$1.5 billion: judge approves the largest copyright settlement of the AI era with Anthropic

By Zendoric · July 21, 2026.

Federal judge Araceli Martinez-Olguin of San Francisco gave final approval on July 20 to the settlement between Anthropic and a group of authors who sued the company for using their works to train Claude, Reuters reported. The figure, $1.5 billion, makes it one of the largest known settlements in U.S. copyright history, and the first of its kind against an AI company to be resolved through a settlement rather than a trial.

The origin of the case lies in a legal distinction that was already outlined in 2025 and is now sealed through the settlement. The judge who presided over the case at the time, William Alsup, ruled that training a model on copyrighted books can fall under the doctrine of "fair use" (the exception that allows protected works to be used without permission in certain contexts, such as research or substantial transformation of the original). But the same judge found that Anthropic did violate rights by storing more than 7 million books obtained from pirate sources in a central library that, according to the ruling, was not necessarily tied to model training. That distinction—training can be legal, accumulating pirated copies is not—is what has structured the negotiation and will likely be cited in every copyright lawsuit against an AI company that follows.

According to one of the authors' attorneys at the hearing, the claims cover more than 92% of the over 480,000 works included in the settlement. It is not unanimous: several authors have expressed disagreement, arguing that the amount is insufficient, that the distribution falls disproportionately on the plaintiffs, and that it leaves out certain rights holders. A group of authors and publishers has opted out and continues to pursue their own lawsuits against Anthropic, still ongoing.

Our take is that this settlement functions as the first real market price for the unauthorized use of protected content in AI training, and that price will anchor future negotiations across the sector, from the lawsuits OpenAI and Meta currently face to image and video litigation like the one Midjourney is fighting over the use of material from studios such as Disney or Warner Bros. Broadly speaking, the industry had been operating under the premise that the legal risk of training on data of dubious origin was abstract and could be deferred; $1.5 billion turns it into a concrete, quantifiable cost item—and above all, a predictable one for anyone considering raising a round or going public with that exposure on the balance sheet.

In the short term, this doesn't resolve the underlying tension between creators and AI companies: many authors will still feel the payout is insufficient given the value their works have contributed to models generating billions in revenue, and parallel lawsuits will keep being filed. It's an uncomfortable transition, and it would be dishonest not to acknowledge that. But in the medium term, precedents like this push toward something healthier than a legal vacuum: data licensing markets with pricing, traceability and compensation, instead of silent extraction. That is precisely the trust infrastructure needed for AI's long-term promise—models trained sustainably, that unleash creative and intellectual capacity rather than eroding it—to be built on foundations that authors, publishers and companies can accept without feeling the game is rigged from the start.