AI vs. Authors Update: Court Approves Historic Anthropic Settlement While Meta Litigation Continues

Key Takeaways

• Anthropic Settlement Receives Final Approval: A federal judge approved the $1.5 billion settlement resolving authors’ claims that Anthropic unlawfully downloaded and retained hundreds of thousands of copyrighted books obtained from pirate libraries.
• Settlement Does Not Resolve the Broader AI-Training Debate: The settlement addresses past copying of specified works, but it does not release claims concerning AI-generated outputs, future conduct or works omitted from the settlement’s works list.
• The Authors’ Case Against Meta Remains Active: Although Meta previously prevailed on the plaintiffs’ fair use theory, the court has allowed the authors to pursue an additional claim relating to Meta’s alleged uploading of copyrighted works through BitTorrent, and the parties continue to litigate discovery issues.
• Acquisition Practices Remain a Significant Risk: For AI developers, the source and handling of training data may present legal exposure independent of whether the ultimate use of that data to train a model is transformative.

In July 2025, we discussed two decisions from the United States District Court for the Northern District of California addressing the use of copyrighted books to train large language models. In Bartz v. Anthropic PBC and Kadrey v. Meta Platforms, Inc., the courts treated the use of books for model training as potentially transformative but expressed serious concerns about the acquisition and retention of books obtained from pirate or “shadow” libraries.

One year later, the cases have followed substantially different paths. The court in Bartz has entered final judgment approving a historic settlement between Anthropic and a class of authors and copyright owners. The Kadrey case, by contrast, remains active and is proceeding through further pleading and discovery.

The Court Approves the $1.5 Billion Bartz Settlement

On July 20, 2026, Judge Araceli Martínez-Olguín granted final approval of the Bartz class action settlement, entering judgment and dismissing the case with prejudice. The settlement established a non-reversionary fund of $1.5 billion for owners of approximately 482,000 eligible works included on a court-approved Works List. The court described it as the largest copyright class action settlement in history.

The settlement followed an earlier ruling that separated Anthropic’s conduct into distinct categories. The court concluded that using lawfully acquired books to train Anthropic’s large language models was fair use because the training process was transformative. It also found that converting purchased print books into digital form for use in a central research library was fair use. But the court rejected Anthropic’s fair use defense for downloading millions of books from LibGen and PiLiMi, two pirate libraries, and retaining copies in a general-purpose central library.

The settlement resolves claims involving specified books downloaded from those libraries. The court found the estimated payment of approximately $3,000 per eligible work — before fees, expenses and allocation among multiple rightsholders — to be meaningful in view of the risks, costs and potential delay of continued litigation. As of April 2026, claims had been submitted for 91.3% of the works on the Works List, while only 350 timely opt-outs covering 1,802 works had been received.

The court overruled objections arguing, among other things, that the settlement amount was inadequate, the notice program was insufficient and the class should include additional works. Of particular importance, the court explained that works not appearing on the Works List are not covered by the settlement. Claims concerning those works therefore remain available rather than being released.

The Bartz Settlement’s Release Is Limited

The settlement does not provide Anthropic with a blanket release for all copyright claims arising from its artificial intelligence systems.

The release applies principally to past claims involving Anthropic’s acquisition and copying of the identified works through the point before any AI output. It does not release claims based on allegedly infringing model outputs, nor does it release claims concerning future conduct occurring on or after August 25, 2025. The court viewed that comparatively narrow release as a benefit to the class because authors retain the ability to challenge future uses or particular outputs.

Anthropic must also destroy original files downloaded from LibGen or PiLiMi and copies originating from those files, subject to specified preservation obligations. Anthropic represented that those datasets were not included in the training corpus of any commercially released large language model.

The court awarded class counsel approximately $101.6 million in attorneys’ fees — substantially less than the requested $187.5 million — plus approximately $2.6 million in incurred expenses and an $18.2 million reserve for future settlement-administration expenses. Ten percent of the fee award will be withheld pending a post-distribution accounting.

The Kadrey Case Remains Active

The authors’ lawsuit against Meta has not reached a similar resolution. In June 2025, Judge Vince Chhabria granted summary judgment to Meta on the plaintiffs’ theory that using copyrighted books to train Meta’s Llama models harmed the market for their works. The court emphasized, however, that its decision resulted from the evidentiary record and arguments presented rather than a categorical conclusion that AI training is always fair use.

The case has since continued on a different theory. In March 2026, the court permitted the plaintiffs to amend their complaint to assert contributory-infringement allegations based on Meta’s use of BitTorrent. The plaintiffs allege that, while downloading books from shadow libraries, Meta simultaneously made portions of those works available to other BitTorrent users, thereby contributing to third-party infringement. The court allowed that theory to proceed while denying the plaintiffs’ request for broader class discovery at that stage.

Discovery concerning Meta’s torrenting practices has continued. The parties have disputed whether internal communications about the legality of torrenting and uploading copyrighted material are protected by the attorney-client privilege. Following in-camera review, the magistrate judge sustained Meta’s privilege claims and later rejected an attempt to invoke the crime-fraud exception for certain materials.

Thus, although the original fair use claim was resolved in Meta’s favor, Kadrey remains ongoing. The remaining allegations may focus less on whether model training itself is transformative and more on the methods used to obtain and distribute copyrighted training materials.

Practical Implications for AI Developers and Copyright Owners

Together, the cases reinforce that “AI training” should not be treated as a single act for copyright-risk purposes. Courts may separately analyze the acquisition of source material, creation of a permanent digital library, ingestion of data for training, operation of peer-to-peer downloading technology and generation of model outputs.

As we discussed in our prior update, for AI developers, a persuasive fair use position concerning model training does not necessarily eliminate risks arising from how the training material was acquired or retained. Companies should document the provenance of training data, distinguish materials licensed or lawfully purchased from materials obtained through scraping or file-sharing networks, and assess whether copies are retained after their intended training purpose has ended. Controls governing BitTorrent and similar technologies should address not only downloading but also any automatic uploading or redistribution.

Developers should also avoid assuming that one favorable training ruling resolves output-related exposure. The Bartz settlement expressly preserves output claims, and future cases may examine whether particular systems reproduce protected expression, generate substantially similar material or interfere with developing licensing markets.

For authors, publishers, and other copyright owners, accurate registration and ownership records remain important. The Bartz class definition depended on copyright-registration timing and work-identification information such as ISBNs or ASINs. Rightsholders should also consider monitoring known training datasets, preserving contracts allocating digital and AI-related rights, and separately evaluating claims involving acquisition, training, outputs and future use.

The final approval of the Bartz settlement closes a significant chapter in the litigation over pirate-library books. It does not, however, establish a universal price for AI training or definitively answer whether copyrighted works may be used without permission. With Kadrey continuing and other lawsuits proceeding across the country, the distinction between transformative use and unlawful acquisition remains one of the most consequential — and unsettled — issues in AI copyright law.

For more information, contact Jay Heidrick, Melenie Van or your preferred Polsinelli attorney.