Anthropic has agreed to pay $1.5 billion to authors whose books were used without permission to train its AI, Claude, following a landmark court ruling that both punishes piracy and sets new standards for AI training. On July 20, 2026, US District Judge Araceli Martínez-Olguín approved the class-action settlement in Bartz v. Anthropic, a lawsuit initiated in 2024 by authors Andrea Bartz and Kirk Wallace Johnson. The case revealed that Anthropic relied on millions of copyrighted books sourced from unauthorized shadow libraries to build its AI.
The court made an important distinction: using pirated materials to train AI models constitutes copyright infringement, while training AI on legally acquired books falls under fair use. As a result, roughly $1.5 billion will be distributed among about 500,000 book titles, averaging $3,000 per book. This sum marks the largest copyright settlement ever recorded in the US. Despite the large payout, only a few hundred authors opted out of the class.
While the settlement might seem like a victory, many authors remain cautious. A one-time payment of $3,000 for a book that took years to write is hardly transformative. The ruling also establishes a precedent that allows AI companies to use legally purchased books for training without paying ongoing royalties, sharing revenue, or disclosing which works were included. Once a publisher sells a copy, the text can be freely consumed by AI systems, potentially generating competing content without further compensation for the author.
From a business perspective, this judgment reduces uncertainty for AI firms. After absorbing a $1.5 billion expense, Anthropic now has a clear legal framework that lowers the cost and risk of acquiring training data. This clarity is expected to speed up collaboration between AI companies and publishers, potentially opening new revenue channels for the publishing industry, even though author payments remain minimal. The ruling also fits into a broader trend of AI companies navigating data usage rights, similar to challenges highlighted by security concerns in AI models.
