Skip to content

Courts on AI

What courts have said about generative AI, decision by decision

Every court-authored US document since November 2022 that discusses generative AI, or addresses the fabricated and misquoted authorities courts have come to associate with it — opinions, orders, concurrences, dissents, administrative orders — read from the court's document, coded by topic, and quoted. United States court documents are public domain, so each page carries the passage itself rather than a paraphrase; a column on every row says whether the court itself names AI. SafeLegalAI records what the court wrote; it does not rank, interpret or advise. Fabricated-citation decisions keep their canonical record in the incident tracker; here they sit beside everything else courts have said.

compiled by the SafeLegalAI editorial teammethodas of 2026-09-08v0.2.2coding CC BY 4.0court text public domain (17 U.S.C. §105; Banks v Manchester)JSON Hugging Face GitHubevidence registerlitigation docketsthe report: 567 decisions, read

decisions, one page each
beyond fabricated citations
courts disclosing their own AI use
evidence rulings worldwide
legal-AI dockets tracked

What the record showsas of 2026-09-08

Counts and shares computed from the dataset at build time. They describe what the record contains; they do not rank, evaluate or advise.

Eleven things courts talk about

A decision can carry several topics; each topic page lists every decision that touches it, newest first, with the passage.

Fabricated or misquoted citations508Decisions in which a court addresses authorities that did not exist or did not say what a filing claimed. Each has a canonical record in the incident tracker; this corpus holds the passage.The court's own use of AI7Opinions, concurrences and orders in which a court says it used a generative-AI tool, or discusses whether courts should.AI-generated evidence and authentication10Rulings on AI-generated, AI-altered or AI-enhanced material offered as evidence, and on deepfake challenges. The evidence register codes these in detail.Discovery and AI-assisted review9Orders on generative AI in document review, technology-assisted review protocols and AI-assisted privilege logs.Privilege and work product over AI use3Whether prompts, outputs and chat histories with AI tools are privileged or protected work product.AI services and unauthorized practice1Decisions on AI services offering legal help to the public and the unauthorized practice of law.Algorithms in criminal justice8Court analysis of risk-assessment tools, forensic software, facial recognition and AI-drafted reports in criminal proceedings.Self-represented litigants using AI33Courts addressing a self-represented litigant's use of AI other than fabricated authorities: AI-drafted pleadings, requests for AI assistance, conditions set on its use.Competence, fees and billing72Fee awards and reductions that turn on AI-assisted work, and competence discussions outside sanctions.Court rules adopted by opinion or order25Courts adopting a rule on AI in proceedings by published opinion or administrative order. The instruments themselves are in the regulation tracker.Substantive AI law courts cite6The handful of AI copyright, authorship and platform decisions that courts in legal-practice cases cite.

Latest decisions

Monthly digests: 2026-09 · 2026-08 · 2026-07 · 2026-06 · 2026-05 · 2026-04 · 2026-03 · 2026-02 · 2025-12 · 2025-10 · 2025-09 · 2025-06

  1. Douglas v. Deutsche Bank National Trust Company

    · District of Columbia Court of Appeals · Fabricated or misquoted citations

    The District of Columbia Court of Appeals strikes Deutsche Bank’s appellee brief after counsel admits that four cited authorities generated through Google’s AI search did not exist. The court also refers the matter to disciplinary counsel and stresses that signed appellate briefs require verified authorities.

  2. In re the Marriage of Monson

    · Iowa Court of Appeals · Fabricated or misquoted citations

    The Iowa Court of Appeals identifies likely AI misuse in a self-represented divorce appellant’s brief, including nonexistent and mischaracterized authorities. The court still considers discernible claims, affirms the decree, and warns that inaccurate AI-assisted filings waste resources.

  3. State v. Presti

    · Intermediate Court of Appeals of Hawaiʻi · Fabricated or misquoted citations

    The Hawaiʻi Intermediate Court of Appeals affirms a criminal judgment while documenting numerous nonexistent or misidentified cases in Cody Presti’s self-represented opening brief. The court explains sanction authority for fake citations but declines to apply civil Rule 11 procedures in the direct criminal appeal.

  4. Reyes v. Bailey

    · United States District Court for the District of New Mexico · Fabricated or misquoted citations

    The District of New Mexico denies self-represented plaintiff Charles Reyes’s Rule 11 motion. The court holds defendants reasonably said Reyes may have fabricated a case citation because Reyes admitted he miscited a case in his amended complaint, and the remaining alleged misrepresentations did not support sanctions.

  5. Noble v. WMC Mortgage Corp.

    · Appellate Court of Illinois, First District · 2026 IL App (1st) 251168-U · Fabricated or misquoted citations

    The Illinois Appellate Court affirms denial of Sanja Noble’s petition and addresses a hallucinated citation in her self-represented brief. The order says Noble cited Ferguson v. Georges, but the reporter citation corresponded to People v. Mott and did not support the standard of review.

  6. Booker v. The Kroger Co.

    · United States District Court for the Northern District of Georgia · Fabricated or misquoted citations

    The Northern District of Georgia sanctions plaintiff’s counsel in Booker v. Kroger after finding fake, false and misleading case authorities and an untruthful denial of AI use. The order imposes an $8,000 court fine and requires documentation of completed ethics and technology CLE.

  7. State v. Booker

    · Ohio Court of Appeals, First Appellate District · 2026-Ohio-3371 · Algorithms in criminal justice

    The Ohio appellate court considers a suppression challenge arising from ShotSpotter and real-time crime-center surveillance. It describes how police used automated gunshot detection and video to locate suspects before affirming the search and conviction.

  8. Williams v. Dallas-Fort Worth International Airport Board

    · United States District Court for the Northern District of Texas · Fabricated or misquoted citations

    The Northern District of Texas dismisses Williams’s Section 1983 claims and warns that plaintiffs’ qualified-immunity briefing relied on nonexistent or inapposite cases. The court identifies Dawkins, Iko and Yates problems and cautions that generative-AI misrepresentations of law may trigger Rule 11 sanctions.

  9. Kiser v. DeSimone Auto Group

    · Superior Court of Pennsylvania · Fabricated or misquoted citations

    The Superior Court of Pennsylvania affirms judgment against Paige and Ty Kiser and identifies an unlocatable Com v. Reid citation in counsel’s appellate brief. The memorandum treats the unsupported argument as waived and quotes prior authority warning that counterfeit citations may be chatbot products.

  10. James v. Conley

    · United States District Court for the Southern District of Florida · AI-generated evidence and authentication

    The Southern District of Florida recommends denying Miami-Dade County’s Rule 11 motion against James on safe-harbor grounds but says counsel relied for months on unverified AI-generated IACP Training Keys. The recommendation calls the Keys AI hallucinations and recommends referral of four signing attorneys.

  11. Adams v. Matrix Providers Inc.

    · United States District Court for the District of Colorado · Fabricated or misquoted citations

    The District of Colorado denies Adams’s post-judgment motions and addresses counsel’s response to a show-cause order over a nonexistent Ricks v. Starbucks citation and other inaccurate authorities. The court admonishes counsel about possible AI-assisted research errors and separately imposes a $1,000 sanction for rule violations.

  12. Unger v. Unger

    · Florida Fourth District Court of Appeal · Self-represented litigants using AI

    The Florida appellate court repeats a prior warning to a self-represented former husband after striking his initial brief with leave to amend. The opinion cautions him about using artificial intelligence in documents filed with the court while resolving contempt issues.

By court level

Federal courts of appeals28
Federal district courts319
Federal bankruptcy courts10
Federal specialty courts10
State supreme courts26
State appellate courts145
State trial courts26
Other2

Legal-AI litigation

Proceedings in which a legal-AI company or product is a party, with dated milestones. All dockets →

How sure this record is

Every row was coded from the court-authored document, not from a summary. Where the document was read matters, so each row says: 300 were read from the court's own copy (the court's site, govinfo.gov or the public RECAP archive) and 266 from a public copy of the court's PDF because the court's own copy was not found online when checked — the court does not publish the order, or its site refuses the identified bot, or the host was down; each such row says which, and is marked so on its page. In 435 decisions the court itself names AI, a chatbot or "hallucination"; in 118 the court says only that the authorities do not exist or are misquoted — those rows are in the corpus because the pattern is the one courts have associated with generative AI, and the ai_mention column lets a reader exclude them.

This release is the first after an independent audit. On 8 September 2026 a 50-row random sample was re-read by an agent that had not written the rows: every quoted passage was verbatim and every case, court and date was right, but 22 of the 50 rows (44%) had at least one error in a coding column — the outcome enum used loosely ("other" as a catch-all; "strike-off", which means removed from the roll, used for a filing being struck), docket fields holding caption fragments, a generic "AI" label recorded as a named tool, and passages that never reached the court's AI reference. Because the errors were systematic, every row was then re-read from the document by a second, independent pass against a written rubric (work/recode/RUBRIC.md); the changes were applied through a validated overlay with a field-level log (work/recode/applied-2026-09-08-*.jsonl).

  • · Independent re-read of a 50-row random sample (seed 20260908). 22 rows with ≥1 coding error (44%); 0 passage or identification errors. Threshold for a full re-read was 5%.
  • · Full second read of every row from the court document, six independent agents, one rubric. 555 of 569 rows re-coded; 14 could not be re-fetched (court hosts unreachable that day) and carry "Second read pending" in notes; 1 row withdrawn because the document was counsel's letter, not a court document. 3,025 field changes across 568 rows: 197 outcomes re-coded (58 from "other" to the new "none"; all 7 "strike-off" rows re-coded — none was a removal from the roll), 343 docket strings normalised, 95 generic "AI" labels moved out of the tool field into the new ai_mention column, 5 placeholder penalties set to null, topics changed on 133 rows.
  • · Official-copy resolution: GovInfo (USCOURTS) API by docket with the quoted passage required to match; state-court archives by hand-briefed agents; CourtListener RECAP search within Free Law Project's published limits. (CourtListener: 106 searches on 8 Sep → 6 rows.) 300 of 566 rows now read from the court's own copy or the public RECAP archive (from 53). 266 remain read from a public copy of the court's PDF: the court does not publish the order online, or its site refuses the identified bot (69 rows, queued for hand download), or the host timed out.
  • · Inter-rater check: one 95-row slice was independently coded twice from the same documents (the first agent finished late; its output was kept as a cross-check, not merged). Agreement on ai_mention 100%, monetary penalty 100%, tool 99%, primary topic 98%, actor 97%, outcome 89% (10 rows, mostly warning vs none/filing-struck), topics 81%, document type 76%. The 10 outcome disagreements are listed for the editor in work/recode/crosscheck-4-disagreements.json; the published coding is the second read's.

The ai_mention column says whether the court itself names AI (435 rows), or only says the authorities do not exist or are misquoted (118); the corpus includes the last group because the pattern is the one courts associate with generative AI, and readers who want only decisions that name AI can filter on it. The outcome column is the most severe consequence the court imposed in that document, on a defined ladder; it is not a finding about any person. A quarterly 50-row independent re-read is standing practice and its result replaces this one.

Found an error? Every decision page has a report link; the repository takes row-error issues. Corrections are made at the source file and released with a note.

How this is made

Leads come from Damien Charlotin's AI Hallucination Cases Database (CC BY 4.0, damiencharlotin.com; each decision re-read and coded by SafeLegalAI), the public CourtListener index (discovery only, within its published limits; a partnership request to Free Law Project is pending), court-website searches and news. Each document is then read — from the issuing court's site, govinfo.gov or the public RECAP archive where the court publishes it, otherwise from the public copy of the court's PDF — and the passage is quoted so that readers can check the coding against it. Rows whose court-authored document has not been read at all are marked provisional and are not indexed until the editor clears them. Method and schema: editorial standards · repository.