Overview
Anthropic, a leading developer of artificial intelligence large language models (LLMs), has reached a settlement in a closely watched dispute with a class of authors and publishers who alleged copyright infringement arising from Anthropic's use of copyrighted materials to train its generative AI models. The settlement, if approved, will be the largest publicly reported copyright recovery in history, and could set a new benchmark for other AI-related litigations or licensing disputes. This alert summarizes the key terms of the proposed settlement and analyzes its broader implications for AI developers, enterprise users, and rights holders in the AI ecosystem.
Background
The settlement stems from Bartz et al. v. Anthropic PBC, a case brought in the Northern District of California. Plaintiffs alleged that Anthropic used copyrighted content from their books to train Anthropic's Claude LLMs without proper authorization. The copyrighted content included books that Anthropic purchased, scanned to create digitized versions, and then ultimately destroyed. In addition, Anthropic allegedly used over seven million digital copies of books acquired from pirating sites such as Library Genesis ("LibGen") and Pirate Library Mirror ("PiLiMi") to train its Claude LLMs. Plaintiffs claimed that Anthropic's practices violated copyright law and sought damages as well as injunctive relief. Anthropic maintained that its conduct fell within the bounds of fair use and was essential for the development of competitive AI technologies.
In June 2025, Judge William Alsup issued a summary judgment which held that Anthropic's use of lawfully purchased books for destructive digitization and for training LLMs was "among the most transformative we will see in our lifetimes" and qualified as fair use, while rejecting the fair use defense for any use of pirated works, finding piracy of copyrighted works to be "inherently, irredeemably infringing," regardless of whether those works were ultimately used for LLM training. The court therefore denied summary judgment for Anthropic as to the pirated works, setting the stage for a high-stakes trial. For a detailed analysis of that summary judgment decision and its implications for stakeholders in the AI ecosystem, see our June 25, 2025 Alert.
Settlement Terms
With the plaintiffs seeking statutory damages—which can range from $750 to $30,000 per infringed work, and higher in cases of willful infringement—Anthropic found itself facing potentially substantial damages as high as in the tens of billions of dollars, and the parties announced in early August 2025 that they had reached a settlement in principle. Publicly available information regarding the settlement, disclosed in a motion filed with the court on September 5, includes the following key provisions:
- Monetary Compensation: Anthropic will pay a minimum of $1.5 billion for its past use of pirated books representing $3,000 for each of the approximately 500,000 copyrighted works that were alleged as having been downloaded from LibGen or PiLiMi and infringed by Anthropic. The parties are still compiling the full list of copyrighted works, and if the list exceeds 500,000, Anthropic will pay an additional $3,000 for each additional work.
- Past Release of Liability Only: Notably, the settlement does not insulate Anthropic from future claims. The plaintiffs have released claims only for Anthropic's past acquisition, retention, and use of the identified works for AI training and related internal research and development activities occurring before August 25, 2025. The settlement does not cover any future conduct or other works owned by class members that are not identified in the final list of works.
- No Release of Claims Based on Infringing LLM Output: Although the claims at issue here related solely to the materials used to train Claude (i.e., the "input" to the LLM), a common issue in many AI-related copyright actions is whether the LLM produces output that infringes copyrighted works. The parties' settlement here does not cover any claims for allegedly infringing outputs from Anthropic's AI models. This means that if Anthropic's models generated infringing outputs, those claims remain actionable, even for outputs that infringe the same identified works.
- Destruction of Materials: In addition to the monetary compensation, Anthropic has agreed to destroy the two libraries that allegedly contain the pirated works, as well as any derivative copies originating from those sources, within 30 days of final judgment or the expiration of any litigation preservation obligations. Anthropic must certify in writing to class counsel that the destruction has been completed and that the allegedly infringing materials are permanently removed from its systems.
Implications for Stakeholders in the AI Ecosystem
This landmark settlement underscores the importance of robust compliance strategies for AI developers and users. The per-work settlement amount of $3,000 is four times larger than the $750 floor for statutory damages, and 15 times larger than what the plaintiffs would have obtained had Anthropic prevailed at trial on its defense of innocent infringement and convinced a court to exercise its discretion to revise down the damages to $200 per work.
Implications for Other Pending Cases Against AI Developers
The Anthropic settlement is likely to serve as a benchmark in other pending copyright actions against AI companies. Plaintiffs in similar cases may point to the $3,000 per-work figure as a starting point for damages negotiations, and courts may look to the structure of this settlement when fashioning remedies. While the settlement itself is not binding precedent, AI companies facing similar claims and risking substantial liability should expect plaintiffs to follow the model set by this settlement and seek to require identification, segregation, and destruction of infringing data sources as part of settlement negotiations. This requirement may also extend to derivative datasets and any internal copies, raising the stakes for data governance and compliance.
Implications for Licensing Negotiations
The summary judgment decision and the settlement also underscore the importance of proactive licensing engagements, and the settlement may accelerate the development of licensing frameworks for AI training data, as developers seek to avoid litigation risk and secure access to high-quality, authorized content. While the law remains unsettled, and fair use outcomes may differ for generative versus non-generative AI models, AI companies face significant copyright-related risk—particularly where pirated works have been used for training LLMs. However, AI companies that engage in proactive licensing engagements may have significant leverage in setting licensing rates that are based on reasonable market royalties rather than statutory damages guidelines, particularly in instances where the AI developer has not engaged in piracy with respect to those works. Enterprises using third-party AI tools may also seek stronger indemnification with respect to the provenance of training data to mitigate risk of potentially infringing outputs.
Strategic Considerations for AI Developers and Enterprise Users
AI developers and enterprise users should prioritize rigorous data governance by maintaining detailed records of data provenance and ensuring that all training data is lawfully acquired. The Anthropic settlement demonstrates that the risks associated with using pirated or unauthorized materials can be substantial.
Companies should also be prepared to act swiftly if infringing datasets are identified, including performing prompt destruction of such data and providing formal certification of compliance if required. Proactive remediation is essential to minimize liability and demonstrate good faith in the event of a dispute.
Given that the risk of copyright claims based on AI-generated outputs remains unresolved, developers and users should implement ongoing monitoring of model outputs for potential infringement, including use of technical safeguards to prevent the regurgitation or close reproduction of protected works.
Finally, enterprise users should seek robust contractual protections and other protections, including clear representations and warranties regarding lawful data sourcing and indemnification provisions for copyright claims to help manage legal exposure and ensure compliance in a rapidly evolving regulatory environment.
Conclusion
Anthropic's settlement marks a significant milestone in the intersection of AI innovation and intellectual property law. Ropes & Gray will continue to monitor this space and provide updates. Clients with questions about AI compliance or risk management are encouraged to contact their Ropes & Gray relationship attorney for tailored guidance.
The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.
