What the courts found.
What you can change.

• Every state and D.C. has at least one decision. Rhode Island, South Dakota, and West Virginia have one each; Alaska and New Hampshire two.
• No decision involving a lawyer yet in Montana, North Dakota, South Dakota, or West Virginia.
• Federal courts publish more searchable orders, so the 68% federal share partly reflects what can be found. The data does not show that state-court practice is safer: state courts handed down three of 2026’s ten largest dollar sanctions against lawyers.
• Appellate courts are in the record too: the Second Circuit’s grievance-panel referral in Park v. Kim (January 2024), the Fourth Circuit’s public admonishment in In re Nwaubani (March 2026), the Nebraska Supreme Court’s dismissal of an appeal in Prososki v. Regan (March 2026), the Sixth Circuit’s $30,000 in fines in Whiting v. City of Athens (March 2026), the Ninth Circuit’s six-month suspensions in Lnu v. Blanche (June 2026), and the California Court of Appeal’s published $10,000 sanction in Noland v. Land of the Free (September 2025).

The pro se majority, briefly

• 817 of the 1,378 U.S. decisions (59%) involve a pro se party. The most common recorded outcome is a warning (413); 59 involve a dismissal of some kind; 41 record a monetary outcome, median about $600 where an amount is stated.
• This site focuses on lawyers because professional duties create a different accountability question. The pro se majority is an access-to-justice story that deserves its own study.

In the lawyers’ own words

“I didn’t know AI could do that” appears in 22 of 450 orders (5%). Counting each decision once, the ignorance family — didn’t know, thought it was a search engine, first time using AI — covers 35 (8%). Four of them:

Four takeaways

• Courts mostly do not record which product it was. Only 86 orders name the product (ChatGPT 35, legal research platforms with an AI layer 23, Gemini or Google 7, Claude 4, Copilot 2). That share, one in five, has been flat since mid-2025. The sanction turns on what the lawyer did after the output arrived, not on where it came from.
• Most of the lawyers caught were solo or in small firms.

  Firm size of the 395 private-practice lawyers placed in one band

  The orders themselves state firm size only 16% of the time; the rest comes from dockets, filings, and firm and bar records (method in About).

• The first order is not always the last. Thirty-seven of the 420 lawyers identified appear in more than one order, thirteen of them only because one case produced two. For the rest, the record shows what happened after a court had already spoken:

  First order regarding the lawyer	What followed
  Facey v. Fisher (N.Y. Sup. Ct., Sept. 15, 2025): sanctions under 22 NYCRR 130-1.1 against the plaintiff’s lawyer, with the defendants’ fees, in an order noting that federal judges had already warned him (“This conduct cannot go unaddressed and without consequences.”)	Five orders in six months. Two weeks later a District of New Jersey warning (Kertesz v. Colony Tire, Sept. 30, 2025: “It is sanctionable conduct”); in December, $6,000, a self-report to the New Jersey and New York licensing authorities, and copies to the client (Gardner v. Combs, Dec. 15, 2025), then the consolidated complaint dismissed without prejudice, with thirty days to amend, after an opposition brief that “continued to use fabricated citations” (Hildebrandt v. SiParadigm, Dec. 31, 2025); in March 2026 a public admonition from the Southern District of New York: “[T]he warnings and sanctions by the courts, as well as the CLE courses, did not have an appreciable impact” (Cartagena v. Dixon, Mar. 10, 2026)
  LiveVideo.AI Corp. v. Redstone (S.D.N.Y.): an earlier grievance-committee referral	June 2026: $85,056 in fees, and the lawyer referred to the district’s grievance committee “again”
  Withers v. City of Aberdeen (N.D. Miss.): show-cause hearing, Jan. 20, 2026	March 2026: the same lawyer filed two more documents citing nonexistent cases in a Louisiana bankruptcy court and was sanctioned there on Apr. 9 (In re Wise, Bankr. W.D. La.); the Mississippi sanctions order of June 8, 2026 notes it
  Coomer v. Lindell (D. Colo.): $3,000 sanctions, July 2025	March 2026: a second show-cause order proposing a “graduated amount of $5,000” “for their continued failure to check their citations”; imposed May 2026

• It is not a novice problem. The failures span the career ladder, from 0 to 56 years since the earliest verified bar admission found. Of the 420 lawyers identified, 394 have a bar-admission year we could verify. At the date of the first order against each of them, the median lawyer had been admitted for 21 years. More than half (54%) had 20 years or more.

  Years since first bar admission, at the first order

The mutation problem

A check that asks only whether the case exists catches one kind of error. The Princeton taxonomy of citation hallucinations, built from actual filings, has five categories, and an existence check catches only one of them: the case that does not exist. The other four survive it: a name and a reporter citation that “refer to two different real cases,” a page that “does not support the quoted language,” a quotation that “does not appear in the cited case,” and a real case that “does not support the proposition.” In 32 of the 450 orders on lawyers (7%), the lawyer told the court that some of the citations had been checked; the rest had not.

In Concord Music Group v. Anthropic (N.D. Cal. May 23, 2025), a lawyer had the right article and the right link and asked Claude to put the citation in proper form. Back came the right journal and year with “a fictitious article name with inaccurate authors (who have never worked together).” The AI’s process transformed the citation from correct to incorrect, and nobody noticed until it was filed. “That is a plain and simple AI hallucination,” the magistrate judge wrote, and yet “the underlying article exists, was properly linked to and was located by a human being using Google search.”

• Parker v. Costco Wholesale Corp. (W.D. Wash. Nov. 7, 2025): drafting with a legal-specific AI tool, “the cases offered by Counsel do exist and some generally stand for the propositions attributed to them through hallucinated quotations. But the quotations Counsel included do not actually exist.” $3,000 to the court, the defendant’s fees, a public reprimand, and a bar referral.
• TOV Realty, LLC v. Suarez (Conn. July 31, 2026): the lawyer wrote accurate documents and used ChatGPT to edit them; “unbeknownst to [the lawyer], the drafts produced by ChatGPT added new case citations or altered existing case citations.” Six hours of CLE and donations to the Connecticut Bar Institute.

Verify every citation after the final AI-assisted pass, including editing and formatting.

Legal research tools still require review

• Retrieval does not eliminate citation errors. The 2024 and 2025 Stanford studies used different tasks and definitions, so their percentages are not a controlled comparison of retrieval with general chatbots. The retrieval-based systems in the 2025 study still made citation errors; that study’s own summary is “no panacea.” Vendor headlines have promised more than the release text (sources below).
• A citator signal does not establish that a quotation or proposition is accurate. “Linked and Shepardized” means the case exists and carries a citator flag. Whether it says what the AI’s sentence claims, and whether it is good law for that proposition, require reading the case.

The largest amounts against lawyers

Every figure above was read out of the order itself. The database records Rivera at $35,603 (the larger of its two awards), LiveVideo at $80,056 (the order says $85,056), and Mattox as a single $28,495.

Money is the minority outcome

The ceiling so far

• Suspension from practice before a court (9 of 450 orders, 2%): six months (Ninth Circuit; Supreme Court of Georgia); one year from the bar of the Middle District of Florida (In re Neusom, 2024); two years from any case in the Northern District of Mississippi (Withers, 2026), and two years from cases assigned to one judge of that district (Billups, 2025).
• Bar suspension on the merits: one, in Colorado in 2023 (one year and one day, ninety days served). A California State Bar Court recommendation of February 2026 (a stayed one-year suspension) awaits the Supreme Court of California.
• Disbarment: none found in any U.S. jurisdiction as of September 2, 2026 (search scope in the sources). The Seventh Circuit told one lawyer in June 2026 that “the next time any incident occurs, he could be subject to disbarment before this court.”

Reported outcomes after referral to the bar

• Candor sets the price. The Ninth Circuit’s six-month suspensions were imposed less for the fabricated citations than for the lawyers’ “repeated failure of candor” about them; a prompt admission and apology, the court said, might have warranted less. The Eastern District of Oklahoma credited a lawyer’s “candor” and chose $3,000 over what it called “professional ruin.”
• Pro hac vice revocations have proved durable. In Mavy v. Commissioner (D. Ariz. Jan. 2026), a district judge vacated a magistrate judge’s Rule 11 finding and most of the sanctions, and left in place the two that require no bad faith: the brief stayed struck and the lawyer stayed out.
• Clients are being told. Forty-two orders required the lawyer to send the client a copy; seven ordered fees refunded or forfeited, including an appointed appellate lawyer in Illinois who disgorged the full fee the county had paid.
• Unverified but reported: a thirty-day suspension of a Minnesota lawyer (Minnesota Public Radio, Aug. 21, 2026) and a six-month suspension from practice in the Middle District of Pennsylvania (a Pennsylvania litigation blog, July 2026). We could not obtain either order.

Opposing counsel has a verification duty too

• Missed the defects. Noland v. Land of the Free (Cal. Ct. App. Sept. 12, 2025): $10,000 against the appellant’s lawyer for a brief in which 21 of 23 quotations were fabricated — and no fees to the respondents, who had not noticed and “appear to have become aware of the issue only when the court issued its order to show cause.”
• Warned, not sanctioned. Nuvola, LLC v. Wright (Minn. Dist. Ct. Nov. 20, 2025): $1,000 against the lawyer who filed the fake cases, and, while expressly declining to sanction his opponent, a reminder that the duty to check runs to the cases in a brief you are answering.
• Both sides sanctioned. Withers v. City of Aberdeen (N.D. Miss. June 8, 2026): four lawyers sanctioned, two per side, after both lead counsel filed AI-drafted documents with hallucinated cases and both local counsel signed without reading. “This case presents the Court with an unusual scenario—attorneys for both litigants engaged in similar sanctionable conduct.” The trial, set for March, was cancelled.

How the warning changed, 2025–26

Five sentences, verbatim and in date order, from a district court’s first warning of 2025 to the latest appellate one.

What the sharp language is for

• The harshest orders go to lawyers who were not candid: who denied using AI and then admitted it, who answered a show-cause order with a second filing full of invented quotations, or who blamed a paralegal, a relative, or a software update. Lack of candor was the aggravating factor courts noted most often (144 of 450), tied with the burden on the court.
• Seventeen of the 22 orders (77%) in which the lawyer admitted the problem at once were coded measured, and several say so. The Eastern District of Oklahoma credited a lawyer’s “candor” and chose a $3,000 fine over “professional ruin.” The Ninth Circuit said a prompt admission and apology might have warranted less than the six-month suspension it imposed.
• Later orders repeatedly cite a small group of earlier sanctions decisions when selecting a remedy: 243 of 450 (54%) cite at least one.

What each is cited for

In the citing courts’ own parentheticals where we have them.

In one court’s words

“We expect that when lawyers submit briefs, they give us their best view of what the law is, and how that law supports their clients. We read those briefs carefully—not just because they are the product of counsel’s time, effort, and skill, but because they help us as we try to reach the right answer. All that is lost if a lawyer decides that it is not worthwhile to do the work to persuade us—if he is willing to sign his name on whatever a machine churns out without so much as checking it for accuracy. We cannot do our job the same way unless lawyers do theirs. Whatever the merits of artificial intelligence, it is no substitute for actual intelligence.”Akerlund v. Atlas Air, Inc., No. 24-11033 (11th Cir. July 10, 2026) (Grant, J.), a published opinion, which announced a referral of the lawyer to the court’s Committee on Lawyer Qualifications and Conduct

Has any one judge seen enough of these to change tone?

Not yet. In the 450 orders, one decision-maker appears five times, the general counsel of the Government Accountability Office in bid protests, and three judges appear four times: district judges in the Western District of Texas, the District of Columbia, and the District of Kansas. At four or five orders there is no trend to report. What the repeats show is judges settling on a standard paragraph: the District of Columbia judge’s four orders, over ten months, each advise the lawyer in nearly the same words that “it is not acceptable to submit filings to the Court containing nonexistent cases and quotations.”

The rules and authorities that matter

The ABA Model Rules are models; the version adopted in the lawyer’s jurisdiction governs, and the numbering and wording vary.

Which authority applies: five questions

Using AI neither establishes nor excuses a violation; it bears on the remedy. A sanctions order, and the response to a show-cause order, turn on five questions that decide which authority applies and what the court has to find.

The remediation checklist courts keep reproducing comes from Wadsworth v. Walmart: withdraw promptly, be candid about the AI use, pay the other side’s fees for the tainted filing, adopt training and safeguards.

What the ABA and the state bars said

These tools run in your browser. Text stays in the current page and is not transmitted or saved by this site. Refreshing clears the workbench. Downloads are created only when you request them.

The check, in three steps

1. Confirm the authority exists. Find it on CourtListener, Google Scholar, the court’s own website, or a subscription service, under the citation given.
2. Compare the quotations and pincites. Every quoted phrase must appear on the page cited; every pincite must land on the words it stands for.
3. Read for the proposition and the treatment. The source must support the point in context, and you must know how it has been treated since. No checker replaces this step.

By role

Open yours.

Citation checkers have proliferated

• Most check existence; few check the quotation or the proposition. Seventeen products are listed below, alphabetically, with what each says it checks. Step 1 is well covered; steps 2 and 3 — the failures that have drawn level with the invented case in lawyer filings — are not.
• Nothing here was tested. Entries reflect vendor descriptions and the identified reporting as reviewed through September 6, 2026; each source note gives the relevant review date. Vendor accuracy claims are not repeated. Before uploading a filing to any of them, read the retention and training terms and consider client confidentiality and privilege.
• Foundation models are not a complete answer either. In the one independent benchmark (Princeton, June 2026), the best-recall system caught about 84% of planted citation errors but raised many false alarms (F1 55%); the best-balanced system caught 63% (F1 69%); every model did worst on wrong pincites. A checker is a screen for step 1, not a substitute for steps 2 and 3.

Feature and pricing descriptions are the inherited vendor-source snapshot through September 6, 2026. These products were not independently tested for this study. A filtered match can include partial, conditional, or vendor-claimed support; read the cell and the source before choosing.

“Yes” is the vendor’s description; “as reported” means independent press coverage. Where reporting and the product page differ (Shepard’s Verify: “cited authority exists and is retrievable” in reporting, versus “flagging unsupported statements” on the product page), the table follows the reporting. No tool in the table can take on the filer’s responsibility to verify.

The database

• What it is. Damien Charlotin’s AI Hallucination Cases is a public list of court and tribunal decisions “where the use of AI, whether established or merely alleged, is addressed in more than a passing reference by the court or tribunal,” limited to cases where the court “explicitly found (or implied)” that a party relied on hallucinated content. Its earliest entry is a Brazilian electoral-court decision of April 2023; the first American entries follow that June. It now covers 40 countries.
• Who decides. Charlotin does not decide whether a filing was hallucinated; the courts do. In his words: “I am not making that judgment, I let the courts and judges make or imply it, which is why the database is necessarily an undercount.” The briefs quietly withdrawn, the fabrications nobody checked, the footnote mentions: none of those are here.
• What each row holds. Case name, court, state, date, the party responsible (pro se, lawyer, judge, expert), the product where named, an itemized list of what was false (fabricated, misrepresented, false quotes, outdated), the outcome, the monetary penalty, whether a professional sanction followed, and a link to the decision where one is available.
• Flags. Nineteen rows (1% of 2,008) are marked as alleged rather than established; twelve (0.6%) carry a vendor-dispute flag; no row carries both. All 31 are counted in the totals, as the database counts them.
• Disclosure. Charlotin also markets a citation-checking tool, PelAIkan. It appears in the checker roundup on the same terms as everything else. He has not reviewed this site and is not responsible for anything in it.

Method

• The reading. Each of the 477 orders was read under a written codebook asking the same fixed questions (who used the AI, how, what explanation was offered, who caught it, what the court did, what authorities it relied on, what other decisions it cited, and what its register was), by one model reader, one pass per order, with the key answers required to quote the passage that supported them. No order was coded twice; register, motive, and explanation fields are one reader’s judgment.
• The checking. 1,565 of the 1,586 quotations the reader produced matched the order text after both were reduced to letters only; the 21 that did not are not used anywhere. Every case name, citation, date, dollar figure, and quotation in the prose was then checked again against the order itself, the docket on CourtListener’s RECAP archive, a court’s own website, or the issuing body’s publication, and the built site was audited and proofread by a second model. Where only press coverage could be found, the text says so. Where nothing could be found, the claim was dropped.

Scope and limitations

• An undercount, and a lag. The database is one person’s reading of the world’s courts, in the languages he reads, from the sources that reach him; in his own words, an undercount. Counts by decision date understate the most recent weeks, which are still being added. The projections on the milestone page are arithmetic on trailing rates, not forecasts of behavior.
• One reader, 450 orders. The coded set covers 449 of 546 U.S. rows on lawyers (82%). The register of an order is a reader’s judgment: the coding was done by a model and checked by quotation, not by a second human reader, and the categories are the ones the codebook offered.
• Left out on purpose. Names of the lawyers sanctioned, except where a case caption is itself a lawyer’s name and the matter is bar discipline of public record (captions are cited as they appear on the public docket, and a caption that names a firm or a party is not redacted); names of small firms and solo practices; any claim about a lawyer’s intent beyond what a court found; any figure from a vendor about its own product’s accuracy; any citation this site could not open.
• Firm size and experience are distributions, not rates. The firm-size and years-since-admission figures on The circumstances describe the lawyers identified in the collected cases. The database has no denominator for briefs filed, lawyers practicing, or AI-assisted work produced, so those figures cannot say which lawyers are more likely to file false authority. The database also tags three rows as lawyer cases that the dockets show were pro se filings; the study charts and coded analysis count them as pro se.
• Searches, not proofs. Negative findings (“no disbarment,” “no AI-specific CLE requirement”) cover public news, bar, and court sources through September 2, 2026, and absence from those sources is not proof of absence. Two claims in circulation were searched for and not found: that a lawyer has been disbarred for this, and that fee awards in early 2026 reached a particular six-figure sum. Two reported suspensions we could not verify (Minnesota; the Middle District of Pennsylvania) are mentioned on the cost page with their source and nothing more.

Credits

Embedded typefaces in Two Thousand Decisions v27

Newsreader
Copyright 2020 The Newsreader Project Authors
http://github.com/productiontype/Newsreader

Source Serif 4
Copyright 2014 - 2021 Adobe Systems Incorporated
http://www.adobe.com/
Reserved Font Name: Source
Project: https://github.com/adobe-fonts/source-serif

IBM Plex Mono
Copyright 2017 IBM Corp. All rights reserved.
Reserved Font Name: Plex

All three typefaces are distributed under the SIL Open Font License 1.1.
The embedded WOFF containers preserve the original font tables and metadata.
No glyphs, font names, or original table contents were changed.

License sources:
https://raw.githubusercontent.com/google/fonts/main/ofl/newsreader/OFL.txt
https://raw.githubusercontent.com/google/fonts/main/ofl/sourceserif4/OFL.txt
https://raw.githubusercontent.com/google/fonts/main/ofl/ibmplexmono/OFL.txt
https://openfontlicense.org/ofl-faq/#22-can-i-make-and-use-woff-web-open-font-format-versions-of-ofl-fonts

-----------------------------------------------------------
SIL OPEN FONT LICENSE Version 1.1 - 26 February 2007
-----------------------------------------------------------

PREAMBLE
The goals of the Open Font License (OFL) are to stimulate worldwide
development of collaborative font projects, to support the font creation
efforts of academic and linguistic communities, and to provide a free and
open framework in which fonts may be shared and improved in partnership
with others.

The OFL allows the licensed fonts to be used, studied, modified and
redistributed freely as long as they are not sold by themselves. The
fonts, including any derivative works, can be bundled, embedded,
redistributed and/or sold with any software provided that any reserved
names are not used by derivative works. The fonts and derivatives,
however, cannot be released under any other type of license. The
requirement for fonts to remain under this license does not apply
to any document created using the fonts or their derivatives.

DEFINITIONS
"Font Software" refers to the set of files released by the Copyright
Holder(s) under this license and clearly marked as such. This may
include source files, build scripts and documentation.

"Reserved Font Name" refers to any names specified as such after the
copyright statement(s).

"Original Version" refers to the collection of Font Software components as
distributed by the Copyright Holder(s).

"Modified Version" refers to any derivative made by adding to, deleting,
or substituting -- in part or in whole -- any of the components of the
Original Version, by changing formats or by porting the Font Software to a
new environment.

"Author" refers to any designer, engineer, programmer, technical
writer or other person who contributed to the Font Software.

PERMISSION & CONDITIONS
Permission is hereby granted, free of charge, to any person obtaining
a copy of the Font Software, to use, study, copy, merge, embed, modify,
redistribute, and sell modified and unmodified copies of the Font
Software, subject to the following conditions:

1) Neither the Font Software nor any of its individual components,
in Original or Modified Versions, may be sold by itself.

2) Original or Modified Versions of the Font Software may be bundled,
redistributed and/or sold with any software, provided that each copy
contains the above copyright notice and this license. These can be
included either as stand-alone text files, human-readable headers or
in the appropriate machine-readable metadata fields within text or
binary files as long as those fields can be easily viewed by the user.

3) No Modified Version of the Font Software may use the Reserved Font
Name(s) unless explicit written permission is granted by the corresponding
Copyright Holder. This restriction only applies to the primary font name as
presented to the users.

4) The name(s) of the Copyright Holder(s) or the Author(s) of the Font
Software shall not be used to promote, endorse or advertise any
Modified Version, except to acknowledge the contribution(s) of the
Copyright Holder(s) and the Author(s) or with their explicit written
permission.

5) The Font Software, modified or unmodified, in part or in whole,
must be distributed entirely under this license, and must not be
distributed under any other license. The requirement for fonts to
remain under this license does not apply to any document created
using the Font Software.

TERMINATION
This license becomes null and void if any of the above conditions are
not met.

DISCLAIMER
THE FONT SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTIES OF
MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT
OF COPYRIGHT, PATENT, TRADEMARK, OR OTHER RIGHT. IN NO EVENT SHALL THE
COPYRIGHT HOLDER BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
INCLUDING ANY GENERAL, SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL
DAMAGES, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
FROM, OUT OF THE USE OR INABILITY TO USE THE FONT SOFTWARE OR FROM
OTHER DEALINGS IN THE FONT SOFTWARE.

This project is built on Damien Charlotin’s freely available AI Hallucination Cases database and the years of work behind it. The 2,000-decision milestone belongs to that record.

• Thanks to Mary E. Matuszak, Director of Library Services at the New York County District Attorney’s Office, whose posts on these decisions as they arrive keep the profession reading them.
• Thanks to Mark Pike, whose Oyez AI? game in the amicus project, which shows one Supreme Court case caption a day and asks whether it is real or generated, inspired the training exercises on this site.
• Original edition (September 5, 2026): built by Kyle Bahr, Claude Fable 5.1, and OpenAI Codex 5.6 Sol. Kyle set the scope and the editorial rules; Claude researched, coded, drafted, and checked; Codex independently audited the built site, proofread it, and researched firm size and years since bar admission beyond the orders; Kyle reviewed the published versions.
• Expanded edition (September 7–9, 2026): built by Kyle Bahr with OpenAI Codex GPT-6 Astra and Claude Fable 5.1. Codex rebuilt the site, ran the source-pronoun review, and designed the Before ChatGPT page; Claude rewrote the exercises for clarity, ran the pre-ChatGPT research with an independent verification pass, and built the contract review page; each reviewed the other’s work; Kyle reviewed every version and decided what stayed.
• Three additions began as LinkedIn comments on the original: thanks to Carolyn Elefant (gender), Doug Hall (Before ChatGPT), and Rafie Faruq (contract review).
• Original edition published September 5, 2026, with corrections through September 6. Expanded edition built September 7–9, 2026.