Anthropic pays $1.5B to settle book piracy claims—AI training on legal content upheld as fair use

In a landmark settlement approved by a federal court in San Francisco, Anthropic agreed to pay book authors $1.5 billion（約2400億円） to resolve a copyright lawsuit stemming from its download of works from the piracy databases LibGen and PiLiMi. Between 2021 and 2022, Anthropic obtained roughly 482,460 listed works from these unlicensed sources. Of that collection, 91.3 percent were claimed by authors in the settlement class, resulting in a payout of about $3,000 per work—four times the statutory minimum. The settlement is the largest copyright settlement in class action history. As part of the agreement, Anthropic must destroy the pirated files and faces ongoing liability: authors retain claims over AI outputs that reproduce their original works and over Anthropic's future conduct.

The settlement, however, tells only part of the legal story. Judge Alsup had previously issued a ruling that stands as a shield for AI training practice more broadly. He determined that training AI on legally obtained books is "transformative—spectacularly so" and constitutes fair use under copyright law. This ruling means that the act of training an AI model on lawful data does not, in itself, infringe copyright. The settlement thus addresses the piracy—the unlicensed acquisition of the books—not the use of those books in AI training. This distinction is significant because it leaves unresolved a broader legal question: whether the mass scraping of internet content without owners' consent, the primary source of training data for most AI labs, counts as lawful acquisition in the first place. The ruling effectively hands AI labs a major victory on the fair use question while leaving the acquisition question open. Whether the scraping itself is legal remains unsettled, and the article suggests that "the fair use debate is likely far from over," implying that future litigation could test the boundaries further and potentially reshape the legal terrain for data acquisition practices.

The settlement represents a watershed moment framed very differently depending on where you stand. For authors and publishers, it is vindication: $1.5 billion（約2400億円） is the largest copyright settlement in class action history, and the per-work payout of roughly $3,000 is four times the statutory minimum, signaling judicial seriousness about author harm. For Anthropic and other AI labs, however, the ruling contains a protective aperture. Judge Alsup's prior determination that AI training on legally obtained books is "transformative—spectacularly so" and constitutes fair use creates a legal scaffold that insulates many training practices. The settlement itself does not target training; it targets piracy—the act of downloading unlicensed works. This distinction is material. If training on lawful data is fair use, then the bottleneck for AI labs shifts from the training step itself to the question of whether their data acquisition was lawful. That question, the article makes clear, remains open. Most AI labs have sourced training data by scraping the web and internet archives without owner consent. The ruling does not resolve whether that scraping is legal; it only clarifies that once the data is lawfully in hand, using it for AI training is likely permissible. The fair use debate is "likely far from over," the article notes, suggesting that litigation over the legality of the scraping itself could emerge separately and test the boundaries further.