AI-Fabricated Citation Cases Every Litigator Should Know

There is now a short, named line of cases that every litigator can be expected to know. Each one began the same way: a lawyer filed a brief citing authority that did not exist, and a court found out. The fact pattern is consistent enough that you can build a defense against it, and the defense is not a tool. It is a written verification protocol.

This piece walks through the cases by name, identifies the single root cause they share, and gives you a protocol you can adopt this week.

Mata v. Avianca: the case that named the problem

The reference point for every conversation about AI hallucination sanctions cases is Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023). Plaintiff's counsel, opposing a motion to dismiss a personal-injury claim, filed a brief built on judicial opinions that ChatGPT had invented, complete with fabricated quotations and internal citations to nonexistent decisions. When opposing counsel and the court could not locate the cases, the lawyer asked ChatGPT whether the cases were real, and the tool assured him they were. They were not.

On June 22, 2023, Judge P. Kevin Castel imposed a $5,000 sanction under Federal Rule of Civil Procedure 11, jointly against the two attorneys and their firm. The court was careful: the sanction was not for using AI. It was for what the lawyers did after the brief was questioned. Judge Castel found subjective bad faith in the failure to check the cases once their authenticity was in doubt, and described one of the AI-generated opinion summaries as "gibberish." This is the foundational entry in the genre of fake citations ChatGPT lawyer stories, and it is worth reading in full because the opinion is measured, not gleeful.

The lesson lawyers took away was narrow but correct: the problem was not the software. It was the absence of any independent confirmation that the cited authorities existed.

Park v. Kim: the appellate court raises the floor

If Mata could be dismissed as a single careless episode, Park v. Kim, 91 F.4th 610 (2d Cir. 2024), closed that escape route. Decided January 30, 2024, the Second Circuit addressed a reply brief in which counsel, Attorney Jae S. Lee, cited a case she later admitted she had generated with ChatGPT and had not read.

The court did not impose a monetary fine itself. Instead it referred the attorney to the court's Grievance Panel under Local Rule 46.2 for further investigation and possible discipline, and it published the referral. The reasoning is the part to remember. The panel held that Rule 11's reasonable-inquiry obligation, at a minimum, requires that "attorneys read, and thereby confirm the existence and validity of, the legal authorities on which they rely." That sentence is now quoted in standing orders and CLE materials across the country. It frames verification not as best practice but as a baseline duty an appellate court is willing to enforce through discipline rather than a one-time fine.

The Michael Cohen near-miss: why intent still matters

Not every fabricated-citation episode ends in a sanction, and the contrast is instructive. In a 2024 matter before Judge Jesse M. Furman in the Southern District of New York, Michael Cohen's motion for early termination of supervised release cited three cases that did not exist. Cohen had found them using Google Bard, which he said in a sworn declaration he believed was a "supercharged" search engine rather than a generative tool that produces plausible-looking but invented text. His attorney, David Schwartz, had passed the citations into the filing without independently confirming them.

Judge Furman declined to impose sanctions. The court reasoned that Cohen had no motive to submit citations he knew were false, given the obvious adverse consequences, and that the record did not support a finding of bad faith. The near-miss is not a safe harbor. Read alongside Mata, it tells you that bad faith can be the difference between a written rebuke and a five-figure sanction, and that the lawyer who signs the filing, not the client who supplied the research, owns the Rule 11 duty. Schwartz avoided a sanction here, but he is the cautionary figure: he filed authority he had not checked.

Wadsworth v. Walmart: the in-house tool does not save you

By 2025 the cases had moved past solo practitioners and free chatbots. Wadsworth v. Walmart, Inc., No. 2:23-cv-118 (D. Wyo. Feb. 24, 2025), involved attorneys from Morgan & Morgan, one of the largest plaintiff-side firms in the country. Their motions in limine cited nine cases. Eight did not exist. They had been produced by the firm's own internal AI platform, not by a public consumer tool.

The court sanctioned the lawyer who prepared the motion, Rudwin Ayala, in the amount of $3,000 and revoked his pro hac vice admission in the case. Two other attorneys whose names appeared on the filing, including T. Michael Morgan, were each fined $1,000, bringing the total to roughly $5,000. The firm itself was not separately sanctioned, in part because it had already moved to require independent verification of AI-assisted research going forward.

Wadsworth dismantles the most common rationalization in firm AI conversations: that a private, in-house, or "legal-grade" tool is safe in a way that ChatGPT is not. It is not. A model trained or marketed for law can still fabricate citations, and the signature block, not the vendor, carries the obligation. The revoked pro hac vice admission is the detail that tends to focus managing partners, because it reaches the ability to appear in the matter at all.

Noland v. Land of the Free: a published appellate warning

The most recent marker is Noland v. Land of the Free, L.P. (Cal. Ct. App., 2d Dist., Div. 3, Sept. 12, 2025), California's first published appellate opinion addressing AI-fabricated citations. The numbers are stark. Of the 23 case quotations in the appellant's opening brief, 21 were fabricated, invented by generative AI tools and never appearing in any real decision.

The court affirmed summary judgment against the appellant, sanctioned her counsel $10,000, referred him to the State Bar of California, and ordered the opinion served on the client. Because it is published, it is citable precedent in California and a template other state courts will look to. Its central holding reads like a protocol in one sentence: no brief, pleading, motion, or other paper should contain any citation, "whether provided by generative AI or any other source," that the responsible attorney "has not personally read and verified."

That is the entire doctrine of this line of cases, compressed into a single duty.

The common root cause, stated plainly

Across Mata, Park, the Cohen matter, Wadsworth, and Noland, the tool varied (ChatGPT, Google Bard, an in-house platform, unspecified generative systems) and the posture varied (trial court, appellate court, sanction, referral, near-miss). The failure did not vary. In every instance a lawyer relied on cited authority that no human had independently confirmed against a primary source before the document was filed.

This is not a technology problem that better software will solve. Generative models produce fluent text that predicts what a citation should look like; producing a real one is incidental, not guaranteed. Courts understand this now, which is why they frame the obligation in terms lawyers already owe under Rule 11 and their candor duties to the tribunal. ABA Formal Opinion 512, issued July 29, 2024, says the same thing through the lens of competence (Model Rule 1.1) and candor (Model Rule 3.3): a lawyer must guard against AI hallucinations forming the basis of representations to a court.

The procedural environment has hardened in parallel. Beginning with Judge Brantley Starr of the Northern District of Texas, whose May 30, 2023 standing order required attorneys to certify either that no part of a filing was drafted by generative AI or that any AI-drafted language was checked for accuracy by a human, individual judges across federal and state courts have adopted AI certification and disclosure standing orders. A widely cited database maintained by researcher Damien Charlotin now catalogs hundreds of decisions worldwide in which courts have addressed hallucinated citations, the large majority of them from 2025. The count is growing, and "I did not know it could do that" is no longer available as a defense.

The fix: a written verification protocol

The cases reward firms that can show a process and punish those who cannot. A protocol does two things at once: it prevents the filing, and if something slips through, it demonstrates the reasonable inquiry that distinguishes Mata from the Cohen outcome. Put it in writing, make it firm policy, and apply it to every AI-assisted output regardless of which tool produced it.

A workable protocol for a 3-to-30-attorney firm:

• Pull every cited authority from a primary source. Open each case, statute, and regulation in Westlaw, Lexis, the official reporter, or the issuing court's site. Confirm the case exists, the citation is correct, and it remains good law.
• Verify every quotation against the actual opinion. A real case can still be paired with an invented quote or a holding it does not support. Match quoted language to the page.
• Assign the verification to the signing lawyer. The signature block carries the Rule 11 duty. Wadsworth shows that names on a filing are accountable even when someone else drafted it.
• Treat in-house and "legal" AI tools exactly like consumer chatbots. No vendor's product is exempt from the read-and-verify step.
• Check your judge's standing orders and local rules before filing. Disclosure or certification requirements vary by courtroom and change frequently.
• Log the verification. A short note in the file that authorities were checked and by whom is cheap insurance and evidence of reasonable inquiry.

None of this prohibits using AI. Each of these courts went out of its way to say the technology is permissible. What is not permissible is filing its output unread. The firms that will avoid joining this list are the ones that write the rule down, train to it, and can prove they followed it.

This is general information for lawyers and law-firm leaders, not legal advice, and it does not create an attorney-client relationship. The authorities are cited so you can read them yourself.