Background
Authors Richard Kadrey, Sarah Silverman, Ta-Nehisi Coates, and others sued Meta Platforms, alleging that Meta downloaded their copyrighted books from internet “shadow libraries”—large unauthorised digital archives—and used them to train its Llama family of large language models (LLMs). The plaintiffs argued that this constituted direct copyright infringement: copying the entirety of their works without licence, payment, or consent. Meta defended on fair use grounds, arguing that training an AI system on existing text is a transformative activity that benefits the public without substituting for the original works.
The case is part of a wave of AI copyright litigation that followed the commercial deployment of generative AI models in 2022–2023. How courts resolve the training-data question is one of the most consequential open issues in copyright law—worth billions of dollars to the AI industry and equally significant to authors, publishers, and other content creators.
The Court’s Holding
Judge Chhabria granted Meta’s cross-motion for partial summary judgment on fair use, finding Meta’s use of the plaintiffs’ books to train Llama was protected. The analysis proceeded under the four-factor fair use framework:
Purpose and character of use: The court found Meta’s use “highly transformative.” An LLM trained on text is not using that text for the same purpose the original author intended—it uses the information, patterns, and structures within the text to develop a system capable of generating new language outputs. The entertainment and information purposes of the original books are incidental to the statistical learning process. Courts have consistently found that using copyrighted works as raw material for a different type of product is more transformative than using them for their original purpose.
Nature of the copyrighted works: Published books are creative works entitled to strong copyright protection, weighing against fair use. But this factor rarely determines the outcome when transformativeness is high.
Amount and substantiality copied: Meta copied entire books, which ordinarily weighs against fair use. However, the court considered this in the context of what was actually used—statistical patterns, not the expressive content of the specific sentences.
Market effect: This is where the case turns decisively. Judge Chhabria found that plaintiffs failed to present evidence of concrete market harm: they did not demonstrate that Llama’s outputs substitute for the original books in any market, nor did they adequately show that a licensing market for AI training has been or will be harmed. The court was sympathetic to a “market dilution” theory—the idea that if everyone uses AI instead of licensed content, demand for original works will dry up—but found plaintiffs had not properly pleaded or proven this theory with evidence. Without concrete market harm, the fourth factor favoured fair use.
Note: As of July 8, 2026, the plaintiffs’ interlocutory appeal of this ruling has been stayed by Judge Chhabria, who indicated the appeal should be presented together with similar rulings in other AI copyright cases as a “tidy package” for the Ninth Circuit to consider.
Key Takeaways
- Under current doctrine, using copyrighted text to train an LLM is likely fair use if the model does not reproduce the original expression and plaintiffs cannot demonstrate concrete market substitution.
- The “highly transformative” finding is significant: LLM training is not simply copying but learning—the output (generated text) does not reproduce the input (original books).
- The court signalled that a well-developed “market dilution” theory might succeed: if plaintiffs can show that widespread AI use systematically reduces demand for licensed content, that could tip the fourth factor.
- This ruling creates tension with cases like Entrepreneur Media v. Meta, where courts have been more receptive to training-data infringement theories, suggesting the area remains unsettled.
- The interlocutory appeal delay means the Ninth Circuit will review multiple AI fair use rulings at once, making the eventual circuit court opinion a major landmark.
Why It Matters
The Kadrey ruling is one of the most significant AI copyright decisions to date. If upheld by the Ninth Circuit, it would validate the business model of every major LLM developer that has used internet-scraped text—without licence or payment—to build billion-dollar AI products. Authors, journalists, musicians, and other rights holders would face a world where their existing works can be used to train AI systems that then compete with them in the content creation market.
The court’s suggestion about the “market dilution” theory is a roadmap for plaintiffs in future cases. If a plaintiff can produce economic evidence that AI model outputs are systematically substituting for licensed content—reducing e-book sales, article subscriptions, or licensing fees—that may tip the balance. The ultimate resolution by the Ninth Circuit, expected to consolidate multiple AI copyright fair use appeals, will be one of the most watched intellectual property decisions of the decade.
Surfaced via Law360 IP (July 2026 appeal timing order).