Meta Wins Legal Battle in Generative AI War: Fair Use Analysis in Kadrey v. Meta Platforms, Inc.

Meta scored a major victory in its war over if its use of copyrighted materials to train its generative AI model is fair use.  In its June 2025 ruling in Kadrey v. Meta Platforms, Inc., the U.S. District Court for the Northern District of California applied the four statutory fair use factors under 17 U.S.C. § 107 to assess whether Meta’s use of copyrighted books to train its large language models (Llama) constituted permissible use. The Court ultimately sided with Meta at the summary judgment stage, but not because the conduct was clearly lawful. Rather, the decision was due to deficiencies in the plaintiffs’ evidentiary presentation.

Under U.S. copyright law, the fair use doctrine allows limited use of copyrighted material without permission from the rights holder. Courts analyze four non-exclusive factors to determine whether a particular use qualifies as fair:

1. The purpose and character of the use
2. The nature of the copyrighted work
3. The amount and substantiality of the portion used
4. The effect of the use upon the potential market for or value of the original work

Below is a detailed analysis of how the Court weighed each factor and the broader implications.

• Factor One: The Purpose and Character of the Use

This factor examines whether the use of the copyrighted material serves a different purpose than the original and whether that purpose is “transformative.” The Court strongly favored Meta on this factor, finding that its use of the plaintiffs’ works was highly transformative.

Whereas the plaintiffs’ books were created to be read for entertainment, education, or literary appreciation, Meta used them to train an AI tool capable of responding to prompts, performing computational tasks, and assisting users across a wide range of domains. The Court likened this to the transformative uses upheld in prior precedent, including Google v. Oracle, where software was repurposed to create a new platform, and Google Books, where books were digitized to enable search functionalities.

Although the use was undeniably commercial because Meta projects generative AI revenues in the trillions, the Court held that commercialism is not dispositive. The transformative nature of the use carried greater weight. Moreover, Meta’s controversial acquisition of the works from “shadow libraries” was not ignored but deemed a secondary consideration in the absence of concrete evidence that this conduct supported broader infringing ecosystems.

Takeaway: Highly transformative use, even if commercial, can weigh strongly in favor of fair use. However, courts may still consider the manner of acquisition and the user’s good faith in marginal cases.

• Factor Two: The Nature of the Copyrighted Work

This factor examines whether the original works are factual or creative, and whether they were previously published. The Court determined that the plaintiffs’ works, including novels, plays, memoirs, and nonfiction, were highly expressive and therefore entitled to robust copyright protection.

Meta tried to analogize its use to prior “intermediate copying” cases such as Sega and Connectix, in which functional code was copied to extract unprotected elements. That argument failed. Unlike software reverse-engineering, training a language model on literary works requires extracting and learning from the creative structure and expression, including word choice, grammar, tone, and syntax. All of these are copyright-protected elements.

Nonetheless, the Court emphasized that this factor generally carries less weight in the fair use analysis, particularly when the works are published and therefore already in public circulation.

Takeaway: This factor will usually favor authors of literary and creative works but rarely dictates the outcome unless paired with compelling arguments under the other factors.

• Factor Three: The Amount and Substantiality of the Portion Used

Here, the Court assessed whether the amount copied was reasonable in light of the transformative purpose. Although Meta copied the plaintiffs’ works in full, which would normally count against fair use, the Court found the copying to be proportional and necessary given the functional requirements of training a large language model.

Full-text ingestion of books allows LLMs to detect long-range linguistic patterns and develop more coherent outputs. The Court likened this to uses in HathiTrust and Google Books, where wholesale copying was justified because the entirety of the works was needed to achieve a legitimate transformative purpose.

Importantly, the Court observed that Meta had implemented post-training mitigations to prevent Llama from regurgitating large segments of training data. The plaintiffs’ expert could not demonstrate that the model reproduced more than 50-word snippets from any book, even under adversarial testing.

Takeaway: Courts may tolerate complete copying of protected works if that use is necessary for a transformative goal, particularly when safeguards prevent downstream market substitution or replication.

• Factor Four: The Effect of the Use Upon the Market for the Original Work

This is the most critical and determinative factor in the Court’s fair use analysis. The central question is whether Meta’s use would serve as a substitute for the original works or materially harm the market for them.

The plaintiffs proposed three theories of harm. First, they argued that Llama could output infringing passages. This failed for lack of factual support. Neither expert could show that the model produced substantial portions of the plaintiffs’ works in response to user prompts.

Second, the plaintiffs claimed that Meta’s unauthorized use undermined a nascent market for licensing books as AI training data. The Court rejected this theory as circular. Fair use analysis does not grant authors the right to monopolize every conceivable use of their work, particularly when the use is transformative.

Third, the plaintiffs suggested that generative AI may flood the market with similar works, reducing demand for original fiction and nonfiction. This was described as a potentially winning argument, especially for lesser-known authors vulnerable to displacement. However, the plaintiffs failed to offer evidence that Meta’s outputs, current or anticipated, would dilute the specific markets for their books.

Takeaway: Market dilution is an emerging and credible theory of harm in the generative AI context, but courts will require detailed, plaintiff-specific evidence showing actual or likely substitution, not generalized speculation.

• Final Thoughts

While the Court granted summary judgment for Meta in this case, its decision was closely tied to the evidentiary failures of the plaintiffs rather than a broad endorsement of using copyrighted works for AI training.

For future plaintiffs, the takeaways are clear. Success will hinge on providing detailed, market-specific evidence showing how AI outputs compete with or substitute for protected works.

Likewise, AI developers should not assume blanket protection under fair use. Even transformative, commercial models may invite liability if they materially impair rightsholders’ economic incentives. Licensing agreements, technical mitigations, and transparency about training sources will remain critical tools in mitigating litigation risk.

Please reach out if you would like to discuss the broader implications of this ruling on your copyright strategy or AI development initiatives.