Until two weeks ago, no U.S. court had ruled on whether training generative AI models on copyrighted works could constitute a fair use, or if the simple act of training such models without a license would constitute copyright infringement. Two new summary judgment rulings out of the Northern District of California have now set the stage for how future courts may interpret fair use in the context of AI training—Richard Kadrey, et al., v. Meta Platforms, Inc. and Andrea Bartz, et al. v. Anthropic PBC. That stage poses significant dangers for technology companies.
In both cases, the court held that using copyrighted books to train large language models (LLMs), like Anthropic’s Claude and Meta’s LLaMA, may qualify as a fair use. Yet neither case constitutes an unalloyed victory for the defendants. Indeed, the two new decisions, while conflicting on important particulars, both suggest that some plaintiffs could be in a strong position to defeat a fair use defense—especially where the copyrighted content is not books. Both decisions offer key insights for rights holders, AI developers, and litigators navigating this unsettled frontier.
The Kadrey and Bartz fair use findings rest on some important commonalities. Both courts held that using copyrighted books to train LLMs is transformative, even when the full text of a work is ingested. Copying the full text of the plaintiffs’ copyrighted books was found to be both necessary and reasonable for LLM training, even though the courts acknowledged the books are creative works, a factor which traditionally weighs against fair use. The purpose of AI training—to enable models to learn statistical patterns of language to generate new content, rather than to reproduce or compete with original works—proved to be determinative in finding a transformative use under the first factor of the fair use analysis. This factor is the second most important factor in a fair use analysis.
While agreeing on the application of these foundational principles, the Kadrey and Bartz decisions sharply diverge in their treatment of data acquisition, the implications of using pirated materials, and market harm.
The most significant difference is how the courts framed their analyses of the fair use defense. In Anthropic’s case in Bartz, the Court divided the AI training process into two distinct stages: data collection and model training. Thus, when the Bartz Court found that training AI models was transformative and a fair use, this was only a partial defense to liability for Anthropic because the Bartz Court separately analyzed how Anthropic acquired its training data. Anthropic used two different approaches for this latter task, leading to different results for the fair use analysis. In the first approach, the Court concluded that permanently converting legally purchased hardbacks into a digitally-formatted “central library” was transformative and a fair use. In the second, the Court took great issue with Anthropic’s use of pirated copies of the Plaintiffs’ novels from online “shadow libraries.” For those copies, it ruled that building and retaining a "central library" of pirated books could not be a fair use. Notably, the Bartz Court referred repeatedly to Anthropic’s independent goal of building a library of “all the books in the world” and to its practice of retaining copies of pirated books, even after Anthropic had decided never to use them in training its models again. Ultimately, because of the two different sources of books used in training Claude, the Bartz Court granted Anthropic summary judgment as to the lawfully acquired and converted hardbacks, but it denied Anthropic’s motion with respect to its use of pirated online materials.
The Kadrey Court, by contrast, collapsed Bartz’s multi-stage analysis into a single step. While the plaintiffs in Kadrey, like in Bartz, alleged that the accused model had been trained on pirated data, the Kadrey Court focused only on how Meta used the data and whether its use in training LLaMA was transformative. The Court did not conduct a separate fair use analysis into how Meta acquired the works. Finding the model’s use of the books was transformative then carried the Court a long ways towards a fair use finding.
The second key distinction between the Anthropic and Meta cases rests in their analyses of market harm. The Kadrey Court emphasized that the harm an infringing work inflicts on the market, or potential market, for the copyrighted work is “the single most important element” in determining fair use and thus focused its most attention on this element. And its ruling was consistent with the concerns it had expressed at the motion hearing: the Court demonstrated substantial apprehension that a model trained with copyrighted books “might harm the market for those works…by helping to enable the rapid generation of countless works that compete with the originals, even if those [new] works aren’t themselves infringing.” But the Court was stymied in analyzing this market dilution theory and could not rule in the plaintiffs’ favor on summary judgment. The plaintiffs there had not pled, argued or offered any evidence to demonstrate a triable issue of fact as to market dilution causing harm to the markets for their works. The Meta Court thus found that the “most important” factor in fair use of market harm favored Meta, and, when combined with its finding that Meta’s use of the books was transformative, that Meta’s fair use defense applied. In short, while the Court dismissed the plaintiffs’ claims because of a lack of evidence, it left the door wide open for future plaintiffs to pursue claims with better developed theories and evidence substantiating market harm.
The Anthropic Court’s approach to the possibility of market dilution by AI-written books was starkly different. While the plaintiffs in that case did argue that AI models trained with their works would generate an “explosion” of competing works, the Bartz Court gave this argument short shrift. It viewed any works generated by AI models as merely being “competition” for the plaintiffs’ books—a type that the Copyright Act is not intended to prevent. The Meta Court, by contrast, viewed such competition as being (at least potentially) an “indirect substitution”, which is protected under the Copyright Act. It noted that the models were more effective because they had been trained on the creative expression in the copyrighted novels and highlighted that AI models, unlike infringing human authors, could “generate literally millions of secondary works, with a miniscule fraction of the time and creativity used to create the original works [on which] it was trained….” This difference in the scale of the threat of market substitution, disregarded by the Anthropic Court, weighed heavily with the Meta Court.
Taken together, the Anthropic and Meta decisions offer a glimpse into what to expect moving forward and into the dangers that technology companies face from copyright suits. While both cases affirm the possibility that AI training can qualify as fair use, they also reflect emerging fault lines along which liability could be found: between lawful and unlawful data acquisition, between training and storage, and between presumed and proven market harm.
In considering these opinions, one also cannot ignore that the plaintiffs in both cases were novelists. Other types of plaintiffs, such as newspapers, visual artists, or music publishers, may find it much easier to prove that the outputs of AI models infringe on their works, are less transformative, and that they dilute the demand for their works in a way that constitutes cognizable market harm. The prospects for a successful fair use defense in such circumstances seems uncertain at best. For litigators and rights holders, these decisions thus mark both a moment of clarity and a preview of the complexities to come.
Key Takeaways
- Use context matters. Courts may separate training from storage or reuse, and assess each under a different legal lens.
- Lawful acquisition counts, especially when a court analyzes acquisition separately from training. Rights holders should monitor whether platforms are sourcing data from authorized, licensed repositories. To discourage courts from analyzing data acquisition separately, companies that have collected unlicensed works should make them available only for the purpose of training their generative AI models. If the companies decide to stop using certain works (or data sets) for training purposes, then those works should be promptly deleted.
- Evidence of market harm is essential. Speculation about market harm or regurgitation is not enough. Rights holders must offer evidence of actual output overlap or market dilution. Rights holders should gather such evidence early in litigation, or even before, as defendants may move quickly for summary judgment in the hope that any market-based evidence that could support a dilution argument is not available to oppose it.
- Future risks and uncertainties persist. Output infringement remains untouched and future decisions focusing on LLM-generated content that mirrors training materials are more likely to find in favor of the copyright holders.
