/>iOn June 23, 2025, Judge Alsup of the Northern District of California issued an opinion in Bartz v. Anthropic, 2025 WL 1741691 (N.D. Cal. June 23, 2025), that may set the stage for many artificial intelligence (AI) copyright disputes in the months and years to come. Plaintiffs—authors and the corporate entities they created to market their copyrighted works—alleged that defendant AI firm Anthropic infringed on their copyrighted works in the development of its chatbot Claude and the underlying large language models (LLMs) that make Claude tick.
Plaintiffs alleged that Anthropic infringed on their rights in three ways. First, they argued Anthropic infringed by using plaintiffs’ books to train Claude’s LLMs. Second, they alleged that Anthropic purchased physical copies of their books, which they then digitized without authorization to create a central digital library. Third, they argued Anthropic used millions of pirated copies of theirs and others’ books to further develop the creation of the central digital library. On summary judgment the Court held that the first two theories of liability constituted fair use, but the third did not.
First, the Court held that “using [the copyrighted] works to train LLMs was transformative – spectacularly so.” Judge Alsup found that Anthropic trained its LLMs not to compete with the copyright holders commercially, but “to turn a hard corner and create something different.” According to the Court, training LLMs with copyrighted works, where the LLMs were designed “to ensure that no infringing output ever reached users,” is fair use in the same way “that training schoolchildren to write well would result in an explosion of new works.”
Second, the Court also found that the digitization of print books that helped build Anthropic’s digital library constituted fair use because “[i]t did not create new copies to share or sell outside” Anthropic. The Court found this digitization project did not have a direct commercial benefit to Anthropic and did not result in the creation of additional copies of books on top of those already purchased, and therefore was not infringing.
That left plaintiffs with one avenue to recovery: the alleged use of pirated books to build Anthropic’s central digital library. According to plaintiffs, Anthropic downloaded millions of pirated copies of books from various online sources, paid nothing for these copies, and maintained them in its central library with no intention of discarding them. The Court found these actions violative of the copyright protections authors are entitled to. According to the Court, such pirated works “displaced demand for [the] Authors’ books – copy for copy” and their use – building a library of works – was not transformative.
Centralized digital libraries are critical to the fulsome development of LLMs. The Anthropic decision provides helpful guidance on how AI companies might be able to create these libraries while staying within the bounds of permissible activities under copyright laws.
