Anthropicの15億ドル著作権訴訟和解が承認、離脱した著者はわずか350人Anthropic’s $1.5B copyright settlement approved; only 350 authors opted out

A federal judge has granted final approval to Anthropic's roughly $1.5 billion settlement with a group of authors who accused the company of using their books without permission to train its Claude AI models. The decision, reported by Ars Technica, closes one of the largest copyright disputes yet to emerge from the generative AI boom and offers an early template for how such claims might be resolved, particularly because only about 350 members of the affected class chose to opt out.

The case centered on how Anthropic acquired the text used to build its language models. Court filings indicated that the company obtained large volumes of books from pirated repositories, the kind of shadow libraries that circulate unauthorized digital copies. Earlier in the litigation, the presiding judge drew a notable distinction: training an AI system on books the company had legally acquired could qualify as fair use, but downloading and retaining pirated copies to assemble a training corpus did not. That ruling narrowed the dispute to the sourcing of the material rather than the act of model training itself, and it appears to have pushed both sides toward a negotiated resolution rather than a risky trial.

Under the approved terms, the settlement fund is set at approximately $1.5 billion and covers a class of roughly 500,000 works. That works out to a payment in the neighborhood of $3,000 per book, a figure that plaintiffs' attorneys have characterized as substantial for a class action of this kind. The agreement also reportedly requires Anthropic to destroy the pirated datasets at issue, addressing the underlying conduct in addition to compensating rights holders.

The low number of opt-outs is significant for interpreting the outcome. In class-action settlements, members who opt out retain the right to pursue their own individual claims, while those who remain in the class accept the settlement terms and give up the ability to sue separately over the same conduct. With only around 350 authors declining to participate out of a class numbering in the hundreds of thousands, the overwhelming majority effectively endorsed the deal. That level of participation strengthens the settlement's finality and reduces the risk of a wave of follow-on lawsuits, though it does not prevent authors outside the class or those covering different works from bringing new claims.

The settlement matters beyond the parties involved because it lands amid a broader legal reckoning over how AI systems are built. Large language models are trained on enormous quantities of text, and the question of whether ingesting copyrighted material without a license constitutes infringement or fair use remains unsettled across the industry. Several high-profile cases are still working their way through the courts, including The New York Times' lawsuit against OpenAI and Microsoft, and Getty Images' litigation against Stability AI over image generation. Authors' groups and other publishers have filed additional suits against multiple AI developers. Because those cases raise overlapping questions, an approved settlement of this size is likely to be studied closely by litigants and companies weighing whether to fight or settle.

It is worth noting the limits of what this outcome establishes. A settlement, unlike a trial verdict, does not create binding legal precedent on the core question of whether AI training on copyrighted works is lawful. The earlier fair use analysis in this case is more consequential as a legal signal, since it suggested that the method of acquiring data may be decisive. Companies that license content directly, or that rely on lawfully obtained and public-domain sources, could face a different calculus than those that used pirated material.

For the wider ecosystem, the resolution reflects a growing trend toward licensing arrangements between AI developers and content owners. A number of companies have struck deals with news organizations, publishers, and stock media providers to secure training data, and platforms have begun building tools to let creators signal whether their work can be used. Whether the Anthropic settlement accelerates that shift or simply reprices the cost of past conduct will depend on how remaining cases are decided. For now, it stands as a marker of how expensive unlicensed training data can become once the courts get involved.