On June 24, 2025, U.S. District Judge William Alsup issued a historic split-decision ruling in the Bartz et al. v. Anthropic PBC case. Anthropic, an American artificial intelligence (AI) safety and research company, was accused of using unauthorized copies of books to train its large language model (LLM), Claude. The plaintiff side consisted of authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson, whose books were among those used by Anthropic.

On one hand, Alsup ultimately ruled that training LLMs on legally purchased works constituted fair use, setting a revolutionary legal boundary in the realm of AI. However, he denied summary judgment on the issue of the seven million pirated books, a case which will now go for jury trial in December 2025. 

Pre-Trial Timeline

From 2021 to 2022, Anthropic obtained an extremely large catalog of pirated books, totaling over seven million by the end. 

First, in early 2021, cofounder Ben Mann downloaded the entire Books3 dataset, around 196,640 pirated books. In June, he obtained 5 million more from Library Genesis (LibGen), and by July 2022, an additional 2 million books joined the database via Pirate Library Mirror (PiLiMi). These are called “shadow libraries,” and they provide free access to large collections of copyrighted materials without authorization, making them popular resources for people who cannot afford content behind paywalls. As such, they are illegal under international copyright law as well as in most countries, including the United States.

Around mid to late 2023, many authors, through GitHub or academic citations, began to discover that their books were being used illegally in training sets. Building on this momentum of legal action against other AI firms like OpenAI and Meta, the plaintiffs filed a formal complaint on October 17, 2023.

The discovery phase of the lawsuit took place from late 2023 to mid 2025. In this time period, both parties exchanged documents and internal emails, recorded testimonies of key witnesses, and requested certain admissions to narrow down disputed issues, revealing crucial facts and solidifying a timeline of events. This led up to the date of summary judgment, a legal decision made by a judge before a case goes to trial. In this circumstance, they determine that certain facts are not in dispute, and the law clearly favors one side, allowing the court to resolve all or part of a case without a jury.

Partial Victory, Pending Trial

In this case, the court evaluated two different categories of training data: legally acquired books and pirated books. Fair use was granted for lawfully obtained texts, with the key conclusion that training AI is a transformative use, not just republishing or copying. This means that the training process adds something new and relevant, with a different purpose or meaning than the original. Moreover, Claude didn’t act as a market substitute for the books, which satisfied the fourth factor of the Fair Use Test: Does the new use harm the market for the original work? Since Claude doesn’t output the books word for word, there is no competition hindering the source material. In the end, Alsup praised the “engineering brilliance” of Claude.

However, his tone changed in regards to the piracy issue, condemning their use as “lawless.” The usage of the content was deemed legitimate, but the illicit acquisition of the content overshadowed the “transformative use.” Furthermore, internal emails revealed that the piracy was intentional to avoid licensing hurdles.

In the end, claims of copyright infringement and piracy will proceed to jury trial, resulting in the partial survival of the authors’ class-action, allowing them to pursue damages. As the first ruling confirming AI training on legal books constitutes fair use, legal clarity was achieved, though Anthropic still isn’t in the clear.

Implications and Outlook: What’s next?

The jury trial in December 2025 will decide whether Anthropic engaged in willful infringement and the amount they must pay for damages, in which case up to $150,000 might be paid per work. If the jury rules in favor of the authors, Anthropic will face massive financial penalties, warning future AI companies to tread carefully in the realm of training liabilities.

Ultimately, though a judicial precedent has been set, the pressure on Congress has increased to define AI copyright rules more explicitly, as questions still exist around AI training infringement. For AI startups, authors, and lawyers alike, this novel territory will bring a set of unique circumstances and challenges.