You Found a Fake Citation in a Filed Brief. Here Is What to Do in the Next 48 Hours

Two lawyers, the same mistake, two outcomes. In February 2026 the Fifth Circuit fined a lawyer $2,500 for fabricated citations and added that had he “accepted responsibility and been more forthcoming, it is likely that the court would have imposed lesser sanctions”. A month later the Sixth Circuit fined two lawyers $15,000 each, plus their opponents’ fees and double costs, after they called its show-cause order “void on its face” and left the court’s question about AI unanswered.

So the first thing to know about a fake citation in a filed brief is that the next 48 hours matter more than the last 48. The error is already in the record. What is not yet in the record is how you respond, and every sanctions opinion since Mata v. Avianca has been written about that.

This is the checklist. Print it, then read the sanctions timeline later, when your hands have stopped shaking.

Hour 0: what to do the moment you find a fake citation in a filed brief

Stop. Do not delete the chat, the draft or the email that brought the problem to your attention. Do not re-file a “clean” version and hope nobody compares them.

Hours 1 to 4: verify every other citation in the filing

One fake citation is rarely alone. In Noland, 21 of 23 quotations in the opening brief were fabricated; in Couvrette, 15 fake citations and eight fabricated quotations sat across three briefs. The court will assume the rest of your filing is suspect, so you must know before it does.

Run the six-layer check on every authority: it exists; names, court, year and reporter match; a citator shows it is good law; the pinpoint says what your sentence says; every quotation matches character for character; jurisdiction and posture fit. A model may build the list; a human does the checking. Never ask the tool whether its own citations are real; the Illinois Appellate Court held that cross-referencing AI output against a research platform is insufficient unless the lawyer confirms the cited text actually appears in the source.

List every case, statute, rule and secondary source cited in the filing below in a table: Citation as written | Sentence it supports (quoted) | Pinpoint given? | Quotation present? | Page of the filing.
Do not tell me whether any citation exists or is accurate; I will check each one in a primary database myself. Do not correct, add or reformat anything. Add a final column headed "Checked by / database / date" and leave it blank.

Filing:
[paste]

Butler Snow, after Johnson v. Dunn, had Morgan Lewis check more than 2,400 citations across 330 filings; none were wrong, and the judge credited it. You are not doing that in four hours, but you are doing this document, and any other filing the same person drafted with the same tool. The comparison of citation checkers tells you which tools speed up the existence layer; none of them replaces reading.

Rule 3.3 and Rule 11: what candour requires

ABA Formal Opinion 512 applies Rules 1.1, 1.4, 1.5, 1.6, 3.1, 3.3, 5.1 and 5.3 to generative AI without dilution; the plain-English summary is on this site. Rule 3.3 requires correcting a false statement of law made to a tribunal. Rule 11, per the Second Circuit in Park v. Kim, requires “that attorneys read, and thereby confirm the existence and validity of, the legal authorities on which they rely”. The California Court of Appeal in Noland wrote the sentence the Sixth Circuit now quotes: no paper filed in any court should contain a citation “that the attorney responsible for submitting the pleading has not personally read and verified”.

Two more lines set the standard. The Fifth Circuit, declining to adopt an AI rule in June 2024: “‘I used AI’ will not be an excuse for an otherwise sanctionable offense.” The Illinois Appellate Court in Scott, July 2026: “The only acceptable standard is zero false citations.”

England is not gentler. The Divisional Court in Ayinde listed its powers as public admonition, costs, wasted costs, strike-out, regulator referral, contempt and referral to the police, and the SRA’s August 2026 warning notice says that “reliance on an output of AI would not be a suitable defence”. The Upper Tribunal’s judicial review form now demands a statement of truth that every authority exists, can be located by its citation and supports the proposition.

How to notify: notice of errata, withdrawal, corrected brief

The vehicle depends on the court and the stage; the content does not. Say what is wrong, withdraw it, say how it happened, say what you have done since, and stop. Do not argue that the point survives without the authority; that can come later, in a properly supported filing.

Draft a notice of errata and corrected citations for filing in [court] in [case], to be served on opposing counsel the same day. Facts: the [motion/brief] filed on [date] cited [N] authorities that do not exist or do not support the propositions ([list]). The draft was prepared with [tool] and the citations were not verified before filing by [me/the signing lawyer].
The notice must: identify each affected citation and the page; withdraw it; state that the error is ours and not the tool's or any junior's; state the verification steps taken since ([describe]) and confirm every remaining citation has been checked in [database] by [name]; apologise once, without qualification; and ask for no relief other than that the court disregard the withdrawn authorities. Under 400 words. Formal, plain, no adjectives.

Expect to appear. In Ontario, Justice Myers ordered the lawyer in Ko v. Li to show cause for contempt; the matter was discontinued only after she admitted that the factum “had indeed been prepared by her staff in part with the use of generative AI, namely ChatGPT”, apologised and undertook continuing-education courses, the court noting that “public shaming near the end of an unblemished career” served deterrence (Cassels summary). In Victoria, a solicitor whose AI case list the judge and her associates could not find gave an “unconditional apology”, paid wasted costs and still lost his principal practising certificate, with two years of supervised practice on top. Judge Castel refused to order an apology in Mata “because a compelled apology is not a sincere apology”. Make yours before it is compelled.

Notify the client (Rule 1.4) and the insurer

The client finds out either way. In Park v. Kim the Second Circuit ordered the lawyer to give the decision to her client; in Johnson v. Dunn the three partners were ordered to send the sanctions order to every client, opposing counsel and judge in their pending cases. Opinion 512 makes disclosure mandatory when the client asks, when client information is put into the tool and when AI use is relevant to the fee, and otherwise calls it a fact-specific judgement; a sanctions hearing in the client’s own case is not a close call; the disclosure guide covers the jurisdictional variations.

Draft a letter to [client] in [matter] reporting that a filing made on our behalf on [date] contained [N] citations that were fabricated or inaccurate, that we discovered this on [date], and what we have done: [notice of errata filed; opposing counsel informed; all remaining citations verified; hearing set for date]. Explain in plain English what a sanctions hearing is and the realistic range of consequences for the case (not for us), without speculation. State that the client will not be billed for any time spent correcting the error. Invite questions and offer a call. No jargon, no defensiveness, under 350 words.

Then the insurer. ALPS’s Chris Newbold told the Daily Record in August 2026 that carriers are “not seeing a lot of what I would call direct AI claims at this point”, and no major US professional-liability writer has filed an explicit AI exclusion; but underwriters have begun asking “Do you police it?”, and broker Lee Norcross’s line is the one to remember: “AI does not have a law license to lose” (Daily Record). A sanction order or fee-shifting award is a circumstance to notify; late notice is how coverage is lost. The malpractice insurance guide lists what carriers now ask for at renewal.

The excuses judges have catalogued (and rejected)

In September 2025, 404 Media asked 18 sanctioned lawyers why they did it. The answers, collected here, are a catalogue of what not to say. Add the excuses the courts themselves recorded and you have the table below. One Hawaii lawyer, fined $100, told the outlet that nearly every lawyer was using AI to some degree and the problem was only getting caught; that is the attitude the Sixth Circuit priced at $15,000 a head.

What judges have rewarded: prompt self-correction

The reverse pattern is just as consistent. The Queens Civil Court in Idehen v. Stoute-Phillip (July 2025) said that had the lawyer “simply come to the hearing acknowledging that he had used Copilot to conduct his research, was unaware that the software could produce fake cases, and apologized for his mistake, the Court would have likely determined that a sanction was sufficient” (Sterne Kessler’s 2025 review). A Massachusetts judge imposing $2,000 in a family case in early 2024 is reported to have said there is nothing wrong with using reliable AI technology, provided the output is verified. Professor Maura Grossman’s summary after Coomer: “You are likely to get a harsher penalty if you don’t come clean.”

If a junior or contractor drafted it: supervision under 5.1 and 5.3

“Someone else wrote it” is not a defence; it is an admission of a second breach. Judge Rankin: “Blind reliance on another attorney can be an improper delegation of this duty and a violation of Rule 11.” Judge Aycock in Withers disqualified local counsel who had let signatures be affixed unread, calling them “a prime example of the risk associated with serving as a rubberstamp”. In Lnu, the briefs had been written by unlicensed law graduates and filed under the lawyers’ names without the citations being read; in Dehghani v. Castro, an immigration lawyer bought the brief from a freelancer. Singapore’s Tan Hai Peng v Tan Cheong Joo ordered personal costs against both the junior and the supervising partner.

Opinion 512 requires managerial lawyers to “establish clear policies regarding the law firm’s permissible use of GAI” and supervisors to make reasonable efforts to ensure compliance. The Upper Tribunal’s headnote in UKUT 81 says supervisors “must ensure that fee-earners under their supervision are aware of the dangers”, and that failure “is likely to result in a referral” to the SRA. Ayinde warned heads of chambers and managing partners that “the profession can expect the court to inquire whether those leadership responsibilities have been fulfilled”. Your notice to the court should therefore name the supervising lawyer’s role honestly. Blaming the junior in public, as in Crabill, is the single worst move available.

After the crisis: the post-mortem template

A policy did not save Butler Snow; Judge Manasco noted the firm had “repeated warnings, internal controls, and firm policies about the dangers of AI misuse… And yet here we are.” What the court credited was the response: an audit, retraining, and a verification requirement with teeth. Judge Wingate’s chambers rule after his own AI-tainted order is the simplest version: a second reviewer, and every cited case printed from the database and attached to the draft.

Write an internal incident record for our firm's file, in neutral language, from these facts: [matter; filing date; who drafted; which tool and tier; the prompt used, if known; who reviewed and what they reviewed for; who signed; how the error was discovered; what we filed and when; client and insurer notified on which dates].
Then draft a one-page corrective plan with: the verification rule going forward (every citation opened in a primary database, checked by, database, date, logged); who may use which tools for court filings; a supervision step for any document drafted by a junior, contractor or AI; and a quarterly review. Mark any professional-conduct rule you mention [VERIFY] for our general counsel to confirm.

Keep the record. It becomes the “pre-filing verification log” and “incident procedure” that now appear on published lists of what to have ready at renewal, and it is the evidence that the next filing was different. If your firm has no policy at all, the one-page AI policy template is the place to start, and the prompt library holds the verification prompts as reusable entries.

Where to go next: the six-layer verification protocol is what stops this happening twice; the sanctions timeline shows every outcome above in date order; and the rest of the verification cluster covers checkers, Europe and the other side’s filings. The verification session of AI Lab for Lawyers ends with exactly this checklist, worked through on participants’ own documents, so that nobody has to search for it in a panic.