In the second landmark decision this week relating to whether use of copyrighted content for training generative AI qualifies as a fair use, Judge Chhabria, in the federal court for the Northern District of California, ordered summary judgment in favor of Meta Platforms Inc. (Meta), finding that Meta’s copying of a group of 13 bestselling authors’ books as training data for use in Meta’s large language training model (LLM) “Llama” was a fair use. Kadrey, et al. v. Meta Platforms, Inc., Case No. 23-cv-0317-VC. This groundbreaking decision out of the NDCA follows Judge Alsup’s ruling earlier this week that Anthropic’s use of legally obtained books for training its LLMs was a fair use, Bartz et al. v. Anthropic PBC, which we covered here.
The orders in both cases determined that the LLM’s use of copyrighted data for training generative AI was “highly transformative” and that the first copyright fair use factor therefore weighed heavily in favor of the AI developers. In both cases, the plaintiffs were unable to demonstrate sufficient market harm to overcome the heavy weight placed on the transformative nature of the AI models. The decisions, however, differed notably as to each judge’s consideration of the source of the copyrighted works and whether the works were obtained through authorized channels or from “pirate websites.”
Use for LLM Training Is Found to Be “Highly Transformative,” and Courts Are Weighing This Fair Use Factor Most Heavily
In Kadrey, Judge Chhabria found that “the first factor focuses on whether the secondary use is ‘transformative,'” and that “there is no serious question that Meta’s use of the plaintiffs’ books had a ‘further purpose’ and a ‘different character’ than the books [and] that it was highly transformative.” This aligns with the decision in Bartz v. Anthropic, where Judge Alsup concluded that the use by Anthropic was “quintessentially transformative.”
Referencing amici arguments that an LLM training on a book is akin to a human reading one, Judge Chhabria refuted this point, finding that “an LLM’s consumption of a book is different than a person’s” because an “LLM ingests text to learn ‘statistical patterns.'” The decision concluded that the comparison between LLMs and humans is inapt, because Meta was not just giving the plaintiff’s books to one person to read, but instead was using large swaths of data to train a generative AI tool. Acknowledging that whether the use is commercial is relevant, and that Meta’s use was commercial, nonetheless the court found the commercial nature of the use was insufficient to overcome the “highly transformative” nature of the technology and “does not tilt the first factor in plaintiff’s favor.”
Judge Chhabria’s Focus on, but Rejection of, Plaintiffs’ Three Theories of Market Harm
In contrast to the three-page analysis of market harm in Bartz v. Anthropic, Judge Chhabria states that the fourth copyright fair use factor is “undoubtedly the single most important element of fair use.” Consistent with this position, he dedicates over twelve pages to considering, and ultimately rejecting, plaintiffs’ theories of market harm.
First, the court found that plaintiffs’ theory that the LLM could reproduce meaningful snippets, and thus serve as a substitute in the market, was “not viable.” Neither expert opined that the tool was able to regurgitate more than 50 words from any of the plaintiffs’ books, “even in response to ‘adversarial’ prompting,” and plaintiffs’ expert conceded that the LLM would not generate “any significant portion” of their books. The same was true in Bartz v. Anthropic, where Judge Alsup concluded that while not demonstrated there, if there was evidence of infringing output, that would be a “different case.”
Second, the court rejected plaintiff’s primary market harm argument that Meta’s actions allegedly diminished plaintiffs’ ability to license their works to the LLM training market. Despite the parties disagreeing on whether the LLM training market exists or is likely to develop, the court held that the market would not be one that plaintiffs “are legally entitled to monopolize.”
Third, the court considered plaintiffs’ theory of market “dilution,” which the court identified as the “only viable theory of market harm”—specifically, the possibility that harm could result from the LLM’s output, since the output could serve as an indirect replacement of plaintiffs’ works. The court acknowledged that market dilution or indirect substitution has not been “particularly important” in copyright cases to date but explained that since this case “involves a technology that can generate literally millions of secondary works, within a miniscule fraction of the time and creativity used to create the original works it was trained on,” and because precedent demands that the doctrine of fair use be flexible to evolving technology, the concept of market dilution is “highly relevant” to this analysis. This diverges from Bartz v. Anthropic, where Judge Alsup rejected this theory since it was “not the kind of competitive or creative displacement that concerns the Copyright Act.” With the stage set, the court determined that market dilution could not be found, however, noting that the theory of market dilution was not raised in plaintiffs’ complaint or affirmative summary judgment motion, and plaintiffs failed to rebut the evidence offered by Meta that there was no such harm.
The Source of the Books Used Was Not Dispositive in This Case
The court rejected plaintiffs’ argument that Meta’s use of “shadow libraries” for training should be dispositive that there was no fair use, finding that it may be relevant, but not dispositive. First, the court found that the source could be relevant to whether the use was a “bad faith” use under the first copyright fair use factor, but that it was insufficient to overcome the transformative nature of the AI technology. Second, the court found it could be relevant if the use encouraged the infringers who created the shadow libraries and perpetuated their copying but found that plaintiffs had not submitted such evidence. Third, the court found that it could be relevant to the potential market harm, but again held that since it was not shown to have benefited the shadow libraries, there was no such harm.
This is in contrast to Bartz v. Anthropic, where Judge Alsup ruled that “Anthropic had no entitlement to use pirated copies for its central library” and the separate use of the works to create a central library was only fair use with respect to works purchased or lawfully accessed.
Looking Forward
These decisions forecast that the transformative nature of the technology will be the hurdle that rights holders must overcome to refute fair use and underscore that any fair use decision will be fact and evidence intensive and highly dependent on the nature of the use and market impact.
