Anthropic built its Claude AI using, among other things, millions of books downloaded from piracy websites. Not a grey area, not an accident. A court found the piracy was, in its words, inherently and irredeemably illegal. That finding scared Anthropic enough to settle before trial.
The result was a $1.5 billion agreement covering roughly 480,000 pirated works. At face value, it sounds significant. In practice, each book gets about $3,000. The lawyers originally asked for $300 million of that fund for themselves, then reduced the request to $187.5 million after judicial pressure. After legal fees, administrative costs, and other expenses, authors and publishers are left splitting roughly $1.29 billion, which works out to approximately $2,900 per title.
For context, US copyright law allows for up to $150,000 per pirated work in cases of willful infringement. Anthropic was staring at theoretical liability north of $70 billion if it had gone to trial and lost badly. The $1.5 billion settlement, painful as it looks in press releases, is a fraction of worst-case exposure. Authors got 2 cents on the dollar of the maximum possible. Anthropic got certainty, protection from future claims tied to past behavior, and a clean break.
The settlement's cracks are now visible in court. On May 14, Judge Araceli Martinez-Olguin declined to give final approval, pressing lawyers on fee structures and payments to the named plaintiffs who kicked off the case. She specifically wanted to understand why objectors were not being given proper room to raise concerns. Fifty-three class members formally objected, and the objections are substantive: some authors discovered that a single copyright registration covering dozens of separately published novels was being counted as one claimable work. Others found that books published under pseudonyms were excluded. International authors without US copyright registration got nothing at all.
Meanwhile, 28 authors including novelist Dave Eggers opted out entirely and filed their own lawsuit against Anthropic in California. They are seeking $150,000 per work per defendant, the full statutory ceiling. Their argument is simple: $3,000 is 2% of what the law actually permits, and they would rather take their chances in court.
This is also not an isolated case. There are now over 70 copyright infringement lawsuits against AI companies in the United States alone. OpenAI, Meta, Google, and others are all defendants in various stages. The music industry has already settled similar cases with AI music generators, with Warner Music Group and Universal Music Group both reaching licensing agreements that include opt-in models giving artists control. The book publishing world is watching to see if Anthropic's deal sets the floor, or whether authors who fight individually can push it higher.
The broader legal picture is genuinely unsettled. In June 2025, one judge ruled that AI training on legally obtained books is fine under copyright law. A few weeks later, a different judge looking at Meta's case reached a different conclusion on a related point. Courts sitting in the same city are disagreeing on the same core question. No appeals court has stepped in yet.
For anyone whose business creates, publishes, licenses, or depends on written content, the practical takeaway is this: the era of AI companies freely training on whatever data they could find is ending. The cost is now measurable in billions. But the final pricing for that access, and who ultimately pays it, is still being argued out in courtrooms.