A federal judge has given final approval to a $1.5 billion class action settlement resolving claims against Anthropic over its use of pirated books to train the Claude AI model. The case, brought by authors including Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson, centered on whether using copyrighted books to train AI systems violates copyright law. [1] [6]
The settlement resolves one of the largest copyright disputes in the AI industry, involving roughly 500,000 works allegedly obtained from pirate websites. While the financial terms are significant, the more lasting impact may be the legal framework the case established for how AI companies acquire and use training data.
The Court’s Fair Use Ruling
Judge William Alsup issued a summary judgment in June 2025 that drew a sharp distinction between the legality of AI training itself and the method of data acquisition. He ruled that using lawfully obtained books to train AI models constitutes transformative fair use, comparing the process to how an aspiring writer reads to create something new. However, he explicitly found that downloading and permanently storing books from pirate websites constitutes copyright infringement.
This ruling established what legal observers describe as a “doctrine of piracy as original sin” meaning that even if the use of the content is transformative, an illegal source cannot be legitimized. The court emphasized that Anthropic had no entitlement to use pirated copies for its central library, noting that creating a permanent, general-purpose library was not itself a fair use excusing the underlying piracy.
The Settlement Terms
Under the proposed agreement, Anthropic agreed to pay approximately $3,000 per work to the rights holders, with the total payout reaching $1.5 billion. The company must also destroy all copies of works obtained from pirate libraries and any derivative copies within thirty days of the final judgment.
Crucially, the settlement only resolves existing claims related to training data up to a specific cutoff date. It does not address potential infringement issues concerning future AI model outputs, leaving that question open for subsequent litigation. The company receives only a limited release covering past training behavior and gains no license for future training, meaning it remains exposed to potential lawsuits over AI-generated content.
How Anthropic Acquired Training Data
According to court filings, Anthropic employed two primary methods to obtain training data. The first involved bulk downloading e-books from pirate websites such as LibGen and PiLiMi, storing them long-term in a “central library.” The investigation found that Anthropic downloaded over 7 million pirated books from these sources.
The second method involved purchasing second-hand physical books and converting them into digital text through destructive scanning, which included cutting the spines and destroying the physical copies. Approximately 2 million physical books were destroyed in this process. While the court ruled that converting legally purchased books into electronic versions for training AI models constituted fair use, the pirated downloads were deemed infringement.
Implications for the AI Industry
The settlement offers a practical framework for how AI companies must approach training-data acquisition. With similar copyright lawsuits pending nationwide against AI developers, including cases involving Disney Enterprises and Meta Platforms, the Anthropic case plots a potential roadmap for resolution.
The key takeaway for AI developers is that how data is acquired matters as much as what is done with it, if not more. Companies can legally train models on copyrighted materials under fair use doctrine, but they cannot rely on pirated sources to build their datasets. This distinction is likely to influence how AI companies source their training data going forward, pushing them toward licensed content and legitimate acquisition channels rather than shadow libraries and pirate sites.
The case also highlights the ongoing tension between AI innovation and copyright law, a balance that courts and legislators will continue to grapple with as generative AI technology advances. For rights holders, the settlement demonstrates that collective action can yield substantial recovery when copyright is violated, even in the rapidly evolving landscape of artificial intelligence.
Sources
- Anthropic to pay authors $1.5 billion to settle lawsuit over pirated …
- $1.5B Anthropic settlement resolves AI training lawsuit
- Anthropic’s AI Training Settlement Offers Tech Sector an IP Map
- Bartz v. Anthropic PBC_Baiduwiki - 百度百科
- Bartz v. Anthropic | BakerHostetler
- Bartz v. Anthropic Settlement: What Authors Need to Know
