- Anthropic will pay $1.5B to settle authors’ claims that it trained Claude on about 465,000 pirated books.
- The deal pays roughly $3,000 per eligible work, requires destruction of infringing datasets, and admits no liability.
- A prior 2025 ruling suggested training on lawfully obtained text may be fair use, but acquiring pirated copies remains actionable.
- Seen as the largest U.S. copyright recovery, the settlement heightens legal and financial risk for AI firms’ data-sourcing practices while leaving future-works and output claims unresolved.
Read More
The settlement between Anthropic and a large class of authors and publishers resolves claims that the company improperly downloaded and stored nearly 465,000 books from pirate sites like Library Genesis and Pirate Library Mirror to train its Claude chatbot, paying roughly $3,000 per work in compensation contingent on judicial approval. While Anthropic maintains fair use of legally acquired materials was upheld in a 2025 ruling, this deal addresses “legacy” claims arising from those pirated works.
For Anthropic, the settlement avoids what legal analysts estimated could have become multiple-billions in liabilities if it lost at trial. The company also must destroy the infringing datasets, but key practices under debate—like use of legally purchased texts or uses of model outputs—remain unchanged. However, this imposes a financial and reputational cost.
For authors, publishers, and IP rights holders, this outcome is a strong signal that unauthorized scraping or piracy for AI training is not risk-free. It demonstrates an enforceable path for compensation and that courts may support settlements that ensure remediation and transparency. The deal gives voice to lesser-known authors through class procedural safeguards and may change bargaining power in licensing.
From a strategic standpoint for the broader AI industry, this sets precedent: companies must be scrupulous about data sourcing, properly license or avoid pirated content, and may face material liabilities. Investors, partners, and AI firms will need to add legal risk premium to models, data acquisition practices, and business valuations. Regulators may treat such settlements as templates, prompting policy responses in copyright law, model transparency, and data labeling.
Open questions remain: How will “future works” be handled? The settlement excludes future copyrights and infringing outputs, but does not resolve broader issues around output liability. Also, what is the precise impact on smaller or international authors whose works may lack registration or visibility? Finally, how this affects competitors like OpenAI, Meta, etc., without similar settlements but similar risk exposures, remains to be seen.
Supporting Notes
- Anthropic will pay authors approximately $1.5 billion to settle claims that it used pirated or unlawfully acquired books to train its Claude AI model.
- Roughly 465,000 books are covered; authors will receive about $3,000 per work.
- The lawsuit was filed by authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson and represents broader rights holders.
- Judge William Alsup previously ruled that while training AI on legally obtained texts may be fair use, use of pirated books (Library Genesis, Pirate Mirror) is not protected and qualifies for liability.
- Anthropic agreed to destroy the infringing datasets as part of the settlement.
- The settlement does not include future works and does not release claims over AI-generated outputs, leaving those issues open.
- This is widely viewed as the largest copyright recovery in U.S. history.
- The class is “opt-out,” giving authors the right to refuse participation if dissatisfied.