IP Law Daily, COPYRIGHT—N.D. Cal.: In a first, training of large language model is found to be fair use, (Jun 24, 2025)

By Matthew Hersh, J.D.

The landmark decision may clear the way for some AI models—but it comes with some important limitations.

The unlicensed use of books to train the large language model powering a popular AI “chatbox” services “was exceedingly transformative and was a fair use under Section 107 of the Copyright Act,” the federal court for San Francisco has held. But the court’s landmark decision—the first to directly opine on whether the training of generative AI models could be considered fair use—comes with some important limitations that may distinguish it from other AI-related cases pending throughout the federal judiciary (Bartz v. Anthropic PBC, No. 3:24-cv-05417-WHA (N.D. Cal. Jun. 23, 2025)).

The decision comes in putative class action lawsuit filed by authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson. The three authors took aim at Anthropic PBC, an AI software firm that runs the chatbox known as Claude. When a user prompts Claude with text, Claude responds with text that mimics human reading and writing. Claude can do so, the authors assert, because Anthropic trained Claude using books and other texts—including many of their own works—selected from a central library Anthropic had assembled. Anthropic assembled this library in two different ways: partly by downloading books from pirate websites, and partly by purchasing used print copies and scanning them into its database.

The lawsuitrevolves around three different sets of allegedly unauthorized uses by Anthropic. First, the lawsuit alleges, Anthropic infringed upon their copyrights by using their works to train the large language model underlying Claude. Second, the lawsuit alleges, Anthropic infringed by taking the used print versions of their books and scanning them into digital form. Third, the lawsuit alleges, Anthropic infringed by downloading pirated versions of their books and retaining them in their database for potential further uses.

Anthropic moved for summary judgment on the defense of fair use, leading to this opinion.

Use of materials for training. The court found that the copying of the authors’ works for the purpose of training Claude was a transformative use. In so doing, the court analogized Claude’s training process to the manner in which humans use books for learning. The authors “cannot rightly exclude anyone from using their works for training or learning as such,” the court noted. “Everyone reads texts, too, then writes new texts. They may need to pay for getting their hands on a text in the first instance. But to make anyone pay specifically for the use of a book each time they read it, each time they recall it from memory, each time they later draw upon it when writing new things in new ways would be unthinkable.” Thus, the “purpose and character” of using works to train LLMs, the court found, “was transformative—spectacularly so.”

Nor was there any reason to think that “computers nonetheless should not be allowed to do what people do,” the court found. Importantly, the court noted, Anthropic was not using artificial intelligence technology to create books that would compete in the market against the authors’ own works. Rather, the court noted, Anthropic’s use “stood sufficiently ‘orthogonal’ to anything that any copyright owner rightly could expect to control.” In this sense, the court noted, the purpose and character of using copyrighted works to train LLMs to generate new text was “quintessentially” transformative. “Like any reader aspiring to be a writer,” the court noted, “Anthropic’s LLMs trained upon works not to race ahead and replicate or supplant them—but to turn a hard corner and create something different. If this training process reasonably required making copies within the LLM or otherwise, those copies were engaged in a transformative use.”

Fair use was also firmly supported by the fourth fair use factor—“the effect of the use upon the potential market for or value of the copyrighted work.” Here. the court noted, the copies used to train the LLM “did not and will not displace demand for copies of [the a]uthors’ works, or not in the way that counts under the Copyright Act.” The authors contended that training LLMs “will result in an explosion of works competing with their works—such as by creating alternative summaries of factual events, alternative examples of compelling writing about fictional events, and so on.” The court was prepared to assume that was true—but it made no difference. “[The [a]uthors’ complaint is no different than it would be if they complained that training schoolchildren to write well would result in an explosion of competing works,” the court reasoned. “This is not the kind of competitive or creative displacement that concerns the Copyright Act.”

Nor did it matter for market effect purposes, the court noted, whether Anthropic’s use “would displace (or will) an emerging market for licensing their works for the narrow purpose of training LLMs.” To be sure, the court found, such a licensing market could well develop in the future. Even so, the court noted, “such a market for that use is not one the Copyright Act entitles Authors to exploit.” All of the cases cited by the authors “contemplated losses of something the Copyright Act properly protected—not the kinds of fair uses for which a copyright owner cannot rightly expect to control,” the court noted. The potential licensing market would play no role here.

Scanning of print copies. The court also found that Anthropic made a fair use of the print copies that it purchased when it scanned them into digital form. Anthropic purchased millions of print copies, the court noted, to build a research library. Importantly, however, the court noted, “it destroyed each print copy while replacing it with a digital copy for use in its library (not for sharing nor sale outside the company).” This was a fair use, the court found. “On the facts here,” the court reasoned, “that format change itself added no new copies, eased storage and enabled searchability, and was not done for purposes trenching upon the copyright owner’s rightful interests—it was transformative.”

Storage of pirated copies. But while Anthropic made a fair use in training its LLMs and converted lawfully purchased books to digital form, it did not have the right to retain copies of the works that it acquired unlawfully. “Anthropic downloaded over seven million pirated copies of books, paid nothing, and kept these pirated copies in its library even after deciding it would not use them to train its AI (at all or ever again),” the court noted. A person who downloads works from a pirate site “has infringed already, full stop.” Nor did it matter that Anthropic might later decide to use these copies in the future for the purpose of training LLMs. “There is no decision holding or requiring that pirating a book that could have been bought at a bookstore was reasonably necessary to writing a book review, conducting research on facts in the book, or creating an LLM,” the court noted. Anthropic would face a trial on these copies.

Analysis. While the decision will be seen as a clear win for the AI industry, it comes with two important limitations.

First, the court emphasized that, unlike in other cases, there was no allegation that Claude ever replicated the authors’ content in its responses to user inputs. “When each LLM was put into a public-facing version of Claude, it was complemented by other software that filtered user inputs to the LLM and filtered outputs from the LLM back to the user,” the court noted. “As a result, Authors do not allege that any infringing copy of their works was or would ever be provided to users by the Claude service.” If the outputs seen by users had been infringing, the court noted, “Authors would have a different case. And, if the outputs were ever to become infringing, Authors could bring such a case. But that is not this case.”

Second, the court emphasized, all of the plaintiff’s works that were used in the training set were originally acquired lawfully through the purchase of the used copies. (Some of the books were included in pirated sets, but a full collection was also purchased legally.) Things would likely have been different if the plaintiffs’ works had been obtained only through piracy, the court warned: “This order doubts that any accused infringer could ever meet its burden of explaining why downloading source copies from pirate sites that it could have purchased or otherwise accessed lawfully was itself reasonably necessary to any subsequent fair use.”

The Case is No. 3:24-cv-05417-WHA.

Judge: Alsup, W.

Attorneys: Justin A. Nelson (Susman Godfrey L.L.P.) for Andrea Bartz. Douglas Andrew Winthrop (Arnold & Porter Kaye Scholer LLP) for Anthropic PBC.

Companies: Anthropic PBC

MainStory: TopStory Copyright AINews TechnologyInternet CaliforniaNews GCNNews