When the Brief Cites a Case That Doesn't Exist

The question

A lawyer files a brief. It quotes a precedent that reads perfectly: the right court, a plausible reporter citation, a clean and quotable holding. There is only one problem. The case does not exist.

Since 2023, that scenario has moved from curiosity to recurring docket event. A generative-AI tool, asked for supporting authority, returns text that is fluent, confident, and — sometimes — invented: a non-existent opinion, or a real case dressed in a quotation it never contained. The lawyer, not reading behind the output, files it.

The rules that govern this did not change when the technology arrived. A lawyer's certification obligations under Federal Rule of Civil Procedure 11 and the duty of candor toward the tribunal predate generative AI by decades. What changed is how easy it has become to file something unverified that looks verified. This piece asks a descriptive question: under those existing duties, how are U.S. courts calibrating consequences when AI-assisted filings contain fabricated or erroneous citations — and what does the 2025–2026 wave of sanctions reveal about the emerging, and decidedly non-uniform, pattern?

The short answer, stated here at the top and supported below: the obligation is non-delegable and runs to the human who signs the filing; the technology changed the workflow, not the duty; and verification — opening the cited authority and confirming it says what the brief claims — remains the constant the rules already required.

The settled duty (the anchor)

Before the wave, there was the rule. Rule 11(b) makes the act of presenting a filing a certification. By signing, filing, submitting, or later advocating a paper, an attorney certifies — after a reasonable inquiry — that its legal and factual contentions have support¹:

The operative standard is the rule's own phrase, "an inquiry reasonable under the circumstances" — what courts commonly describe as an objective-reasonableness test, a label that is the courts' shorthand rather than the rule's words. Subjective good faith is not the measure; a sincere but unverified belief that a cited case exists does not satisfy the certification. When the standard is breached, Rule 11(c) authorizes a sanction, after notice and an opportunity to respond, that is "limited to what suffices to deter repetition."¹

The professional-conduct rules run alongside Rule 11. Most directly on point is the duty of candor toward the tribunal under Model Rule 3.3²:

Two distinct prongs matter here, and they should not be blurred. A fabricated citation is, most naturally, a false statement of law to the tribunal under Rule 3.3(a)(1) — and that subsection reaches past the bare prohibition: by its terms it also forbids a lawyer to "fail to correct a false statement of material fact or law previously made to the tribunal by the lawyer." So the duty to fix a fabricated citation, once the lawyer learns of it, is built directly into the candor rule. Rule 3.3(a)(3) addresses a separate track — offering evidence the lawyer knows to be false — and the rule separately obliges a lawyer who comes to learn that offered evidence is false to "take reasonable remedial measures, including, if necessary, disclosure to the tribunal." The false-statement track and the evidence track are related but not identical, and they should be kept distinct.² (These are model templates; each state's adopted version controls in that jurisdiction.)

Competence supplies a third hook. Model Rule 1.1's Comment [8], as amended in 2012, ties competence to technology³:

The supervision rules — Model Rule 5.1 (responsibilities of partners, managers, and supervisory lawyers) and Model Rule 5.3 (responsibilities regarding nonlawyer assistance) — supply the duty to supervise delegated work. Their black-letter text does not mention artificial intelligence. The reading that they reach a lawyer's use of an AI tool is drawn not from the rules' own words but from the American Bar Association's mapping of existing duties onto generative AI in Formal Opinion 512 (July 29, 2024)⁴:

Opinion 512 creates no new rules. It is a non-binding advisory opinion that maps the existing duties — competence, confidentiality, communication, supervision, meritorious claims, candor, and fees — onto generative-AI use. It is on point precisely because it treats AI as a tool whose use is governed by duties that already existed.⁴

The origin point: Mata v. Avianca (2023)

The case that set the template arose in the Southern District of New York. In Mata v. Avianca, Inc., attorneys Steven A. Schwartz and Peter LoDuca, with the firm Levidow, Levidow & Oberman, P.C., submitted an affirmation that cited and quoted non-existent judicial opinions — fabricated decisions attributed to judges who had written no such thing — generated by ChatGPT, and then stood by them after their authenticity was questioned.⁵ Judge P. Kevin Castel grounded the sanction on Rule 11(b)(2) and (c), "or, alternatively, [the court's] inherent authority." The two standards can seem to collide — Rule 11(b)'s certification is judged objectively, yet the court applied a heightened subjective-bad-faith standard — but they answer different questions: the objective standard is what the rule asks of the filer, while the bad-faith finding followed from the court's raising sanctions on its own initiative and invoking its inherent authority, and attached to the respondents' standing by the fabrications once their authenticity was questioned. The court imposed a monetary penalty:

The $5,000 was joint and several across all three respondents — both attorneys and the firm — not a penalty on one lawyer. A separate and distinctive remedy required the respondents to send letters to the plaintiff and to each judge who had been falsely identified as the author of a fabricated opinion. Mata established the proposition the later cases all assume: the output of the tool is the lawyer's responsibility, and signing it is a certification about it.⁵

The 2025–2026 sanctions wave (the live frontier)

What was an isolated cautionary tale in 2023 is now a documented, recurring phenomenon. The independent "AI Hallucination Cases" database maintained by Damien Charlotin catalogs legal decisions worldwide in which a court found that a party relied on generative-AI hallucinated content — typically fabricated citations. As of 2026-06-15, and updated daily, the database reports¹¹:

That figure is a worldwide count from a third-party dataset, cited here for scale rather than as a precise national tally; it should be read with its as-of date, because it moves. What the volume establishes is that the problem is not a handful of outliers.

The duty is model-agnostic

One feature of the wave is worth stating plainly, because it is easy to get wrong: the duty does not change with the brand of the tool. In Mata, the fabricated opinions came from ChatGPT. In Kaur v. Desso (N.D.N.Y. 2025), a federal habeas matter, counsel's supplemental brief contained fabricated quotations attributed to real, correctly named Supreme Court cases; counsel's sworn response to the court's order to show cause stated that he had used "the generative AI tool Claude Sonnet 4" to assist with research and drafting.⁹ The court imposed a monetary penalty:

Two points of precision. First, the model attribution rests on counsel's own sworn representation as recited by the court; the court did not independently audit which model produced the text — exactly as the ChatGPT attribution in Mata reflects what the record showed, not a forensic finding. Second, the court treated fabricated quotations from real cases no differently from invented cases: it saw "no reason to distinguish between the submission of fabricated cases and the submission of fabricated quotations from real cases."⁹ The sanction — the monetary penalty plus proof of a continuing-legal-education program on AI — flowed from the lawyer's conduct, not from the identity of the software. The analysis in each case is of the human filer's duty; the tool's name is a fact in the record, not the thing being judged. (This article is itself produced with AI assistance under a verification discipline, described below — which is one more reason the model-neutral frame is the honest one.)

Candor on discovery: the Prince Global self-report

Not every episode is a contested sanction. In the Chapter 15 case In re Prince Global Holdings Limited (Bankr. S.D.N.Y. 2026), the law firm Sullivan & Cromwell self-reported to Chief U.S. Bankruptcy Judge Martin Glenn that its April 9, 2026 emergency motion contained inaccurate citations and other errors — dozens of them, itemized by paragraph in a "Schedule A," a subset of which the firm attributed to AI hallucinations and others to manual or clerical error⁸:

The firm filed a corrected motion and a redline the same day, and the court granted the corrected provisional-relief motion days later. The episode is included not for any monetary penalty — the salient facts are the prompt self-report and correction, and the firm's own statement that its AI-use policies were not followed. How a filer responds once an error is discovered is, as the calibration section below shows, one of the things courts weigh.⁸

The same standard, a wide range of severity

The clearest lesson of the 2025–2026 appellate decisions is that courts agree on the rule and diverge on the consequence. At the severe end, the Sixth Circuit, in Whiting v. City of Athens, Tennessee (2026), confronted a record it described in stark terms⁶:

The court drew its remedies from two sources. It ordered the lawyers to reimburse the appellees' fees and to pay double costs — double costs being the cost penalty Rule 38 makes available for a frivolous appeal — while the $15,000-each sanction it called punitive rested on the court's inherent authority, which requires a finding of bad faith; it also referred the lawyers to the chief judge for possible discipline. Note the court's own framing — "fake citations and misrepresentations of fact," a bundle of problems across three appeals, not a single slip.⁶

At the lenient end, the Seventh Circuit, in Dec v. Mullin (2026), faced a brief that "contained non-existent citations and a false quotation, seemingly generated by artificial intelligence" — and imposed no monetary sanction at all⁷. The court admonished counsel and declined a show-cause order, citing counsel's acceptance of responsibility, a sincere apology, the small number of errors, and the fact that they mainly supported an undisputed standard of review. Same governing duty as Whiting; a very different result on very different facts.⁷

Prospective controls: individual-judge standing orders

Some judges have responded by issuing standing orders that regulate AI use before a problem arises. Two early examples are instructive precisely because their rationales differ. Judge Brantley Starr of the Northern District of Texas issued a mandatory certification order in 2023 — widely cited as the first of its kind — requiring each attorney to certify either that no portion of a filing was drafted by generative AI or that any AI-drafted language was checked for accuracy by a human; it warns that AI can "make stuff up — even quotes and citations."¹⁰ Judge Stephen A. Vaden of the U.S. Court of International Trade — a federal trial court, not a district court — issued an order in 2023 aimed at a different risk¹⁰:

Starr's order targets the hallucination-and-accuracy risk; Vaden's targets the confidentiality risk. Both are individual-judge standing orders, not court-wide or district-wide rules, and the two rationales should not be conflated or generalized into a claim that "the district courts are adopting" any single approach.¹⁰

How courts are calibrating

Reading the surveyed decisions together, the cases share a standard but reach different consequences, and the divergence is not random. Several factors recur as the variables that move a result from an admonishment toward a five-figure punitive sanction. They are best understood as competing, fact-dependent considerations rather than a settled formula.

Factor	Toward leniency	Toward severity
Volume & pattern	A few isolated errors (Dec)	Over two dozen fabrications and misrepresentations across multiple filings (Whiting)
Intent / knowledge	Carelessness, promptly owned	Subjective bad faith; standing by the fabrication when questioned (Mata)
Candor on discovery	Prompt self-report and correction (Prince Global); acceptance of responsibility and apology (Dec)	Defiance, or persistence after authenticity is challenged
Prejudice / materiality	Errors in a non-dispositive section, supporting an undisputed point (Dec)	Fabrications going to load-bearing propositions
Repetition	A single lapse	Conduct repeated across briefs or appeals (Whiting)

The same factors explain why the basis for a sanction can shift — Rule 11 and inherent authority in the trial courts, Appellate Rule 38 and inherent authority on appeal, the professional-conduct rules in parallel — while the underlying expectation holds steady. What a court is measuring is whether the lawyer made the inquiry the rules require and how the lawyer behaved once the problem came to light.

One question the surveyed decisions do not resolve is worth naming, because it is where the live debate sits. They agree that verification is required and that a fabricated citation is sanctionable; what they have not settled is how much verification "an inquiry reasonable under the circumstances" demands when the draft in front of the lawyer was produced by an AI tool — whether checking the output against a conventional database is enough, and whether a tool's own assurances of accuracy can ever be part of a reasonable inquiry. The decisions surveyed here police the failure to verify at all; the contours of sufficient inquiry for AI-assisted work remain unsettled.

The through-line

Across every decision surveyed here, one observation holds without exception. The duty is non-delegable and runs to the human who signs the filing. The technology changed the workflow — it did not change the obligation. A generative-AI tool can draft, summarize, and suggest; it cannot certify, and it bears no duty of candor. That duty belongs to the lawyer, exactly as it did before the tool existed.

Stated as a neutral observation rather than advice: verification is the constant. The act the rules have always required — open the cited authority and confirm it says what the brief claims it says — is the same act that would have prevented every fabricated-citation sanction in this survey. The cases are new; the duty they enforce is old.

Scope and limits

How this was verified

A piece about fabricated citations has to hold itself to the standard it describes. Every named case, rule, quotation, and figure above was checked against its primary source — the rule and Model-Rule text, the ethics opinion, and the court orders and opinions themselves — before it was allowed into the article. Anything that could not be confirmed was cut, not softened. Two moments from that process are worth surfacing, because they are the demonstration in miniature.

A fabricated citation, caught inside the research. During verification, a working source list for this article contained a Sixth Circuit decision under a case name and reporter citation that — on checking against the court's records — corresponded to no real decision. It was a fabricated citation that had surfaced inside the research for an article about fabricated citations — not staged for effect, but the ordinary kind of error the checking exists to catch. It was removed. The actual Sixth Circuit authority, Whiting v. City of Athens, was then confirmed from the court's own published opinion and used in its place. The point of the piece, demonstrated against the piece itself: a confident-looking citation is not a verified one until someone opens the source.

Produced under Echo Angel Studio's governed-research discipline. Sources verified as of 2026-06-15. This is neutral research, independently verifiable against the primary documents linked below — not legal advice.

Authorities & sources

Each load-bearing claim above is keyed by superscript to the primary authority it rests on. Links go to the primary document.

1. Federal Rule of Civil Procedure 11(b)–(c). The certification standard ("an inquiry reasonable under the circumstances") and the sanctions provision (a sanction "limited to what suffices to deter repetition"). law.cornell.edu/rules/frcp/rule_11
2. ABA Model Rule of Professional Conduct 3.3 — Candor Toward the Tribunal. Rule 3.3(a): no knowing false statement of fact or law to a tribunal; the remedial-measures duty on offered false evidence. Model template; each state's adopted version controls. americanbar.org — Model Rule 3.3
3. ABA Model Rule 1.1, Comment [8] (with Rules 5.1 and 5.3). Competence includes "the benefits and risks associated with relevant technology" (clause added 2012); Rules 5.1 (partners, managers, and supervisory lawyers) and 5.3 (nonlawyer assistance) supply the supervision duties. americanbar.org — Model Rule 1.1
4. ABA Formal Opinion 512 — Generative Artificial Intelligence Tools (July 29, 2024). The ABA's first formal ethics opinion specifically on generative AI; maps existing Model Rules onto GAI use; creates no new rules; non-binding advisory. americanbar.org — Formal Op. 512 (PDF)
5. Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023) (Castel, J.). The foundational ChatGPT fabricated-citation case; $5,000 jointly and severally on two attorneys and the firm under Rule 11(b)(2) and (c) or inherent authority (subjective bad faith); letters to the falsely identified judges. law.berkeley.edu — Mata v. Avianca (PDF)
6. Whiting v. City of Athens, Tennessee, Nos. 24-5918/5919, 25-5424 (6th Cir. 2026). The severe end: "over two dozen fake citations and misrepresentations of fact"; $15,000 each on two attorneys under Fed. R. App. P. 38 and inherent authority, plus fees, double costs, and a disciplinary referral. opn.ca6.uscourts.gov — Whiting (PDF)
7. Dec v. Mullin, No. 25-2417 (7th Cir. 2026) (Brennan, J.). The lenient end of the same standard: "non-existent citations and a false quotation, seemingly generated by artificial intelligence"; admonishment, no monetary sanction. media.ca7.uscourts.gov — Dec v. Mullin
8. In re Prince Global Holdings Limited, No. 26-10769 (MG) (Bankr. S.D.N.Y. 2026). Sullivan & Cromwell's self-report to Chief Judge Glenn of inaccurate citations and errors (dozens itemized in a "Schedule A," a subset attributed to AI; others manual/clerical); "The Firm's policies on the use of AI were not followed"; corrected motion granted days later. firm disclosure letter (PDF) · court opinion (PDF)
9. Kaur v. Desso, No. 9:25-cv-726 (AMN) (N.D.N.Y. July 9, 2025) (Nardacci, J.). Federal habeas matter; fabricated quotations attributed to real Supreme Court cases; counsel's sworn response stated he used Claude Sonnet 4 (a representation the court did not independently audit, exactly as the ChatGPT attribution in Mata reflected the record rather than a forensic finding); $1,000 plus a CLE program; the court drew "no reason to distinguish" fabricated cases from fabricated quotations of real cases. govinfo.gov — Kaur v. Desso (PDF)
10. Standing orders: Judge Brantley Starr (N.D. Tex., 2023) and Judge Stephen A. Vaden (U.S. Court of International Trade, 2023). Individual-judge standing orders with distinct rationales: Starr's mandatory AI certification targets the hallucination/accuracy risk; Vaden's disclosure-and-certification order targets the confidentiality risk. The Court of International Trade is a federal trial court, not a district court. txnd.uscourts.gov — Starr certification · cit.uscourts.gov — Vaden order (PDF)
11. Damien Charlotin, "AI Hallucination Cases" database. A curated, daily-updated database of legal decisions worldwide in which a court found reliance on generative-AI hallucinated content. Cited for worldwide scale ("1621 cases identified so far") with its as-of date; no U.S. sub-count is asserted (the per-jurisdiction breakdown is not primary-verifiable here). damiencharlotin.com/hallucinations