Just Take the Cash
The Anthropic copyright settlement pays authors for access to their books and nothing more.
Copyright litigation in the United States has produced a settled price for the use of a book as training data. The sum is large, applies once, and reaches only copies made before a cutoff last August. A trained model’s conduct in the market afterward lies outside the transaction.
The verdict
Judge Araceli Martínez-Olguín granted final approval last week to the $1.5 billion settlement between Anthropic and a class of authors who accused the company of assembling its training corpus from pirated books. The award covers more than 482,000 works at roughly $3,000 each and stands as the largest copyright recovery on record. About 91 percent of those works have been claimed by their authors or publishers.
The size of the award follows from a ruling that went against the authors. Judge William Alsup held in June 2025 that training a language model on books was transformative fair use, and that the authors could show no market harm because no market existed for licensing books as training material. Anthropic’s liability survived on narrower ground. The company had downloaded more than seven million pirated books and kept them in a permanent internal library that no training run required.
The settlement follows the contours of that finding. It resolves liability for a defined set of pirated works acquired before August 2025, requires the destruction of those copies, and leaves Anthropic free to train on lawfully obtained editions of the same titles. It creates no forward-looking license, imposes no rule on what the models produce, and binds no other company.
Another piracy scare
Roughly 125 copyright cases against AI developers remain active in American and foreign courts, and the surviving theory in most of them concerns acquisition. Once two federal judges in San Francisco found the training itself defensible, plaintiffs redirected their claims toward how the works were obtained, concentrating on shadow libraries such as LibGen and Anna’s Archive, the sources from which those corpora were assembled.
Authors who opted out of the Anthropic class sued six developers at once last December, Anthropic among them, arguing that any company downloading a dataset over a torrent network simultaneously distributes it to every other peer in the swarm. Meta has responded in its own case that seeding is an unavoidable byproduct of the protocol and belongs inside the fair use finding that it has already won.
Each of these claims measures the same two quantities, the number of copies made and the method by which they were obtained. A judgment therefore scales with facts that were fixed long before any case reaches trial. Nothing in that calculation attaches to what a model earns after it is deployed.
The disappearing eyeballs
Earlier this month the French competition regulator published an assessment of the layer where those earnings are collected. The Autorité de la concurrence’s opinion on the AI agents sector is its third study of the AI value chain and the first to look downstream of model training, at the point where agents reason, plan, and carry out tasks on a user’s behalf. OpenAI, Google, and Anthropic together account for more than 84 percent of the market that it describes.
An agent that fully answers a question removes the reason to visit the source. Traffic passed from an agent to an e-commerce site currently runs below 5 percent, a figure that the Autorité expects to reach 20 to 25 percent by 2030 as agents become a primary channel for reaching consumers. Sites lose the behavioral data attached to any purchase completed inside an agent’s interface. The developer of the agent also sets the conditions under which a site’s content appears in a response at all, including how many sources a user sees and how those sources are ranked.