On July 20, 2026, the Federal District Court for the Northern District of California gave final approval to a settlement in the case of Bartz v. Anthropic (developer of the Claude artificial intelligence model), under the terms of which Anthropic must pay 1.5 billion US dollars to authors and publishers.
This sum is being called the largest known payout in a copyright infringement case in US history. It covers more than 480,000 books, and the compensation will amount to approximately 3,000 dollars per work before deduction of the expenses provided for in the settlement agreement.
However, the significance of this case is not limited to the record payout amount. Its main legal outcome is that an American court drew a distinction between two fundamentally different questions: the use of works to train artificial intelligence, and the legality of the method by which the developer obtained those works.
Beginning of the Case
In August 2024, American writers Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson filed a lawsuit against Anthropic in the Federal District Court for the Northern District of California.
The plaintiffs alleged that the company, without the rights holders’ permission, had downloaded millions of books, including from the pirate online libraries LibGen and PiLiMi, compiled them into its own digital library, and used the works to develop and train the Claude large language models.
The court subsequently found that the case met the requirements to proceed as a class action, under which nearly 120,000 authors and other rights holders filed claims for compensation, with the case as a whole covering 482,460 books.
Anthropic, for its part, insisted that training artificial intelligence on texts is a transformative use and falls under the Fair Use doctrine — the US legal doctrine permitting the good-faith use of works without the rights holder’s prior permission.
Thus, at the center of the dispute were two interrelated but legally distinct questions: whether training a model on copyrighted books constitutes infringement, and whether a developer can invoke Fair Use if the copies of the books themselves were obtained unlawfully.
Proving Anthropic’s Use of the Texts
During the discovery stage — the procedure under American procedural law for the disclosure of evidence — the parties gained access to documents and information in each other’s possession.
To support their claims, the plaintiffs relied on Anthropic’s internal documentation, employees’ email correspondence, technical materials, and sworn testimony given by company employees.
This evidence showed that Anthropic had deliberately sought out large collections of books to build training datasets. Internal discussions referenced the resources LibGen and PiLiMi, and the company had created a so-called “central library” — a centralized repository of e-books accessible to its engineers.
In total, Anthropic downloaded more than seven million books from pirate sources. Some of these materials were used to train the models, while other copies were stored in the library for possible future use.
At the same time, Anthropic did not deny the fact of downloading substantial volumes of books. Therefore, the dispute gradually shifted from the question of whether the company had used the relevant works to the question of the legal characterization of its actions.
The Court’s Summary Judgment Ruling of 2025
In June 2025, Judge William Alsup issued a ruling on summary judgment — a simplified resolution of the case without a full trial.
The court gave a separate legal assessment to three categories of Anthropic’s conduct:
- The court found that using books to train large language models serves a fundamentally different purpose than their ordinary use by readers, and is therefore transformative. In the court’s view, the AI model does not reproduce the books or create substitutes for reading them. Instead, it analyzes texts to identify linguistic patterns and learn to generate new content. For this reason, the court concluded that the process of training a model may, under certain conditions, fall under the Fair Use doctrine of good-faith use;
- The court found it permissible to purchase print books and scan them in order to replace physical copies with digital copies in the company’s internal library. In this case, Anthropic had lawfully acquired the relevant copies, and the digitization did not increase the number of copies it possessed.
- The court declined to extend Fair Use to the downloading of millions of books from pirate sources. The creation of a permanent digital library from unlawfully obtained copies was a separate act that could not automatically be justified by the fact that some of the materials might subsequently be used to train artificial intelligence.
Thus, Anthropic secured an important victory regarding the training of the models themselves, but not regarding the manner in which a significant portion of the books had been obtained.
Further Proceedings and Conclusion of the Settlement
The 2025 ruling did not determine the final extent of Anthropic’s liability and did not conclude the proceedings concerning the pirate library.
The next stage was to be a full trial. During it, it would have been necessary to establish which specific works had been unlawfully downloaded, who held the rights to them, whether the infringement had been willful, and what compensation Anthropic should pay.
Of particular importance was the question of willful copyright infringement.
Under US law, a rights holder may, under certain conditions, seek not only compensation for actual damages but also statutory damages provided for by law. Where willful infringement is established, the amount can reach 150,000 US dollars per work.
Given the scale of Anthropic’s library, even compensation calculated for only a portion of the downloaded books could have resulted in liability amounting to tens of billions of dollars. This is precisely why the planned trial posed an extremely serious financial risk for the company. For Anthropic, therefore, the settlement made it possible to cap its potential liability and avoid litigation over the willfulness of the infringement as well as years of appeals.
However, continuing the dispute also carried risks for the plaintiffs. Authors and publishers would have had to prove ownership of the rights to each work, the fact that a specific book had been included in the pirate library, compliance with the requirements for obtaining compensation, the willful nature of Anthropic’s conduct, and the amount of compensation due. A separate difficult issue was determining the scope of class action participants and those entitled to compensation — in particular, authors and publishers, the delineation of their respective rights, and the distribution of any recovery among rights holders. Even in the event of victory, the final receipt of funds could have been delayed for years due to appellate review.
For this reason, the settlement replaced these uncertain prospects with a guaranteed fund of 1.5 billion US dollars. For Anthropic, it established a predictable limit on financial risk, while for rights holders it secured real compensation without the need to continue a complex and costly judicial proceeding.
In September 2025, the court granted preliminary approval of the proposed settlement and authorized the process of notifying class members and accepting compensation claims to begin. Because this was a class action, an agreement between the class representatives and Anthropic was not sufficient on its own. The court had to determine whether the terms of the settlement were fair, reasonable, and adequate for all persons whose rights it affected.
In May 2026, a separate hearing was held on final approval of the settlement. The court considered the number of claims filed, objections from class members, the terms for distributing compensation, the question of excluding certain individuals from the group of claimants, and the amount of attorneys’ fees.
Following Judge William Alsup’s retirement, the case was taken over by Judge Araceli Martínez-Olguín. On July 20, 2026, she gave final approval to the settlement and concluded the class action proceedings.
At that point, authors and publishers had asserted rights to more than 91 percent of the over 482,000 books covered by the settlement. At the same time, the court reduced the attorneys’ fees requested by plaintiffs’ counsel so that a larger share of the compensation fund would remain available to class members.
After the settlement was approved, Anthropic Deputy General Counsel Aparna Sridhar noted that the court’s ruling confirmed that training AI models on books can fall within the Fair Use doctrine, and also stated that more than 91% of the authors and publishers covered by the settlement had already received the payments owed to them.
The Legal Significance of Bartz v. Anthropic case
The settlement does not mean that Anthropic was finally found liable by the court for all the alleged infringements. The settlement provides for the termination of the dispute without a full trial of the questions of willfulness and the final amount of damages. At the same time, it does not overturn the legal conclusion the court reached in 2025 regarding Fair Use.
The main significance of Bartz v. Anthropic lies in its distinction between two aspects of AI developers’ conduct.
The use of lawfully obtained works to train a model may, depending on the specific circumstances, be found to constitute transformative good-faith use.
However, the creation of training datasets or internal libraries from pirated copies is a separate act. The unlawfulness of the origin of such copies does not disappear simply because they are subsequently used for a technologically new or transformative purpose.
Therefore, the main practical question for AI developers today is not only what protected materials are used for, but also where they were obtained, under what conditions they are stored, and whether the company can document the lawfulness of their origin.
In preparing this article, materials from the following sources were used:
Associated Press, Reuters, Anthropic Copyright Settlement Website