2025 Summer Associate Wade Marshall contributed to this article.
Recently, two Northern District of California decisions revealed fault lines in the forming fair use terrain for GenAI copyright infringement actions. Both cases Bartz et al v. Anthropic PBC and Kadrey et al v. Meta Platforms, Inc. present similar fact patterns: for-profit GenAI companies using copyrighted books to train their LLMs. And both decisions grant summary judgment in favor of those companies, finding that their uses of copyrighted books for training LLMs were fair use. But the courts’ analyses are notably different. In Bartz, Judge Alsup issued a technophilic opinion, extolling the transformative nature of this technology (No. 3:24-05417-WHA, D.I. 231 (N.D. Cal. June 23, 2025)). A few days later, in Kadrey, Judge Chhabria expressed sharp disagreement with Judge Alsup’s approach in Bartz and gave greater consideration to the market effect of GenAI infringement (No. 3:23-03417-VC, D.I. 601 (N.D.Cal. June 25, 2025)). Both courts step away from the significant Thomson Reuters Enter. Centre GmbH v. Ross Intell. Inc. decision rejecting a fair use defense in the District of Delaware earlier this year.
Background
In Bartz, Anthropic placed pirated or purchased-and-destructively-scanned copies of books into a central research library and used some of them to train Anthropic’s LLM Claude. In Kadrey, Meta had trained its LLM Llama mostly on data from Common Crawl as well as websites and book databases, but Meta also procured books from sources that make creative content free to download, regardless of whether that content is copyrighted. A notable fact in both cases is that neither LLM could be used to read or meaningfully access the plaintiff authors’ books.
Bartz Prioritizes Transformation Over Market Effect
The Bartz court conducted separate fair use analyses for each of Anthropic’s alleged uses of copyrighted materials. It found the doctrine covered training LLMs and converting purchased print library copies into digital library copies but not building a central library using downloaded pirated copies. For Anthropic’s use of books for training data, the court makes a sweeping declaration that “the purpose and character of using copyrighted works to train LLMs to generate new text was quintessentially transformative,” even “spectacularly so.” For Anthropic’s format conversions, the court found “the mere conversion of a print book to a digital file to save space and enable searchability was transformative for that reason alone.” Judge Alsup’s analysis anthropomorphizes the AI model, making comparisons to how individuals read, mentally process, and emulate books without engaging in infringing activities. The Bartz court also distinguished Ross. On the issue of “purpose and character,” it held that in Ross, “what was trained was ‘not generative AI’” but rather, “a competing AI tool for finding court opinions in response to a given legal topic,” which was not transformative. The Bartz court instead said that the “better analogue to our facts would be an AI tool trained — using court opinions, and briefs, law review articles, and the like — to receive legal prompts and respond with fresh legal writing.”
The Bartz court leaned less on the fourth fair use factor (i.e., the effect of the use upon the potential market for or value of the copyrighted work), which had been dispositive in Ross. For the training-data use, the court reasoned that the Copyright Act seeks to advance original works of authorship, not to protect authors against competition. Notably, the court rejected the marketplace dynamic that Ross considered, which is that authors could have participated in the emerging market for licensing their works for training LLMs.
But the court reached a different result with respect to the pirated library copies, holding that “[s]uch piracy of otherwise available copies is inherently, irredeemably infringing even if the pirated copies are immediately used for the transformative use and immediately discarded.” The court also noted that, for the fourth factor, pirated copies did have a market displacement effect. Thus, the court “doubt[ed] that any accused infringer could ever meet its burden of explaining why downloading source copies from pirate sites that it could have purchased or otherwise accessed lawfully was itself reasonably necessary to any subsequent fair use.”
Kadrey Reasserts Market Effect’s Primacy
In Kadrey, the court similarly found the use of plaintiffs’ books to train its LLM to be “highly transformative.” But the court emphasized that “[u]nder the fair use doctrine, harm to the market for the copyrighted work is more important than the purpose for which the copies are made.” Specifically, the fourth factor—market harm—was emphasized as the counterweight to the first factor. On this point, Judge Chhabria criticized the Bartz opinion, accusing it of “blowing off the most important factor in the fair use analysis” by “focus[ing] heavily on the transformative nature of generative AI while brushing aside concerns about the harm it can inflict on the market for the works it gets trained on.”
The Kadrey court observed three ways to show market harm: (1) the model regurgitates or outputs substantially similar works; (2) unauthorized copying harms the market for licensing their works for AI training; and (3) the model generates work that are similar enough in subject matter or genre that they compete with the originals. The court found the first inapplicable because the LLM did not allow users to generate any meaningful portion of the plaintiffs’ books. The court also found the second inapplicable, marking a departure from Ross. While Ross recognized that market as the one harmed, Judge Chhabria had a different view, finding that “to prevent the fourth factor analysis from becoming circular and favoring the rights-holder in every case, harm from the loss of fees paid to license a work for a transformative purpose is not cognizable.” But the third route—allowing the rapid generation of innumerate competing works—was a plausible argument for market harm. But plaintiffs did not raise it. The court observed that market dilution is highly relevant in the LLM context given the technology’s potential to flood the market with competing works and that indirect substitution can still count as market harm.
Judge Chhabria repeatedly noted the lack of evidence proffered by plaintiffs on Meta’s alleged use of pirated copies and their impact on the market. Thus, this fact did not move the needle as much as it could have in the analysis.
Judge Chhabria arrived at the same result as Bartz, holding that “this ruling does not stand for the proposition that Meta’s use of copyrighted materials to train its language models is lawful.” Rather, “[i]t stands only for the proposition that these plaintiffs made the wrong arguments and failed to develop a record in support of the right one.”
Conclusion & Takeaways
Within the span of a few months, Ross, Bartz, and Kadrey have given parties three different directions for fair use analysis in GenAI copyright infringement cases. Below are a few key points for parties to keep in mind:
- Transformative use is strongly favored when GenAI systems employ mitigating software to prevent infringing or overly presumptuous borrowing in their outputs.
- Market effect can be a necessary counterweight to transformative use to defeat a fair use defense.
- Market harm may be shown more effectively through market dilution than hypothetical loss of licensing fees, which may not be cognizable in certain courts.
- The use of pirated copies, even if immediately used for the transformative use and thereafter discarded, might not qualify as fair use.
