AI Litigation Tracker: Federal Court Cases, Timelines, Product Liability and Legal Updates

Which AI cases are included in the tracker

The tracker indexes civil actions filed in U.S. federal district courts and state jurisdictions involving artificial intelligence technologies. Eligible matters include claims about generative ai training datasets, algorithmic discrimination, copyright infringement, trade secrets, consumer privacy violations, product liability for conversational agents, and unauthorized automated data access. Regulatory enforcement actions and agency policy initiatives are cataloged separately, so a rulemaking consultation never gets mistaken for a live docket.

Inclusion boundaries are applied consistently. A matter enters the tracker when a complaint is docketed, not when a policy proposal, consultation or guidance note is published. Employment-screening, resume-parsing, monitoring and notetaking matters carry a state-level tag, because a meaningful share of AI litigation now starts outside federal court.

Counts differ across public trackers because inclusion rules differ. Some monitor only copyright dockets. Others fold in privacy, deepfake, antitrust, securities and consumer-protection claims, and a few keep settled or appeal-stage matters inside the active set.

How to read case status and court records links

Latest legal updates on AI court cases

Recent ai litigation tracker legal updates show federal courts tightening scrutiny on training data procurement and on generative outputs.

• Antitrust and competition. In United States v. Google (D.D.C., Sept. 2, 2025), the court declined Chrome and Android divestiture but imposed behavioral remedies restricting exclusive contracts covering Google Search, Chrome, Google Assistant and Gemini.
• Discovery. In the S.D.N.Y. multi-district litigation, the district court affirmed a magistrate order compelling production of a 20-million de-identified ChatGPT log sample (January 2026). OpenAI's objection was denied in April 2026.
• Court procedure. Southern District of Texas General Order 2025-04 requires counsel and self-represented litigants to verify the factual and legal accuracy of AI-assisted filings.
• Patent inventorship. USPTO 2024 guidance permitted patenting AI-assisted inventions where one natural person made a significant contribution; the 2025 revision reaffirmed that AI is a tool and conception must be human.
• Music and video models. Sony v. Uncharted Labs (Udio) survived a motion to dismiss in April 2026, and Disney et al. v. MiniMax & Hailuo AI moved into motion-to-dismiss briefing the same month.

Case statuses: from complaint to judgment

Federal civil litigation moves sequentially from filing to final verdict under the Federal Rules of Civil Procedure. A case begins with a complaint, progresses to Rule 12(b)(6) motions to dismiss, enters Rule 26–37 discovery, and reaches Rule 56 summary judgment before trial. Governance teams read these stages to judge when a third-party model faces heightened injunction or settlement risk.

Each stage carries a different governance meaning. Litigation stage, not headline volume, should drive internal escalation.

Training data, output and copyright use in AI cases

Copyright analysis separates training data ingestion from model output generation. The U.S. Copyright Office 2025 Part 3 Report notes that training implicates reproduction rights under 17 U.S.C. § 106 during ingestion, including downloading, transferring, format conversion and temporary copies. Output claims work differently: they require proof of substantial similarity and actual copying of protected expression. That is the analysis deciding whether outputs from AI image and art generators or AI voice generators can be used commercially without risk.

Empirical research supports the plaintiffs' output theory in a narrow but important way. Large language models can memorize and, under specific prompting, emit verbatim or near-verbatim fragments of training data. That capability is the technical bridge between an ingestion claim under § 106 and a regurgitation claim tested against substantial similarity. It also explains why a court such as LG München I treated memorization itself as reproduction.

UK analysis frames the question slightly differently. It emphasizes that copies contained in a model or in its outputs may engage communication, distribution and secondary-infringement rules, so the location of the copy matters as much as the training act. Small distinction on paper. Large consequence for a group operating on both sides of the Atlantic.

Federal court decisions and case summaries

Every case summary must separate factual allegations from binding judicial holdings. In Thaler v. Perlmutter (D.C. Cir., March 18, 2025), the court affirmed that copyright requires human authorship, ruling that pure AI outputs are ineligible for registration while AI-assisted works may qualify where a human is the author using AI as a tool. In Thomson Reuters v. Ross Intelligence (D. Del., Feb. 11, 2025), by contrast, the court rejected a fair use defense where AI training used proprietary Westlaw headnotes to build a commercial substitute.

Two further decisions matter for annotation discipline. In AI Visualize v. Nuance Communications (Fed. Cir., April 4, 2024), the court affirmed dismissal of AI-related patent claims as ineligible under 35 U.S.C. § 101. In Percipient.ai v. United States, a precedential Federal Circuit opinion issued June 7, 2024 was later vacated by order on November 22, 2024, with the appeal reinstated. A useful reminder, actually: a case summary must always state the live posture, never a superseded panel holding.

In Andersen et al. v. Stability AI, the products at issue include diffusion-based image systems distributed to consumers. That is why enterprise creative teams evaluating AI art generators should track this docket alongside licensing terms.

Five-step integration workflow

Map vendors to dockets.For every model in the inventory, record the developer entity, any parent, and all active matters naming that entity, with case number, court and claim class. Refresh on the same five-business-day cadence used for docket verification.

Assign a legal-risk tier.Score each model on four factors: whether the vendor faces an adverse holding rather than a mere allegation, whether the claim reaches training data your use case depends on, whether outputs are customer-facing, and how strong the indemnity is.

Bind the tier to validation.Raise validation frequency and expand the documentation request when the tier rises. An adverse Rule 56 ruling against a vendor should trigger targeted re-validation of provenance and output-filtering controls, not a full revalidation of everything.

Define escalation triggers in the risk appetite statement.Pre-agree with the board risk committee which docket events require notification (motion to dismiss denied), which require reassessment (summary judgment), and which require contingency execution (injunction, or a settlement that alters model availability).

Report to the audit committee in docket language.Each quarterly pack should state case number, court, judge, stage, last-verified date, and the allegation/holding label. Those are the fields an examiner can independently confirm on PACER.

Vendor contract and indemnity checklist (five clauses derived from 2025–2026 rulings)

• Provenance attestation vendor warrants lawful acquisition of training corpora and represents that no known pirated repository was retained. That is the precise distinction on which Bartz v. Anthropic turned.
• Uncapped or separately capped IP indemnity copyright indemnity carved out of the general liability cap, covering statutory damages and defense costs for outputs generated by the licensed model.
• Litigation notification vendor notifies the institution within a defined window of any new complaint, adverse order, injunction or settlement materially affecting the licensed model.
• Output-filtering and memorization controls contractual commitment to regurgitation mitigation, with evidence available for validation, addressing the memorization risk documented in the German and SDNY output rulings.
• Continuity and substitution rights termination without penalty, model-substitution assistance and data-export rights if an injunction or settlement restricts the model.

Escalation runbook: what a model risk manager does on an injunction alert

1. Confirm the order on the primary docket (PACER/CM/ECF) and capture the PDF for the evidence file.
2. Determine scope: which model versions, jurisdictions and use cases the order reaches.
3. Identify all inventory entries and downstream processes dependent on the affected model.
4. Apply the pre-approved control: restrict to non-customer-facing use, switch to the substitute model, or suspend.
5. Notify legal, procurement and the model risk committee; record the decision and its rationale in the inventory.
6. Reassess residual risk and, where exposure is material and unhedged, discuss an operational-risk capital add-on with finance.
7. Log the closure date and the next verification date.

One caveat on this runbook. Steps four and five collapse if nobody owns the substitute model in advance, which is the most common failure we see described in governance reviews.

OpenAI litigation landscape

OpenAI faces consolidated class actions in the Southern District of New York, including Authors Guild v. OpenAI and The New York Times v. OpenAI, both inside Judge Stein's multi-district litigation. Plaintiffs allege unauthorized copying of millions of books and news articles to train ChatGPT, alongside DMCA § 1202 and trademark-related theories. OpenAI defends on fair use, absence of substantial similarity, and lack of market substitution.

In October 2025 the court denied dismissal of output-infringement claims, holding that short summaries of plaintiffs' works may infringe absent fair use. Discovery disputes through mid-2026 keep circling training logs and model data retention, including the affirmed order to produce a 20-million de-identified log sample. A July 2026 Manhattan filing further accused OpenAI of discovery misconduct. Related matters include Silverman v. OpenAI, filed in N.D. California and narrowed at the dismissal stage before being drawn into the broader consolidation, plus OpenAI, Inc. v. Open Artificial Intelligence, Inc. (N.D. Cal., No. 4:23-cv-03918) on the trademark side. In February 2026, the California federal court dismissed xAI v. OpenAI, ending that trade-secret action at the district-court level.

Meta and dataset liability

In Kadrey v. Meta Platforms, Inc. (N.D. Cal.), Judge Vince Chhabria granted summary judgment for Meta on plaintiffs' reproduction claims, holding on June 25, 2025 that training LLaMA on book datasets was fair use. The DMCA/CMI-removal theory then failed because the underlying copying was not actionable infringement. The court left open a separate theory, though, on dataset distribution via torrents, and plaintiffs pleaded that Meta used the piracy-linked Books3 corpus.

(Updated status) The matter did not end in 2025. Plaintiffs moved to amend their consolidated complaint in December 2025; the court "reluctantly granted" leave in March 2026 and the amended complaint was filed in April 2026. Plaintiffs' bid to certify the summary-judgment issues for interlocutory appeal was denied in July 2026. In Europe, French publishers (SNE v. Meta, March 2025) challenge Meta under EU text and data mining rules, and a further publisher class action naming Meta was filed in SDNY in May 2026 over books and journal articles used to train Llama.

Anthropic legal proceedings

Anthropic reached a milestone in Bartz v. Anthropic (N.D. Cal., No. 3:24-cv-05417, filed Aug. 19, 2024), where the court approved a $1.5 billion class action settlement on July 20, 2026 covering author book claims, after a final fairness hearing in May 2026. Judge William Alsup ruled on June 23, 2025 that model training was "exceedingly transformative" fair use, but held that storing more than seven million pirated books in a central library created separate infringement liability. Two findings, one docket, opposite directions.

Meanwhile, Concord Music Group v. Anthropic PBC (No. 5:24-cv-03811) remains active in N.D. California over song-lyrics training, brought by Concord, Universal Music Group and ABKCO. Anthropic moved for summary judgment on fair use in April 2026 and the fair-use hearing was reset to October 21, 2026. Separately, in Reddit v. Anthropic, the court tentatively granted Reddit's motion to remand in March 2026, keeping the platform's terms-of-service and scraping claims in state court.

(Updated) A previously published sentence in this section referring to an unrelated third-party domain has been removed as an editorial error; the original wording is retained for transparency in Appendix A. Enterprise buyers evaluating any AI vendor, including smaller design and generation tools such as those reviewed in our Canva AI generator overview, should request the same evidence set used above: legal-entity registration, training-data provenance attestation, indemnity scope, and audit rights.

Chatbot harm and product liability

Litigation targeting consumer-facing LLMs has shifted toward defective design and negligence rather than speech. In wrongful-death and self-harm actions against companion and assistant AI developers, including Garcia v. Character.AI, Raine, Soelberg, Peralta and Gavalas v. Google, plaintiffs allege that engagement-maximization architectures create foreseeable risks of self-harm. Courts are testing whether Section 230 immunity extends to dynamically generated conversational output, and the emerging theory frames engagement optimization itself as a design defect. For banks deploying conversational agents in collections, advice or servicing, this line of cases is the closest available proxy for duty-of-care expectations.

Agentic AI and CFAA compliance

As autonomous agents scrape web data and execute API calls, corporate targets are reaching for the Computer Fraud and Abuse Act. In Amazon v. Perplexity, the court is evaluating whether AI agents that bypass robots.txt and anti-scraping protocols incur civil CFAA liability. That ruling lands directly on automated market-intelligence, pricing and vendor-monitoring workflows. Adjacent matters include X v. Bright Data on scraping, Ambriz v. Google on the CIPA capability test, the Otter.ai notetaker MDL, and a widening wave of CIPA/ECPA claims against AI voicebots.

Governance translation: an agent acting on external systems is a digital worker. It needs a named owner, an approved role, access limits, an escalation path, an audit trail, and a shutdown mechanism. No evidence, no autonomy.

Privacy, biometrics and data scraping

Training-data acquisition claims increasingly plead privacy statutes instead of copyright: BIPA actions over biometric identifiers, right-of-publicity claims over voice and likeness cloning, and the consolidated Clearview AI litigation. Deepfake and AI-generated defamation claims form a parallel track covering celebrity likeness, voice cloning and political synthetic media.

State AG enforcement and AI statutes

State attorneys general have become the primary enforcement channel. Key proceedings include Texas v. Pieces Technologies (generative-AI accuracy claims in healthcare), Vermont v. Clearview AI, civil investigative demands issued in Texas to Meta and Character.AI, and xAI v. Bonta. Compliance duties now flow from the Colorado AI Act, Texas TRAIGA, California SB 243 and the New York RAISE Act.

Infrastructure, antitrust and securities

Data-center build-out is generating its own docket: Clean Air Act citizen suits over on-site gas turbines, exclusionary-zoning challenges to hyperscale campuses, and utility-regulator proceedings over load and tariffs. Alongside these sit antitrust matters (United States v. Google), trade-secret disputes among labs, securities claims over AI disclosure, and employment cases over automated hiring.