Second District Court Rules AI Training Can Be Fair Use
Summary
On June 25, 2025, Judge Vince Chhabria in the Northern District of California issued an order in Kadrey v. Meta Platforms, granting partial summary judgment to Meta on fair use as a defense to the plaintiffs’ claims that training was copyright infringement. This decision follows a recent, similar finding in Bartz v. Anthropic that AI training on copyrighted works can be fair use, although the two decisions focus on different aspects of the fair use analysis.
Plaintiffs alleged that Meta’s use of pirated copies of their books to train its large language model (LLM), Llama, constituted copyright infringement. While the court clarified that its ruling does not stand for the proposition that any use of copyrighted materials to train its LLMs is lawful, it found that the plaintiffs failed to rebut Meta’s fair use defense or present evidence in support of plaintiffs’ arguments. The court evaluated the four fair use factors as follows:
- Purpose and Character. This factor favored Meta, because training LLMs that “can be used to generate diverse text and perform a wide range of functions” is highly transformative. Because Meta’s ultimate use was transformative, the court found that Meta’s downloading of the books from “shadow libraries” was also transformative.
- Nature of Work. This factor favored plaintiffs, because their books—mostly novels, memoirs, and plays—are highly expressive works. The court rejected Meta’s argument that the training process only wanted to access the “functional elements” of the works, finding that statistical relationships are the product of creative expression.
- Amount Used. This factor favored Meta, because its LLMs do not output any meaningful amount of the plaintiff’s books. The court noted that “the amount copied doesn’t seem especially relevant in this case” because whether a whole or partial work was copied, Llama does not provide the copyrighted material to the public. The court also found that copying full texts was reasonable given the needs of LLM training and that the amount made available to the public was minimal.
- Market Impact. This factor also favored Meta. The court identified three potential theories of harm raised by the parties, attributing the first two to the plaintiffs’ complaint and the third to Meta’s expert: (1) harm from the model regurgitating the works; (2) harm to the market for licensing the works for AI training; and (3) harm based on competition from similar works generated by the model. For all of these, the court found plaintiffs’ argument “half-hearted” and that the plaintiffs “fail[ed] to present meaningful evidence.” The court found that the first theory fails because Llama does not allow users to generate any substantial portion of the books. The second fails because the market for licensing for AI training is not one the plaintiffs are legally entitled to monopolize. The court found that the third theory—which was not raised by plaintiffs—failed because of the lack of concrete evidence of market substitution or dilution. Because Meta offered at least some evidence that its copying has not caused market harm, the court concluded that this factor favored Meta.
In dicta, Judge Chhabria contemplated AI training cases where plaintiffs have better-developed records on the market effects that could rebut a fair use defense. In pure hypothetical, the opinion stated that: “it’s hard to imagine that it can be fair use to use copyrighted books to develop a tool to make billions or trillions of dollars while enabling the creation of a potentially endless stream of competing works that could significantly harm the market for those books.” The court also contemplated cases that might present an even stronger case for fair use, such as using copyrighted books to train an LLM for nonprofit purposes or where plaintiff’s works are unlikely to face meaningful competition from AI-generated ones.
Ultimately, because Meta’s use was highly transformative, the court found that plaintiffs needed to win decisively on the fourth factor to win on fair use. Absent evidence of market dilution, the court found that Meta was entitled to summary judgment on its fair use defense to the claim that copying plaintiffs’ books for use as LLM training data was copyright infringement. Similar to the Bartz case, plaintiffs’ claims regarding the distribution of copies via torrenting remain at issue.
