Written by: Haim Ravia, Dotan Hammer
Two recent decisions from the U.S. Federal District Court for the Northern District of California analyzed the complex issue of whether using copyrighted books to train artificial intelligence (AI) models constitutes fair use under copyright law.
In Bartz v. Anthropic PBC, Judge Alsup ruled on Anthropic’s use of millions of copyrighted books to train its Claude Large Language Model (LLM). Judge Alsup found the use of books to train specific LLMs to be “exceedingly transformative” and a fair use under copyright law. This was because the LLMs produced new text responses without directly outputting infringing copies of the original works to users.
Judge Alsup also determined that converting purchased print books into digital copies for Anthropic’s central library was fair use, as it facilitated storage and searchability without redistribution. However, Judge Alsup denied fair use for the millions of books Anthropic had pirated for its central library. He deemed this initial piracy, to create a permanent, general-purpose library, “inherently, irredeemably infringing” and not justified by subsequent transformative uses. This led to a trial on damages for piracy.
In Kadrey v. Meta Platforms, Inc., Judge Chhabria agreed that Meta’s use of copyrighted books from “shadow libraries” to train its Llama LLMs was “highly transformative”. The crucial point was the “effect of the use upon the potential market,” the most important fair use factor. Judge Chhabria emphasized that generative AI could “dramatically undermine the market” for original works by flooding it with competing AI-generated content. Judge Chhabria explicitly critiqued Alsup’s analogy that this harm is akin to schoolchildren learning to write, calling it “inapt”. However, Judge Chhabria granted summary judgment for Meta because the plaintiffs failed to provide sufficient evidence that Llama’s outputs would dilute the market for their specific works, relying instead on “speculation” or non-cognizable theories of harm. Chhabria suggested that with a better-developed record on market effects, plaintiffs would “often win”.
Both judges agreed on the transformative nature of LLM training but differed significantly on the treatment of library acquisition and the burden of proving market harm from AI-generated competitive works.
Click here to read the decision in Bartz v. Anthropic PBC.
Click here to read the decision in Kadrey v. Meta Platforms, Inc.
