How to avoid AI-hallucinated citations in court filings

Lawyers file AI-hallucinated citations when a generative AI tool invents a case or quotation and nobody pulls the real authority and reads it before the filing is signed. Courts respond with fines, letters to clients and judges, public reprimands, disqualification and referral to the bar. The fix is a fixed routine: every cited authority gets pulled from a research database and read by a named person before anyone signs.

How common is this?

As of November 2, 2026, the AI Hallucination Cases Database records 897 court and tribunal decisions addressing AI-hallucinated material from lawyers, such as invented case citations and false quotations. 603 of them are in the US. The first lawyer decision it records is dated April 14, 2023, and it records lawyer decisions in every month since then through October 2026.

The pace has picked up. The database counts 25 such lawyer decisions in 2024, 359 in 2025 and 507 in the first ten months of 2026. It records a monetary sanction or costs order in 265 of the 897 decisions and a professional sanction in 161. These figures change as the database grows. The Record page has current figures and the breakdown by country, state and month.

How do fake citations end up in a filing?

The cases show a few repeated patterns.

The tool was treated as a research database. In Mata v. Avianca, the court found the lawyers submitted non-existent opinions, with fake quotes and citations, that ChatGPT had created, then kept standing by them after the court questioned whether they existed (page 1). The lawyer who did the research asked ChatGPT “Is Varghese a real case”, and it answered that the authorities were real and could be found on Westlaw and LexisNexis (page 19). Asking the tool to check itself is not a check.

Someone else added the citations. In Johnson v. Dunn, a supervising partner added the citations while revising another lawyer’s drafts. The lawyer who drafted, signed and filed the motions said he reviewed the edits for grammar and typos but did not conduct an independent review of the legal authorities added (page 8). The firm, Butler Snow, already had AI rules. When it gave its lawyers an AI research tool in January 2025, it adopted a policy that “[a]ll outputs must be reviewed and verified by the responsible attorney” before filing (page 21). The policy did not stop it.

The tool was built for lawyers. In Wadsworth v. Walmart, a lawyer uploaded a draft to MX2.law, which the court said appears to be an in-house database launched by the firm of the lawyer and his supervisor, Morgan & Morgan, and asked it to add case law (page 4). The motions cited nine cases, but eight did not exist (page 3).

Quotations were made up inside real cases. In Noland v. Land of the Free, the California Court of Appeal found that the opening brief contains 23 case quotations, 21 of which are fabrications (page 10). Many cited cases were real but did not discuss the topics they were cited for. The court said the errors went undetected because counsel “did not read the cases the AI tools cited” (page 2).

What have courts done about it?

Case	Court and date	What went wrong	What the court did
Mata v. Avianca	S.D.N.Y., June 22, 2023	ChatGPT-generated opinions, defended after the court questioned them	$5,000 penalty; letters to the client and to each judge falsely named as an author (page 34)
Wadsworth v. Walmart	D. Wyo., Feb. 24, 2025	Eight of nine cited cases did not exist	Drafting lawyer’s pro hac vice admission revoked and $3,000 fine; $1,000 fine each for two co-signers (page 14)
Johnson v. Dunn	N.D. Ala., July 23, 2025	Five problematic citations across two motions	Public reprimand, published order, disqualification from the case and referral to the Alabama State Bar for three lawyers (page 2)
Noland v. Land of the Free	Cal. Ct. App., Sept. 12, 2025	21 of 23 quotations fabricated	$10,000 sanction; counsel to serve the opinion on the client; clerk to send it to the State Bar (pages 2 to 3, 30)

The direction is toward harsher sanctions. In Johnson, the judge wrote that fabricated authority demands substantially greater accountability than the reprimands and modest fines that have become common (page 1). The court also noted that the client, a state agency, had kept the lawyer on after learning of the misconduct (page 2).

Not checking your opponent’s citations can also cost you. In Noland, the court declined to order sanctions payable to opposing counsel, partly because they did not alert the court to the fabricated citations and appeared to learn of them only from the court’s order to show cause (page 30).

Which rules apply?

Rule 11. In federal court, signing or filing a paper certifies that, after “an inquiry reasonable under the circumstances,” the legal contentions are warranted by existing law or by a nonfrivolous argument for changing it (Rule 11(b)(2)). A sanction must be limited to what suffices to deter repetition of the conduct and can include nonmonetary directives, a penalty paid into court or fees (Rule 11(c)(4)). The Mata court put it this way: existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings (page 1).

Signing is the trigger, not who used the tool. The Wadsworth court fined two lawyers who did not draft the motions because they signed without checking. Its advice: make a reasonable inquiry into the law before signing (or giving another permission to sign) a document (page 10).

Rule 11 has gaps. It does not apply to disclosures and discovery requests, responses, objections, and motions under Rules 26 through 37 (Rule 11(d)), which is why the Johnson court said Rule 11 “does not apply” there. Finding that its local rule did not clearly reach the conduct either, it sanctioned the lawyers under its inherent authority (pages 31 to 32). State courts apply their own rules. The Noland court cited California’s frivolous-appeal statute and its Rules of Court, which allow sanctions for unreasonable rule violations (pages 19 to 21).

Rule 3.3, candor toward the tribunal. State professional conduct rules set the duty of candor, and the wording differs from state to state. Alabama’s says a lawyer shall not knowingly make “a false statement of material fact or law to a tribunal.” (quoted in Johnson, page 28). The Johnson court declined to apply it, because Rule 3.3 “forbids only knowing misstatements of law” and these lawyers had not verified the citations at all (page 31). New York’s version, quoted in Mata, also bars a lawyer from failing to correct “a false statement of material fact or law previously made to the tribunal” (page 23). Once you know a citation is fake, the duty to fix it is direct. For the ABA’s ethics guidance on candor and AI, see our guide to ABA Formal Opinion 512.

What checking routine works for a small firm?

The California Court of Appeal set the standard plainly: no filing should contain citations that the attorney responsible for submitting the pleading has not personally read and verified (pages 2 to 3). The steps below are our own routine for meeting that standard. No court or rule prescribes them. They apply to every filing, whether or not anyone used AI, because you cannot always tell where a citation came from.

1. Freeze the draft and list every authority. Build a list of every case, statute, rule and quotation, including string cites and footnotes. Anything added after this point goes back through the list.
2. Pull each authority from a research source, not from the AI tool. Open it on Westlaw, Lexis, Fastcase, the court’s website or another official source. If you cannot find it, treat it as fake until proven otherwise. Do not ask the AI tool whether the case is real. That is what failed in Mata.
3. Match the citation details. Case name, volume, reporter, page, court and year must all match the opinion you pulled.
4. Read the cited page. Confirm the case says what the brief says it says. Noland’s brief cited real cases for points they never discussed.
5. Match every quotation word for word. Search the opinion for the exact quoted words. A quotation that cannot be found gets cut or replaced.
6. Check subsequent history. Run each case through a citator to confirm it is still good law.
7. Check every edit, from anyone. Citations added by a partner, co-counsel or a reviewer get the same check. In Johnson, the gap was a supervisor’s revisions that nobody re-checked.
8. Record who checked. Keep a short log in the file: authority, source pulled from, checker’s initials, date. The lawyer who signs reviews the log before signing. In Wadsworth, lawyers who signed without checking the law were fined.
9. Check the other side’s citations too. Run the same check on authorities in the opposing brief that matter to your argument, and tell the court if one is fake.
10. Correct fast if something slips through. Notify the court and withdraw or correct the filing as soon as you find the problem. In Mata, standing by the fake opinions after the court questioned them made things worse.

A firm policy alone is not enough. Butler Snow had one. What protects a firm is a check that someone has to do and sign off on before every filing.