at AI research tools and how the better ones handle citation accuracy, see AI legal research tools compared. For the broader AI ethics framework, see ABA Formal Opinion 512 explained.

AI Hallucinations in Legal Filings: When Courts Sanction Lawyers