BLG Alert: Is AI training fair use under U.S. copyright law?

Overview

Recent weeks have seen big developments in the rapidly evolving tug of war between copyright owners and those who use their works, sometimes without permission, to train AI models. Courts have not yet settled the question whether it is fair use, and thus legal, to train AI models on copyrighted material owned by others. But several courts have now weighed in.

Background

Earlier this year, a Delaware district court held that AI training is not fair use. Thomson Reuters Enter. Ctr. GMBH v. Ross Intelligence Inc., No. 20-cv-613, 2025 WL 458520, at *10 (D. Del. Feb. 11, 2025) (Bibas, J., sitting by designation from the Third Circuit). In another case, in which our firm represented the plaintiff, the court deferred the issue until trial. UAB "Planner5D" v. Meta Platforms, Inc., No.19-cv-3132 (N.D. Cal. Mar. 27, 2025), Dkt. No. 627 (Orrick, J.) (tentative ruling). The case settled shortly after that order.

Two new thoughtful orders recently wrestled with these issues. Bartz v. Anthropic PBC, No. 24-cv-5417, 2025 WL 1741691 (N.D. Cal. June 23, 2025) (Alsup, J.); Kadrey v. Meta Platforms, Inc., No. 23-cv-03417, 2025 WL 1752484 (N.D. Cal. June 25, 2025) (Chhabria, J.). These two decisions gave narrow fair use wins to AI companies, but in different ways that leave many questions unanswered. Still, they offer some guidance to litigants in this new and evolving area of law.

The Latest Opinions

Both Bartz and Kadrey involved large language models trained on extensive libraries of digitized books. In Bartz, OpenAI used both pirated digital books downloaded from “shadow libraries” and scans of lawfully acquired physical copies. In Kadrey, Meta relied entirely on pirated books.

In Bartz, Judge Alsup found OpenAI’s scanning was fair use, but the downloading of pirated books was not. OpenAI’s scanning was fair use, according to the court, because it was simply a “format change”: a non-infringing physical copy was destroyed and replaced with a digital one. In contrast, the court doubted that downloading pirated copies could be fair use even if solely for AI training: “Such piracy of otherwise available copies is inherently, irredeemably infringing even if the pirated copies are immediately used for the transformative use and immediately discarded.” But having articulated a broad no-fair-use rule for pirated works, Judge Alsup stepped back, avoiding the question by issuing a narrower ruling. He grounded his finding of no fair use on OpenAI’s maintenance of a central digital library of the pirated works, an infringing act not solely tied to AI training.

In Kadrey, Meta also used pirated copies of books for training, but Judge Chhabria reached the opposite conclusion on the significance of the pirating. He found it was fair use to train Meta’s LLM on pirated copies and that it was also fair use for Meta to create and maintain its central library of the pirated copies. Such a library of pirated works was fair use even though many of the works were not used for training, but were only studied for suitability or kept for potential study.

Judge Chhabria emphasized, however, that his holding was based on the limited record before him, noting that the plaintiffs had waived their best argument against fair use. Future plaintiffs, Judge Chhabria explained, likely could defeat a fair use defense by establishing that the defendant’s product “will likely flood the market with similar works, causing market dilution.” Interestingly, Judge Alsup considered and rejected this very argument in Bartz. He likened it to arguing “that training schoolchildren to write well would result in an explosion of competing works,” Judge Chhabria disagreed with this analogy, explaining that “using books to teach children to write is not remotely like using books to create a product that a single individual could employ to generate countless competing works with a miniscule fraction of the time and creativity it would otherwise take.”

Kadrey and Bartz each gave narrow fair use wins to generative AI companies, but the law remains unsettled and the landscape murky. Judge Alsup made it fairly clear in Bartz that he would not have found fair use to train on pirated books, suggesting that, unlike Judge Chhabria, he might have denied fair use in Kadrey. Judge Chhabria, meanwhile, proposed in Kadrey that sufficient evidence of market dilution could defeat fair use. This raises a question whether, in a case like Bartz involving digitization of physical works, Judge Chhabria might have ruled differently.

What’s clear from these two contrasting decisions is how early we are in the evolution of the law of AI. Like AI itself, the law in this area is poised for rapid and consequential growth.