Anthropicの15億ドル著作権訴訟和解が承認されるAnthropic’s landmark $1.5B copyright settlement is approved

A federal court has granted final approval to Anthropic's roughly $1.5 billion settlement with a group of authors and publishers who accused the company of using their copyrighted books without permission to train its AI models. The agreement ranks among the largest ever reached over the data used to build generative AI systems, and it offers one of the first concrete price signals for how much unauthorized training material may ultimately cost the companies that rely on it.

The case centered on Anthropic, the developer of the Claude family of AI assistants, and claims that it obtained large volumes of books from so-called shadow libraries, online repositories that host pirated copies of copyrighted works. According to figures cited during the litigation, the settlement is expected to cover roughly 500,000 works, which translates to a payment of approximately $3,000 per book. The deal also reportedly requires the destruction of the pirated datasets in question, though the payments are structured to compensate rights holders whose works were included.

The settlement follows a closely watched ruling by the presiding judge that split the legal questions into two parts. Training an AI model on books the company had lawfully acquired was found to be transformative and therefore likely to qualify as fair use, a doctrine in U.S. copyright law that permits certain uses of protected material without a license. Downloading and storing pirated copies, however, was treated as a separate and potentially infringing act. That distinction is significant because it suggests courts may scrutinize how training data is acquired as much as how it is ultimately used, an interpretation that could shape industry practices even where model training itself is defended as fair use.

The court's approval process was not entirely smooth. The judge initially raised concerns about aspects of the proposed settlement, including how class members would be notified and how claims would be validated, before signing off on the final terms. Such scrutiny is common in class action settlements of this scale, where the fairness of the distribution to affected parties is a central consideration.

For context, Anthropic is one of several leading AI developers facing litigation over training data. OpenAI, Microsoft, Meta, Google, and Stability AI have all been named in lawsuits brought by authors, artists, news organizations, and other rights holders. The New York Times, for example, has an ongoing case against OpenAI and Microsoft, and various groups of authors have filed suits alleging their books were used without consent. Because many of these cases hinge on similar questions about fair use and the provenance of training data, the resolution of the Anthropic matter is likely to be referenced by parties on both sides of those disputes.

It is worth noting what a settlement does and does not establish. Because the parties agreed to resolve the case rather than litigate it to a final verdict, the outcome does not create binding legal precedent in the way a court judgment would. The earlier fair use ruling remains the more legally consequential element, while the settlement primarily resolves the piracy-related exposure. Even so, the dollar figure provides a reference point that other plaintiffs and defendants may weigh when assessing the risks and costs of similar claims.

The broader industry backdrop helps explain why the case drew such attention. Large language models are trained on enormous corpora of text, and the exact composition of those datasets has often been opaque. Datasets assembled from web scrapes and book collections have circulated widely among researchers, and some were later found to contain pirated material. As legal and regulatory pressure has grown, several AI companies have begun signing licensing agreements with publishers, news outlets, and content platforms, signaling a shift toward paid data arrangements rather than reliance on freely scraped material.

The approved settlement appears to reinforce that trend by attaching a tangible cost to the use of unauthorized works. Whether it prompts a wider move toward licensing, changes how companies document their data sources, or simply raises the stakes in ongoing litigation will likely become clearer as the remaining cases against other developers proceed. For now, the outcome stands as a notable marker in the still-evolving relationship between generative AI and copyright law.