July 22, 2025

Judge Lets Llama March Forward

By Emily Poler

As I noted in my previous post, there have been two recent decisions involving fair use and AI. Last time around, I wrote about the case brought by a group of authors against Anthropic. This time, we turn to the other case where there was a recent decision — Kadrey v. Meta Platforms, Inc. — which was also brought by a number of writers. 

To cut to the chase, in Kadrey (a/k/a the Meta case), the judge granted Meta’s motion for summary judgment on the issue of fair use, finding that Meta’s use of Plaintiff’s copyrighted works to train its large language model (LLM) — Llama — was highly transformative and, therefore, Meta’s processing of Plaintiff’s work was fair use.

As in the Anthropic case, Meta ingested millions of digitized books to train Llama. It obtained these books from online “shadow libraries” (i.e., pirated copies). Unlike the judge in the Anthropic case who found fair use only where Anthropic paid for the initial copy and actually used the works in developing its LLM, the judge in the Meta case was unfazed by Meta’s use of pirated works. According to the judge in the Meta case, the fact that Meta started with pirated works was irrelevant so long as the ultimate use was transformative. 

In other words, the ends justify the means, which seems like a pretty novel way of deciding fair use issues. 

Also of interest: the judge in the Meta case didn’t spend any time discussing the exact nature of Meta’s use. Instead, he assumed that the use was Llama itself. This stands in pretty sharp contrast to the judge in Anthropic who spent quite a bit of time looking at the various ways Anthropic used and stored the downloaded books. This seems not great because the intermediate steps (downloading, storing, cataloging, and making pirated works available to internal users) represent copyright infringement on their own. The court here, however, largely glossed over these issues because all of the “downloads the plaintiffs identify had the ultimate purpose of LLM training.” 

With that said, the judge in the Meta case did invite other authors to bring a second case against Meta and provided those putative plaintiffs with a roadmap of the evidence that would support a ruling that Meta’s use of their works was not fair use. Here, the judge suggested a novel way of looking at the fourth fair use factor, which focuses on market impact, proposing that indirect substitution could weigh against a finding of fair use. That is, the judge said a court could consider whether someone, for example, buying a romance novel generated by Llama substitutes for a similar novel written by a human, here writing that other plaintiffs could pursue “the potentially winning argument — that Meta has copied their works to create a product that will likely flood the market with similar works, causing market dilution.” 

While this certainly has some appeal, it also seems a little unwieldy. Does everyone who writes a romance novel get to point to content generated by LLMs and say that’s substituting for their work? What happens if a book spans genres? How much of a market impact is required? 

Overall, the cases against Anthropic and Meta represent some pretty big wins for AI platforms at least as far as copyright infringement goes. However, there are still plenty of areas of uncertainty that should keep things very interesting as these cases march on.