AI Copyright Case Tracker: Courts Split as Suits Multiply

A Fast-Growing Docket Redraws the AI Copyright Map

Generative AI's collision with copyright law has moved well past theory. More than 70 infringement lawsuits have been filed against AI developers, alleging that copyrighted books, articles, images, music and code were scraped and used without permission to train large language models and image generators 9. Law firms and trackers, including BakerHostetler's widely cited compendium, now treat this litigation as a connected chain of liability running from content creators and publishers through dataset compilers, model developers, and finally to the businesses and consumers who deploy AI outputs 112. What began as a single cluster of author lawsuits against OpenAI, Meta and Stability AI has splintered into dozens of cases spanning books, news, visual art, music, open-source code and legal research 532.

The First Merits Rulings Split the Difference

For most of 2023 and 2024, courts mainly ruled on procedural motions to dismiss. That changed in 2025, when judges in the Northern District of California issued the first substantive, merits-stage fair-use decisions on AI training 108. In Bartz v. Anthropic, Judge William Alsup held in June 2025 that training Claude on lawfully acquired books was fair use, calling the process "exceedingly transformative" because the system learned patterns from the works rather than reproducing them as substitutes 1311. But Alsup drew a sharp line: Anthropic's practice of downloading and retaining pirated copies of as many as seven million books in a permanent digital library was not protected, regardless of whether some of those copies were later used for training 1310. That distinction exposed Anthropic to potential statutory damages that could have reached into the billions, given that willful infringement can carry penalties of up to $150,000 per work 14.

A day later, Judge Vince Chhabria reached a similar outcome for Meta in Kadrey v. Meta, granting summary judgment on the authors' training claims involving writers such as Richard Kadrey, Sarah Silverman, Ta-Nehisi Coates and Junot Díaz 1611. Yet Chhabria was careful to frame the win narrowly, stating that the ruling did not establish that Meta's conduct was lawful — only that these plaintiffs had made the wrong arguments and failed to build a record on market harm 16. He also warned that generative AI could flood markets with cheaply produced substitutes for human-created work, undermining the incentives copyright law is meant to protect 1610. Legal commentators describe the two rulings as converging on the significance of lawful data sourcing while diverging on how strictly courts should scrutinize claims of market harm 107.

Piracy Becomes the Costliest Issue

The piracy finding against Anthropic proved consequential. Facing a trial scheduled for December 2025 to determine damages, Anthropic instead reached a settlement with the author class, agreeing in principle to a resolution reported at $1.5 billion — roughly $3,000 per affected book — marking the first major settlement in the wave of generative-AI copyright litigation 141518. The court granted preliminary approval in October 2025, with final approval reported by mid-2026 after roughly 350 authors opted out to pursue individual claims 1817. Meanwhile, the piracy-based theories in the Meta case remain alive, with plaintiffs seeking Ninth Circuit review 17.

Legal Research Becomes the Clearest Loss for AI Developers

While the book-training cases produced mixed but generally favorable outcomes for AI companies, Thomson Reuters v. ROSS Intelligence went the other way. In February 2025, a Delaware federal court rejected ROSS's fair-use defense after finding it copied Westlaw headnotes to build a directly competing legal-research product 181712. Courts and analysts point to this case as the clearest instance where an AI product's status as a commercial substitute for its source material defeated the fair-use defense, in contrast to the more generalized transformation found in the Anthropic and Meta rulings 177. No federal appellate court has yet ruled on the core AI-training fair-use question, leaving ROSS's pending appeal and other cases positioned to potentially deliver the first circuit-level guidance 18179.

News Publishers Fight Over Inputs and Outputs

News organizations have pursued a distinct set of claims, often centered on the Digital Millennium Copyright Act's protections against removal of copyright-management information. The New York Times' consolidated case against Microsoft and OpenAI alleges that millions of articles were used to train Copilot and ChatGPT, which the paper says can reproduce, closely paraphrase, or misattribute its journalism 1210. In August 2026, a court dismissed the Times' contributory infringement and trademark dilution claims while allowing direct infringement, vicarious infringement, and DMCA claims to proceed 17. Related suits from the Daily News, the Center for Investigative Reporting, and The Intercept and Raw Story raise similar theories, though with mixed results 12. OpenAI failed to dismiss part of The Intercept's lawsuit, with Judge Jed Rakoff finding the outlet had plausibly alleged harm from removal of copyright-management data, even as a related claim from other publishers was previously dismissed 2019. Courts have increasingly required an "identicality" showing — that AI outputs closely match the original training text — before allowing DMCA claims to survive, creating inconsistent results across similar cases 19.

Visual Art, Music and Code Add New Dimensions

Beyond text, artists in Andersen v. Stability AI allege that billions of scraped images were used to build Stable Diffusion, Midjourney and DreamUp, with copyright claims surviving a motion to dismiss even as DMCA claims were rejected 12. Getty Images' related suit against Stability AI raises the stakes further, alleging misuse of more than 12 million photographs along with claims that outputs reproduced Getty's watermark, adding trademark and trade-dress theories to the copyright dispute 12. Music publishers, including Concord Music Group, allege that Anthropic copied song lyrics to train Claude, again pairing copyright claims with DMCA allegations 12. Programmers behind Doe v. GitHub argue that Copilot and Codex were built using open-source code in violation of licensing terms, folding contract law into the copyright fight 12. Meanwhile, in the entertainment sector, Disney v. Midjourney signals a broader shift toward litigating AI-generated outputs themselves rather than solely the training process 17.

Class Actions Concentrate Financial Stakes

Across these disputes, class-action procedure magnifies the financial exposure facing AI developers, since a single model may have ingested works belonging to enormous numbers of rights holders. The Anthropic settlement illustrates how quickly potential liability can escalate once a class is certified, and how such exposure can push companies toward settlement even before appellate courts resolve the underlying fair-use questions 1814.

A Broader Liability Landscape Beyond Training Data

Copyright is not the only frontier. Courts are also weighing contributory and vicarious liability theories for AI-generated outputs, contract claims tied to open-source licenses, and even professional-responsibility questions — including a ruling by Judge Rakoff that documents produced with commercial AI tools are not automatically shielded by attorney-client privilege 12. These threads intersect with the accelerating adoption of AI tools inside law firms themselves, where generative AI is increasingly used for legal research, document review and summarization even as firms remain cautious about confidentiality, accuracy and formal governance policies.

Where the Law Stands

Taken together, the rulings suggest an emerging, if unsettled, framework: lawfully sourced training data and non-substitutive outputs strengthen a fair-use defense, while piracy and direct market competition sharply increase liability. Commentators caution that 2026 will likely bring further pivotal rulings as more cases reach summary judgment or trial, with the fate of pending appeals — particularly in ROSS and Kadrey — poised to determine whether these district-court distinctions harden into binding appellate doctrine 91817.