As lawsuits over artificial intelligence and copyrights continue to unfold, two recent federal court decisions from the Northern District of California provide early insight—conflicting at times—into how judges will begin to draw the boundaries of “fair use” in AI training. In Bartz v. Anthropic PBC and Kadrey v. Meta Platforms Inc., the courts reached different outcomes on whether copying books to train large language models (“LLMs”) violates copyright law. Together, the cases offer preliminary guidance for authors and technology companies navigating this yet-unsettled legal landscape.
Bartz v. Anthropic: Fair Use—But Not for Pirated Books
In Bartz v. Anthropic, the court held that using copyrighted books to train large language models (“LLMs”) qualifies as fair use when the books are lawfully obtained. But, the use of unlawfully obtained, pirated copies of those books did not similarly constitute fair use.
Background: Authors Sue Anthropic Over Use of Books
Plaintiffs—authors of various nonfiction and fiction books—sued Anthropic PBC, the developer of the AI platform Claude, for copyright infringement. Anthropic had acquired copies of the authors’ books through two primary means: (1) downloading pirated copies from “shadow libraries” online, and (2) purchasing print books, scanning them into digital form, and destroying the print copies. The company stored the digitized books in a central repository and used subsets to train Claude’s LLMs. Plaintiffs alleged this unauthorized copying constituted copyright infringement.
The Court’s Ruling: Fair Use Applies in Part
The court evaluated several distinct uses of the copyrighted works:
- Use in Training LLMs: The court held that Anthropic’s use of copyrighted books to train LLMs was “spectacularly” transformative and qualified as a fair use. The court emphasized that the LLMs did not output infringing content to users and that the training process was akin to how human authors learn from reading prior works. As such, using the books to develop AI-generated content constituted a new and different purpose.
- Scanning of Purchased Books: The court also found that scanning lawfully acquired print books into digital format for internal use (and then destroying the print copies) was a fair use. Citing precedents involving format-shifting and archival use, the court noted that Anthropic replaced each purchased print copy with a corresponding digital version stored in its internal research library—without distributing or sharing the digital copies.
- Retention of Pirated Copies: However, the court rejected Anthropic’s fair-use defense for pirated copies retained in its central library. The court found that building a permanent digital library from unauthorized copies was not transformative and amounted to direct infringement. The fact that Anthropic later used some of the pirated books to train AI models did not retroactively justify the unlawful acquisition.
Notably, the court rejected the argument that the LLMs would displace the authors’ works in the marketplace by creating competing works. The court concluded that training LLMs to produce distinct content that competes for readers was no different than humans who read a book and then write their own books, drawing on their full body of knowledge.
Mere days after Bartz was decided, another judge in the Northern District of California issued an opinion on AI and fair use, using a different approach.
Kadrey v. Meta Platforms Inc.: Authors Lose on Market Harm
In Kadrey v. Meta Platforms Inc., the court ruled that the plaintiff authors did not prove Meta infringed their copyrights by using books to train its large language models (“LLMs”). The court recognized that AI training on copyrighted works can pose serious risks to the market for creative content but concluded the authors in this particular case failed to present evidence of meaningful harm.
The Dispute
Thirteen prominent authors—including Sarah Silverman and Ta-Nehisi Coates—sued Meta for copyright infringement after the company downloaded hundreds of their books from shadow libraries and used the books, among other materials, to train its LLMs (known as Llama).
The authors argued that Meta’s copying was not protected by fair use and that Meta’s models harmed the market for their books.
The Court’s Decision
The court sided with Meta, granting summary judgment on the core copyright claim. But the decision turned less on whether AI training can be fair use and more on whether the plaintiffs marshaled enough evidence in the case.
Key takeaways from the ruling:
- Transformative Use: The court found Meta’s use of the books to train general-purpose AI tools to be “highly transformative.” The models are designed to perform many tasks that are very different from the authors’ intended use of their books. [Notably, the Kadrey decision did not take issue with the downloading of pirated copies from shadow libraries, whereas Bartz concluded fair use did not apply where pirated copies of books are used for training].
- Minimal Output Overlap: The authors claimed that Meta’s LLMs could regurgitate their books, but both sides’ experts agreed the models could only reproduce tiny snippets—no more than 50 words—even under aggressive prompting.
- Licensing Market Not Protected: The court rejected the authors’ argument that Meta’s failure to license their books for the LLM training constituted market harm. Courts do not recognize a lost opportunity to license a transformative use.
- Missed Opportunity on Market Dilution: The authors might have succeeded by showing that Meta’s models are capable of generating works that compete with their books in the marketplace. But the court found they failed to present evidence on that front. [Notably, the Bartz decision had rejected this market-dilution argument].
A Narrow Win for Meta
The Kadrey ruling is narrow: It does not hold that all LLM training is fair use, or that Meta’s conduct was necessarily lawful in other contexts. Rather, the court emphasized that these authors “made the wrong arguments and failed to develop a record in support of the right one.”
Comparing the Two
- Both courts agreed that using books to train AI models can be transformative and may qualify as fair use.
- Bartz indicates that it matters how content is acquired—pirated copies are not shielded by fair use. Kadrey, on the other hand, did not take issue with the manner in which the content was acquired.
- Kadrey indicates that plaintiffs must prove market harm. Simply pointing to lost licensing opportunities will not cut it; evidence that AI-generated works substitute for the originals may be required. Bartz, of course, rejected the market-dilution argument.
What This Means Going Forward
For authors and rights holders, Bartz underscores the need to police infringing content online to minimize the risk that pirated versions of works are available for unauthorized uses. Kadrey suggests it may be important to develop strong evidence showing how AI training and outputs undermine real markets for creative works. Future lawsuits may focus on competitive harm from AI-generated substitutes.
For AI developers, the rulings provide some comfort but also important guardrails:
- Using lawfully obtained materials for transformative training is more defensible.
- Reliance on pirated copies remains a major liability risk.
- Courts may examine whether outputs compete with copyrighted works.
Bottom Line: These decisions provide early insight into the different ways in which courts may ultimately address the intersection of AI and copyrights. Despite their conflicting nature, the decisions offer some important guidance. Courts are open to fair use arguments for AI training, but the outcome may be highly fact dependent—with particular focus on the source of training data and/or the impact on markets for original works. That said, the decisions will likely be appealed. And there are others in the pipeline. Stay tuned.
