reportAI in CourtsAI Hallucinations
566 US court decisions on AI: fabricated citations and what else courts recorded
As of 8 Sep 2026, 566 US court documents on generative AI and related AI questions show 507 fabricated-citation rulings.
Edited and verified by SafeLegalAI · updated
Researched with AI assistance · sources verified by SafeLegalAI · How this was made ↓
In the United States, the SafeLegalAI court-decision corpus holds 566 court-authored documents on generative AI and related AI questions as of 8 September 2026. The United States record is concentrated in 507 fabricated-citation decisions dated from February to September 2026, and the remaining 59 rows show what the faster sanctions record misses: courts disclosing their own AI use, testing AI evidence, deciding privilege, writing rules in opinions, citing substantive AI law and tracking legal-AI dockets.
This report asks what the 566-document corpus can show before every row has been promoted into the verified incident file: the monthly shape of US fabricated-citation decisions, plus court passages on how judges are using, limiting and classifying AI in proceedings.
Key findings
- As of 8 September 2026, the corpus holds 507 fabricated-citation decisions from February to September 2026: 56 in February, 112 in March, 91 in April, 101 in May, 44 in June, 35 in July, 66 in August and 2 in September.
- The actor field in each fabricated-citation row shows 310 of the 507 decisions involved a self-represented litigant, 192 involved a lawyer, and the other five involved a firm, an expert or another actor.
- The outcome categories sum to 507: 208 warnings, 74 with no consequence, 65 pending, 31 fines, 28 referrals, 26 filings struck, 23 costs orders, 21 dismissals, 20 sanctions, 7 other outcomes and 4 suspensions. No row is coded strike-off.
- Federal courts account for 334 of the 507 fabricated-citation rows, state courts for 172 and other bodies for 1; federal district courts alone account for 293.
- The court’s document names a tool in 38 of the 507 fabricated-citation rows and names none in 469. Generic AI labels are not counted as tools in v0.2.0; a separate
ai_mentionfield shows that in 111 fabricated-citation rows the court never itself names AI and says only that authorities do not exist or are misquoted. - The 59 non-fabrication decisions are led by 18 rows on self-represented litigants using AI other than fabricated authorities, 10 on court rules adopted by opinion or order, 7 on AI-generated evidence and authentication, 6 on criminal-justice algorithms, 5 on substantive AI law, 4 on courts’ own AI use, 4 on discovery, 3 on fees and 2 on privilege.
- Five decisions disclose a court’s own AI use: Snell v. United Specialty Insurance Co., United States v. Deleon, Ross v. United States, Smith v. P.A.M. Transport, Inc. and Gouveia v. Meridian Financial Investments, LLC.
- The litigation sidecar holds 9 legal-AI dockets as of 8 September 2026; Thomson Reuters v. ROSS is the only docket coded as on appeal, with Third Circuit No. 25-2153 argued on 11 June 2026 and no opinion found by the dataset build.
Why this question
The verified hallucinations file answers a different question from this corpus. It records incidents after editorial verification. This corpus records court-authored documents at scale, with the court’s discussion of AI captured directly in the row, and lets a reader see the monthly shape of fabricated-citation decisions while many rows are still awaiting editor review.
That distinction matters because the 2026 US record moved faster than the verified file could absorb it. Legal AI incident report, Q3 2026 covers the verified incident file through early September, and Three years of legal AI hallucinations gives the earlier baseline. This report instead uses the larger court-decision corpus to count where the decisions arrived, who filed the material, what courts did, which tools were named and which courts recur.
The non-fabrication rows answer the second half. They show courts speaking about their own use of language models, authentication of altered videos, the privilege status of Claude and ChatGPT conversations, protective orders for confidential information submitted to AI systems, state high-court rules adopted by order, and AI copyright decisions that legal-practice courts now cite. A separate report addresses judicial rules and adjudicator hallucinations in The bench’s record; another addresses the privilege and discovery side in AI chat logs, note-takers and privilege. The purpose here is narrower: what these 566 rows show when read as one US courts dataset.
Method and data
The decision dataset is exposed through /courts, decision pages, topic pages such as /courts/topics/fabricated-citations, monthly digests such as /courts/digest/2026-08, and the JSON export /courts/opinions.json. Each row records the court, date, topic labels, disposition, quoted court passage, tool and source link; fabricated-citation rows add actor, outcome, monetary penalty and tool details.
The corpus title is broader than generative AI alone. The criminal-justice algorithm rows and some substantive AI law rows concern AI systems that are not generative; the corpus includes them because they are coded as related AI court questions.
The litigation dataset is exposed through the court-litigation pages. It contains 9 dockets: unauthorized practice, legal database, copyright and AI training, regulator and trade secret matters. It is broader than the US decision corpus because legal-AI litigation crossed national lines before the court-decision harvest was built.
Every table below was computed by a Node script over the two JSON exports. The run log contains the scripts and output. Of the 566 decision rows, 300 were read from the court’s own copy, public RECAP archive, govinfo or another primary repository; 266 were read from a public copy of the court PDF.
Leads from Damien Charlotin’s AI Hallucination Cases Database (CC BY 4.0), damiencharlotin.com; each decision re-read and coded by SafeLegalAI
Other leads came from court and docket searches recorded in each row.
| Verification status | Decisions |
|---|---|
| Court document read from an official, RECAP or other primary repository | 300 |
| Public copy of court PDF read | 266 |
A v0.2.0 re-release followed an independent audit in which a 50-row random sample found 22 rows, or 44%, with at least one error in a coding column; every quoted passage and every case, court and date was right. Every row was then re-read from the court document against a written rubric: 555 of 569 rows were re-coded, 14 could not be re-fetched that day and keep first-pass coding as “Second read pending,” and one row was withdrawn because the document was counsel’s letter, not a court document. The new ai_mention column records whether the court itself names AI, chatbot use or hallucination; the outcome ladder is strike-off > suspension > referral > fine > costs-order > sanctions > dismissal > filing-struck > warning > pending > none > other; and pro-se-ai-use is confined to AI use other than fabricated authorities.
Across all rows, ai_mention is explicit in 435 decisions, none in 118 and unset in 13 rows where the second read is pending.
The fabricated-citation rows are corpus records, not automatically verified incidents. The Incident Tracker is the verified file; 505 of this corpus’s 507 fabricated-citation rows have a tracker slug, but this report counts rows from the court-decision corpus and does not describe them as verified incidents unless the tracker does.
The counts describe the record collected, not the universe of US court uses of AI, and the monthly counts describe what has been collected, not incidence. Unpublished orders, sealed proceedings and dockets without accessible PDFs are outside the count. Tool attribution is taken only when the court names a product as used; a generic AI label is not a tool. Court names were normalised only for abbreviated and spelled-out United States variants. Monetary penalties are counted only where the row contains an amount.
Fabricated-citation decisions arrived in waves, not evenly
The corpus’s fabricated-citation rows all fall between February and September 2026, and the monthly count is uneven. March and May together account for 213 of 507 rows; July has 35. September has 2 because the dataset was built on 8 September, not because the month was complete. Earlier rows in the all-decisions column are non-fabrication rows.
| Month | Fabricated-citation decisions | All decisions |
|---|---|---|
| 2024-02 | 0 | 1 |
| 2024-05 | 0 | 1 |
| 2024-09 | 0 | 1 |
| 2024-10 | 0 | 2 |
| 2025-01 | 0 | 1 |
| 2025-02 | 0 | 2 |
| 2025-03 | 0 | 1 |
| 2025-06 | 0 | 4 |
| 2025-09 | 0 | 1 |
| 2025-10 | 0 | 2 |
| 2025-12 | 0 | 2 |
| 2026-02 | 56 | 60 |
| 2026-03 | 112 | 118 |
| 2026-04 | 91 | 96 |
| 2026-05 | 101 | 103 |
| 2026-06 | 44 | 49 |
| 2026-07 | 35 | 44 |
| 2026-08 | 66 | 76 |
| 2026-09 | 2 | 2 |
The count is a collection rate. It should not be read as a misconduct rate. A month with more rows can mean more opinions, more docket access, more Charlotin leads or more backfill. The fabricated-citation record was already visible in February, peaked in March, stayed near 100 rows in April and May, then split between a slower June-July period and a renewed August harvest.
The non-fabrication rows cluster differently. February 2026 includes Matter of M.S., United States v. Heppner, Warner v. Gilbarco and Oklahoma’s AI filing rule. July and August include the newest appellate and trial-court rows outside sanctions: Bryan v. City of Philadelphia, ShotSpotter appeals, fee arguments about Strongsuit and several self-represented AI-use decisions.
The users were mostly self-represented, while penalties clustered around counsel
The actor table answers who put the fabricated material into court filings or decisions. Self-represented litigants account for 310 rows, lawyers for 192, and five rows are coded to a firm, an expert or another actor. The outcome table shows a second pattern: warnings are the common response, while fines, referrals, costs and dismissals are smaller subsets.
| Actor in fabricated-citation rows | Decisions |
|---|---|
| Self-represented litigant | 310 |
| Lawyer | 192 |
| Firm | 2 |
| Other actor | 2 |
| Expert | 1 |
| Outcome in fabricated-citation rows | Decisions |
|---|---|
| Warning | 208 |
| No consequence noted | 74 |
| Pending | 65 |
| Fine | 31 |
| Referral | 28 |
| Filing struck | 26 |
| Costs order | 23 |
| Dismissal | 21 |
| Sanctions | 20 |
| Other | 7 |
| Suspension | 4 |
The requested warning, fine, referral and dismissal categories together account for 288 rows. The remaining outcomes include sanctions and costs orders, which are separate codes because some decisions imposed a sanction without a separately coded fine, and some shifted costs without the sanctions label. Sixty-one rows carry a fixed monetary penalty, totaling USD 338,497.09.
No fabricated-citation row is now coded strike-off, which is reserved for removal from the roll. Rows that had used that code loosely were re-coded to filing-struck, sanctions or dismissal according to the consequence the court imposed in that document.
The named-tool record is thinner than the actor record. The court’s document names a tool in 38 of 507 fabricated-citation rows and names none in 469. Generic labels, such as AI, GAI or generative AI, are no longer counted as tools unless the court names a product. A separate ai_mention field shows that in 111 of the 507 fabricated-citation rows the court never itself names AI; it says only that the authorities do not exist or are misquoted.
| Tool category | Decisions |
|---|---|
| Not named | 469 |
| ChatGPT / OpenAI | 17 |
| Westlaw CoCounsel | 4 |
| Google AI / Gemini / Bard | 3 |
| LexisNexis / Protégé | 3 |
| Claude | 2 |
| Perplexity | 2 |
| Casetext; vLex | 1 |
| Centient AI | 1 |
| Deepseek | 1 |
| Eve Legal | 1 |
| Fastcase | 1 |
| First Drafts | 1 |
| MX2.law | 1 |
The product table is conservative. It does not infer ChatGPT from the word AI or allocate a multi-product row to each product. It answers what the court record named, not which tools were actually used.
Federal district courts supplied most of the fabricated-citation record
Federal courts supply 334 of the 507 fabricated-citation decisions, state courts 172 and other bodies 1. The main court-level finding is narrower: federal district courts account for 293 rows. The state side is distributed across appellate, supreme and trial courts.
| System | Fabricated-citation decisions |
|---|---|
| Federal | 334 |
| State | 172 |
| Other | 1 |
| Court level | Fabricated-citation decisions |
|---|---|
| Federal district courts | 293 |
| State appellate courts | 131 |
| State trial courts | 25 |
| Federal courts of appeals | 21 |
| State supreme courts | 16 |
| Federal bankruptcy courts | 10 |
| Federal specialty courts | 10 |
| Other | 1 |
The top state table counts the location recorded for each row. It therefore includes a row for federal decisions with no state code. California, New York, Texas, Illinois and Florida are the top five state-coded locations in the fabricated-citation data.
| Location recorded | Fabricated-citation decisions |
|---|---|
| California | 48 |
| New York | 42 |
| federal no state | 32 |
| Texas | 30 |
| Illinois | 25 |
| Florida | 23 |
| Arizona | 20 |
| Michigan | 18 |
| Indiana | 15 |
| Pennsylvania | 15 |
| Washington | 15 |
| Ohio | 14 |
| Alabama | 13 |
| Oregon | 13 |
| Maryland | 12 |
Recurring courts are concentrated in federal districts. The Central District of California and Southern District of New York each appear 17 times after name normalisation, the Eastern District of Michigan and Northern District of Illinois 15 each, and the District of Arizona 11.
| Court, normalised name | Fabricated-citation decisions |
|---|---|
| United States District Court for the Central District of California | 17 |
| United States District Court for the Southern District of New York | 17 |
| United States District Court for the Eastern District of Michigan | 15 |
| United States District Court for the Northern District of Illinois | 15 |
| United States District Court for the District of Arizona | 11 |
| United States District Court for the District of Nevada | 9 |
| United States District Court for the Northern District of California | 9 |
| United States District Court for the Western District of Washington | 9 |
| United States District Court for the District of Oregon | 8 |
| United States District Court for the Northern District of Alabama | 8 |
The fifty-nine non-fabrication rows show courts using AI, policing evidence and writing rules
The 59 non-fabrication rows are not one subject. They include direct judicial disclosure of AI use, evidentiary disputes over generated or altered material, privilege and discovery disputes over AI conversations, court rules written by opinion, criminal-justice algorithm decisions and substantive AI law.
| Primary topic | Decisions |
|---|---|
| Self-represented litigants using AI | 18 |
| Court rules adopted by opinion or order | 10 |
| AI-generated evidence and authentication | 7 |
| Algorithms in criminal justice | 6 |
| Substantive AI law courts cite | 5 |
| The court’s own use of AI | 4 |
| Discovery and AI-assisted review | 4 |
| Competence, fees and billing | 3 |
| Privilege and work product over AI use | 2 |
The primary court’s-own-use rows begin in the Eleventh Circuit. In Snell v. United Specialty Insurance Co., Judge Newsom’s concurrence linked the experiment to ordinary meaning. The source is the Eleventh Circuit PDF:
Perhaps in a fit of frustration, and most definitely on what can only be described as a lark, I said to one of my clerks, “I wonder what ChatGPT thinks about all this.” So he ran a query: “What is the ordinary meaning of ‘landscaping’?” Here’s what ChatGPT said in response: “Landscaping” refers to the process of altering the visible features of an area of land, typically a yard, garden or outdoor space, for aesthetic or practical purposes.
In United States v. Deleon, the same judge described a second use and the problem of variable model answers. The source is the Eleventh Circuit PDF:
“Those, like me, who believe that ‘ordinary meaning’ is the foundational rule for the evaluation of legal texts should consider—consider—whether and how AI-powered large language models like OpenAI’s ChatGPT, Google’s Gemini, and Anthropic’s Claude might—might—inform the interpretive analysis.” Id. at 1221. With the benefit of a little perspective, and incorporating by reference here all the caveats that I expressed there, I stand by what I said.
The District of Columbia Court of Appeals then treated ChatGPT more skeptically in Ross v. United States. The source is the D.C. Courts PDF:
I do not mean to suggest that ChatGPT is a good proxy for what is, and isn’t, common knowledge—it is definitely not. It can answer questions that are far from common knowledge. For instance, when I asked it what the forty-first element in the periodic table is, it responded (apparently correctly) with Niobium (Nb).
The fourth primary court-use row, Smith v. P.A.M. Transport, Inc., is a Sixth Circuit concurrence using ChatGPT in a statutory-meaning discussion. A fifth disclosure appears in Gouveia v. Meridian Financial Investments, LLC, a fabricated-citation row where the court included a ChatGPT-generated limerick while cautioning a self-represented appellant. Together, the five rows show disclosure, not a hidden methodology. The count is still small.
| Court disclosed AI use | Decisions |
|---|---|
| True | 5 |
| False | 561 |
Evidence, privilege and rules are being decided outside the sanctions lane
The non-fabrication rows matter because they contain rules of proof and confidentiality as well as discipline for bad citations. New York’s high court addressed authentication in a deepfake setting; federal courts split on AI-chat privilege and work product; Oklahoma and Arkansas wrote AI duties by published court order.
| Decision | Date | Main topic | Tool named | Court URL |
|---|---|---|---|---|
| Matter of M.S. (M.H.) | 2026-02-17 | AI-generated evidence and authentication | none | source |
| Bryan v. City of Philadelphia | 2026-07-17 | AI-generated evidence and authentication | none | source |
| United States v. Heppner | 2026-02-17 | Privilege and work product over AI use | Claude | source |
| Morgan v. V2X, Inc. | 2026-03-30 | Discovery and AI-assisted review | ChatGPT; Claude; Gemini; Harvey.AI; Bard | source |
| In re Addition of a New Rule to the Rules of the Court of Criminal Appeals | 2026-02-18 | Court rules adopted by opinion or order | none | source |
| In re Adoption of Arkansas Supreme Court Administrative Order No. 25 | 2025-12-11 | Court rules adopted by opinion or order | none | source |
In Matter of M.S., the New York Court of Appeals rejected authentication built on matching general details in a video to the scene. The source is the New York official report page:
The fact that much of the video apparently accurately depicted the home is not sufficient—as it was not in Patterson—to authenticate the video.
In Bryan v. City of Philadelphia, the Third Circuit recorded a prisoner’s allegations that police body-camera footage had been manipulated with AI. The source is the Third Circuit PDF:
The arresting officer manipulated body camera footage using artificial intelligence, body camera footage for all other on-scene officers was also altered, dash camera footage went “missing,” and footage that played in court was altered to remove scenes.
In United States v. Heppner, the Southern District of New York held that Claude exchanges were neither privileged nor work product. The source is the CourtListener RECAP PDF:
Because Heppner’s use of Claude fails to satisfy either of these rules, the AI Documents do not merit the protections Heppner has claimed.
In Morgan v. V2X, Inc., the District of Colorado protected a pro se litigant’s work product but changed the protective order. The source is the govinfo PDF:
No party or authorized recipient may input, upload, or submit CONFIDENTIAL Information into any modern artificial intelligence platform, including any generative, analytical, or large language model-based tool (“AI”), unless the AI provider is contractually prohibited from: (1) storing or using inputs to train or improve its model; and (2) disclosing inputs to any third party except where such disclosure is essential to facilitating delivery of the service.
Oklahoma’s criminal appellate court adopted a filing rule in 2026 OK CR 7. The source is the Oklahoma State Courts Network page:
When generative artificial intelligence (“generative AI”) has been used in the drafting of any document for filing in this Court, the party, or their counsel, shall ensure that any portion of the document produced or modified by generative AI, whether in whole or in part, has been verified as accurate by a person responsible for the document.
Arkansas adopted Administrative Order No. 25 in In re Adoption of Arkansas Supreme Court Administrative Order No. 25. The source is the Arkansas judiciary document page:
Everyone participating in the court system must be mindful of the following when entering client or court data into any electronic system that generates responses or uses generative artificial intelligence (GAI): (a) Certain GAI tools retain the data submitted into their system and use it to keep building their large language models (LLM), or what you would consider their database.
The unauthorized-practice subject does not appear as the main topic in the 566 decision rows. It appears in the litigation sidecar instead, where three dockets are coded as consumer unauthorized-practice matters: smartlaw, MillerKing v. DoNotPay and Faridian v. DoNotPay.
Substantive AI law appears as a source for legal-practice courts
The corpus contains 5 rows on substantive AI law, not practice misuse. They matter here because courts and litigants in legal-practice disputes cite AI copyright and authorship cases when arguing about training data, legal databases, legal research tools and machine output.
| Substantive AI law row | Date | Court | Tool or system named |
|---|---|---|---|
| Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc. | 2025-02-11 | U.S. District Court for the District of Delaware | ROSS AI search tool |
| Thaler v. Perlmutter | 2025-03-18 | U.S. Court of Appeals for the D.C. Circuit | Creativity Machine |
| Bartz v. Anthropic PBC | 2025-06-23 | U.S. District Court for the Northern District of California | Claude |
| Kadrey v. Meta Platforms, Inc. | 2025-06-25 | U.S. District Court for the Northern District of California | Llama |
| United States v. Anderegg | 2026-08-25 | U.S. Court of Appeals for the Seventh Circuit | Stable Diffusion |
In Thomson Reuters v. ROSS, the Delaware court described an AI legal-search product that was not generative AI. The source is the District of Delaware PDF:
Ross was using Thomson Reuters’s headnotes as AI data to create a legal research tool to compete with Westlaw. It is undisputed that Ross’s AI is not generative AI (AI that writes new content itself). Rather, when a user enters a legal question, Ross spits back relevant judicial opinions that have already been written.
In Bartz v. Anthropic, the Northern District of California described Anthropic as an AI software firm whose core offering is Claude, and described the books at issue as part of a central library used to train language models. The source is the govinfo PDF.
In Thaler v. Perlmutter, the D.C. Circuit decided the human-authorship question. The source is the D.C. Circuit PDF:
The Creativity Machine cannot be the recognized author of a copyrighted work because the Copyright Act of 1976 requires all eligible work to be authored in the first instance by a human being.
These substantive rows should not be mixed with sanctions rows. They decide copyright, authorship or criminal-evidence questions. Their value in this report is that they are the external AI-law citations a legal-practice court may import when a dispute concerns legal databases, legal-search systems or generated images.
The nine legal-AI dockets sit beside the court-opinion corpus
The litigation sidecar is a posture table rather than an incident table. It records 9 dockets involving legal-AI products, legal databases, consumer legal services, trade secrets and regulator claims. Three are judgments, two are dismissed, and one each is active, settled, on appeal or under a consent order.
| Case | Court | Filed | Category | Status | Current posture as coded |
|---|---|---|---|---|---|
| West Publishing Corporation v. LegalEase Solutions, LLC | U.S. District Court for the District of Minnesota | 2018-05-24 | Copyright and AI training | Judgment | Consent judgment and stipulated permanent injunction entered May 5, 2020; case closed. |
| Éditions Dalloz, Lexbase, LexisNexis, Lextenso, Lamy Liaisons v. Forseti (Doctrine) | Cour d’appel de Paris, Pôle 5, Chambre 1 | 2018-10 | Legal database terms | Judgment | Paris Court of Appeal judgment dated May 7, 2025 found unfair competition and awarded damages to legal publishers; official court page was blocked to automated access. |
| Thomson Reuters Enterprise Centre GmbH and West Publishing Corp. v. ROSS Intelligence Inc. | U.S. District Court for the District of Delaware; U.S. Court of Appeals for the Third Circuit | 2020-05-06 | Copyright and AI training | On appeal | District-court summary judgment issued February 11, 2025; Third Circuit No. 25-2153 was argued June 11, 2026 and no opinion was found as of 2026-09-08. |
| Hanseatische Rechtsanwaltskammer Hamburg v. Wolters Kluwer Deutschland GmbH (smartlaw) | Bundesgerichtshof | 2021-09-09 | Unauthorized practice / consumer | Judgment | BGH judgment issued September 9, 2021 affirmed dismissal of the bar chamber’s challenge and held the smartlaw document generator was not an RDG legal service. |
| MillerKing, LLC v. DoNotPay, Inc. | U.S. District Court for the Southern District of Illinois | 2023-03-15 | Unauthorized practice / consumer | Dismissed | Complaint was dismissed for lack of Article III standing; a November 28, 2023 order clarified any dismissal would be without prejudice, and no Seventh Circuit appeal was found. |
| Jonathan Faridian v. DoNotPay, Inc. | U.S. District Court for the Northern District of California | 2023-04-07 | Unauthorized practice / consumer | Dismissed | Public docket leads report a stipulated dismissal with prejudice on July 30, 2024; free final dismissal paper was not located. |
| In the Matter of DoNotPay, Inc. | Federal Trade Commission | 2024-09-25 | Regulator action | Consent order | FTC final decision and order issued January 14, 2025; DoNotPay owes monetary relief and must comply with notice, substantiation, reporting, and recordkeeping terms. |
| Canadian Legal Information Institute v. Caseway AI | Supreme Court of British Columbia | 2024-12 | Legal database terms | Settled | CanLII and Caseway announced on March 20, 2026 that all matters in VLC-S-S-247574 were fully and finally resolved on confidential terms. |
| EvenUp, Inc. v. Butler Labs, Inc. | U.S. District Court for the Northern District of California | 2025-09-26 | Trade secrets between vendors | Active | Docket leads report Judge Yvonne Gonzalez Rogers denied EvenUp’s preliminary-injunction motion on March 19, 2026; no free court PDF was located. |
Thomson Reuters v. ROSS is the docket to watch because it is the active appellate test in this sidecar. The litigation row says the Third Circuit argument occurred on 11 June 2026 and that no opinion was found by 8 September 2026. The decision row for the district-court opinion remains in the substantive AI law table above.
What to watch
The September 2026 digest is incomplete. The dataset date is 8 September, and only two fabricated-citation decisions appear for the month. Later September dockets could change the monthly table without changing any conclusion about February through August.
The provisional incident queue is the second watch item. The court-decision corpus has 507 fabricated-citation rows; 505 of them have tracker slugs that can be reconciled with the incident workflow. As those rows move into verified status, the verified tracker will gain more actor, outcome, tool and penalty fields that can be compared with this corpus.
The Third Circuit’s decision in No. 25-2153, Thomson Reuters v. ROSS, had not been located when the dataset was built on 8 September 2026; the litigation page will be updated. The next opinion there will affect legal-database and AI-training disputes beyond the parties to that case. The Arizona, New York, Colorado and federal evidence-rule records outside this corpus will also determine whether AI-generated or AI-altered evidence is treated as an authentication problem, a discovery problem or both.
Appendix A — data tables
A1. Main topic counts
| Main topic | Decisions |
|---|---|
| Fabricated or misquoted citations | 507 |
| Self-represented litigants using AI | 18 |
| Court rules adopted by opinion or order | 10 |
| AI-generated evidence and authentication | 7 |
| Algorithms in criminal justice | 6 |
| Substantive AI law courts cite | 5 |
| The court’s own use of AI | 4 |
| Discovery and AI-assisted review | 4 |
| Competence, fees and billing | 3 |
| Privilege and work product over AI use | 2 |
A2. Fabricated-citation rows by tool category named by the court
| Tool category | Decisions |
|---|---|
| Not named | 469 |
| ChatGPT / OpenAI | 17 |
| Westlaw CoCounsel | 4 |
| Google AI / Gemini / Bard | 3 |
| LexisNexis / Protégé | 3 |
| Claude | 2 |
| Perplexity | 2 |
| Casetext; vLex | 1 |
| Centient AI | 1 |
| Deepseek | 1 |
| Eve Legal | 1 |
| Fastcase | 1 |
| First Drafts | 1 |
| MX2.law | 1 |
A3. Courts with the most fabricated-citation decisions
| Court, normalised name | Decisions |
|---|---|
| United States District Court for the Central District of California | 17 |
| United States District Court for the Southern District of New York | 17 |
| United States District Court for the Eastern District of Michigan | 15 |
| United States District Court for the Northern District of Illinois | 15 |
| United States District Court for the District of Arizona | 11 |
| United States District Court for the District of Nevada | 9 |
| United States District Court for the Northern District of California | 9 |
| United States District Court for the Western District of Washington | 9 |
| United States District Court for the District of Oregon | 8 |
| United States District Court for the Northern District of Alabama | 8 |
A4. Litigation sidecar status and qualifier counts
| Status | Dockets |
|---|---|
| Judgment | 3 |
| Dismissed | 2 |
| Active | 1 |
| Consent order | 1 |
| On appeal | 1 |
| Settled | 1 |
| Category | Dockets |
|---|---|
| Unauthorized practice / consumer | 3 |
| Copyright and AI training | 2 |
| Legal database terms | 2 |
| Regulator action | 1 |
| Trade secrets between vendors | 1 |
The complete row-level dataset is the JSON export at /courts/opinions.json. No appendix table above exceeds 60 rows.
Correction, 8 September 2026: an earlier version described Damien Charlotin’s AI Hallucination Cases Database as CC0. Its author confirmed on 8 September 2026 that the licence is CC BY 4.0; the attribution above has been corrected and the dataset re-released as v0.1.2 with the same change.
Revision, 8 September 2026: an independent re-read of a 50-row sample found coding errors in 22 rows (44%); every row was then re-read from the court document and the dataset re-released as v0.2.0, then v0.2.1 the same day (566 decisions). All counts in this report were recomputed; the outcome, actor and tool tables changed materially. The first version’s figures are preserved in the dataset’s v0.1.x releases. Details: /courts#how-sure.
Sources
Primary court documents linked in this report:
- Snell v. United Specialty Insurance Co., U.S. Court of Appeals for the Eleventh Circuit (28 May 2024): PDF
- United States v. Deleon, U.S. Court of Appeals for the Eleventh Circuit (5 September 2024): PDF
- Ross v. United States, District of Columbia Court of Appeals (20 February 2025): PDF
- Smith v. P.A.M. Transport, Inc., U.S. Court of Appeals for the Sixth Circuit (25 September 2025): PDF
- Gouveia v. Meridian Financial Investments, LLC, District Court of Appeal of Florida, Fourth District (25 March 2026): CourtListener PDF
- Matter of M.S. (M.H.), New York Court of Appeals (17 February 2026): official report
- Bryan v. City of Philadelphia, U.S. Court of Appeals for the Third Circuit (17 July 2026): PDF
- United States v. Heppner, U.S. District Court for the Southern District of New York (17 February 2026): RECAP PDF
- Warner v. Gilbarco, Inc., U.S. District Court for the Eastern District of Michigan (10 February 2026): RECAP PDF
- Morgan v. V2X, Inc., U.S. District Court for the District of Colorado (30 March 2026): govinfo PDF
- In re Addition of a New Rule to the Rules of the Court of Criminal Appeals, Oklahoma Court of Criminal Appeals (18 February 2026): OSCN page
- In re Adoption of Arkansas Supreme Court Administrative Order No. 25. Artificial Intelligence, Supreme Court of Arkansas (11 December 2025): Arkansas judiciary page
- Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., U.S. District Court for the District of Delaware (11 February 2025): PDF
- Bartz v. Anthropic PBC, U.S. District Court for the Northern District of California (23 June 2025): govinfo PDF
- Thaler v. Perlmutter, U.S. Court of Appeals for the D.C. Circuit (18 March 2025): PDF
- Kadrey v. Meta Platforms, Inc., U.S. District Court for the Northern District of California (25 June 2025): govinfo PDF
- United States v. Anderegg, U.S. Court of Appeals for the Seventh Circuit (25 August 2026): Seventh Circuit page
- In the Matter of DoNotPay, Inc., Federal Trade Commission final decision and order (14 January 2025): FTC final-order PDF
Dataset JSON URLs:
- SafeLegalAI court decisions JSON export: /courts/opinions.json
- SafeLegalAI court decisions landing page: /courts
- SafeLegalAI litigation sidecar rows, including source links and review notes: /courts/litigation/thomson-reuters-v-ross
Archive snapshots:
- Many fabricated-citation rows were read from the court PDF held by Damien Charlotin’s AI Hallucination Cases Database (CC BY 4.0); row-level mirror links and SHA-256 hashes are in /courts/opinions.json.
Reuse this research
Quote it, cite it, forward it — CC BY 4.0 for the data and figures; the linked official documents are the record. Suggested citation and a link to this exact report: