No harm, no win: A cautionary tale of Kadrey v. Meta Platforms, Inc.

Meta prevailed on fair use, technically speaking.

In Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417-VC, 2025 WL 4123456 (N.D. Cal. June 25, 2025), the Northern District of California addressed whether Meta's use of copyrighted books to train its LLAMA language models constituted fair use under 17 U.S.C. § 107.

Plaintiffs, a group of prominent authors including Sarah Silverman and Junot Díaz, alleged that Meta acquired their works through shadow libraries, incorporating them into the datasets that were used to train Meta's LLAMA 1, 2, and 3 models. Meta acknowledged both downloading the copyrighted books via BitTorrent and incorporating them into the datasets. Still, Meta argued that its use of the copyrighted works was transformative and protected by the fair use doctrine.

The court concluded that while Meta did copy plaintiffs' copyrighted works without permission, plaintiffs' failure to present a viable theory of market harm -- particularly the absence of evidence that LLAMA substitutes for their books -- proved fatal to their case. Dkt. 598 at 40.

Court's analysis of fair use

The court reminded us that "[w]hat copyright law cares about, above all else, is preserving the incentive for human beings to create artistic and scientific works." Dkt. 598 at 1.

With that principle in mind, the court analyzed Meta's conduct under the four statutory fair use factors set forth in 17 U.S.C. § 107 and interpreted by the U.S. Supreme Court in Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569 (1994). Id. at 13-36. These factors -- (1) the purpose and character of the use, (2) the nature of the copyrighted work, (3) the amount and substantiality of the portion used, and (4) the effect of the use upon the potential market -- must be weighed together and considered in context.

While Meta argued that its use was transformative and socially beneficial, the court emphasized that the analysis must ultimately serve the core goal of copyright law: encouraging and protecting creative labor. Id. at 1. With that lens, the court walked through each factor -- placing particular weight on the fourth. Id. at 13-36.

1. Purpose and character. The court found Meta's use of the copyrighted works highly transformative. While the books were created to be read, LLAMA's use was to learn patterns and relationships between words for the purpose of generating new content. Dkt. 598 at 17. That the books were copied verbatim was not decisive; the use, in Judge Chhabria's view, had "further purpose and different character." Id. at 16.

Even though Meta's LLAMA models are part of a massive commercial enterprise, the court held that the highly transformative nature of the use outweighed the commercial aspect. Id. at 15-22.

2. Nature of the work. The second fair use factor typically weighs against fair use when the original works are highly creative, as the plaintiffs' books unquestionably are. Meta argued otherwise, contending that it used the books only to extract unprotected "functional elements" like grammar, syntax, and word frequency -- relying on 9^th Circuit "intermediate copying" cases such as Sega Enterprises Ltd. v. Accolade, Inc., 977 F.2d 1510 (9th Cir. 1992), and Sony Computer Ent. v. Connectix Corp., 203 F.3d 596 (9th Cir. 2000). In those cases, fair use applied where copyrighted code was copied solely to access non-expressive elements for compatibility purposes. Id. at 123.

But the court found the analogy inapplicable. Meta's use depends on expressive content: LLAMA learns from the creative structure of the plaintiffs' writing -- word order, vocabulary, and syntax -- all of which are protected forms of expression. See Harper & Row, Publishers, Inc. v. Nation Enters., 471 U.S. 539, 548 (1985). Dkt. 598 at 23-24.

Meta also invoked Google Books, 804 F.3d 202 (2d Cir. 2015), but the court distinguished it. That search index was content neutral; LLAMA, by contrast, needs high-quality expression to function effectively. Thus, Meta's use could not be described as neutral or detached from the plaintiffs' creative work. Id. at 24.

Still, the court treated this factor as neutral overall. Because the works were already published and this factor plays a modest role in fair use analysis, it did not weigh heavily in the outcome. See VHT, Inc. v. Zillow Grp., Inc., 918 F.3d 723, 744 (9th Cir. 2019). Id.

3. Amount and substantiality. The court found that this factor favored Meta because the "amount Meta copied was reasonable given its relationship to Meta's transformative purpose." Id. at 25. Meta copied entire books -- hundreds of them -- yet the court emphasized that "feeding a whole book to an LLM does more to train it than would feeding it only half of that book," thus making the whole sale copying "reasonably necessary" in the eyes of the court. Id.

4. Market effect. Harkening back to Bartz v. Anthropic PBC, the court rejected Judge Alsup's suggestion in Bartz that training AI on copyrighted books is no more harmful than using them to teach "schoolchildren to write." Judge Chhabria called that analogy "inapt," emphasizing that generative AI enables the rapid production of countless competing works with minimal human input -- posing a qualitatively different threat to human authorship. Dkt. 598 at 3.

Unlike Bartz, which leaned heavily on the transformative nature of AI tools, the Kadrey court focused squarely on the fourth fair use factor: market harm. Quoting Harper & Row, the court reiterated that market effect is "more important than the purpose for which the copies are made." Id. at 3. It acknowledged the theoretical risk that AI models trained on copyrighted books could "flood the market" with similar works and displace demand for human-authored novels. Id. at 4.

Yet that argument played little role in the outcome -- because plaintiffs failed to make it. Instead, they relied on two theories the court dismissed as "clear losers": that LLAMA could regurgitate small snippets of their works, and that unauthorized use diminished their ability to license books for AI training. The court found both speculative and unpersuasive.

Ultimately, this factor weighed in favor of Meta -- not because market harm was absent in theory, but because the plaintiffs failed to present meaningful evidence of substitution or displacement. The ruling underscores that in future AI copyright litigation, plaintiffs must do more than gesture toward potential harm; they must build a factual record showing actual or likely market interference.

The court also granted Meta's motion for summary judgment on the plaintiffs' Digital Millennium Copyright Act claim, finding no evidence that Meta intentionally removed or altered copyright management information in violation of 17 U.S.C. § 1202.

Summary judgment narrow in scope -- big questions remain

The court did not address plaintiffs' distribution claims, which allege that Meta unlawfully re-uploaded pirated files of the books via BitTorrent during the training dataset acquisition process, as neither side moved for summary judgment on that issue. Id. at 14. That issue -- centered on whether Meta "seeded" or "leeched" copyrighted works during its torrent downloads -- remains open for further discovery and adjudication.

More broadly, the court left open significant questions about the legality of training AI models on copyrighted content. The ruling does not establish that Meta's practices are lawful; it merely reflects the plaintiffs' failure to present the right arguments or evidence. Because the decision was limited to the individual claims of 13 authors and not a certified class, future litigants -- armed with better-developed theories of market harm or different factual records -- may yet shift the legal landscape.

The case offers a warning: Plaintiffs cannot rely on abstract fears or policy rhetoric to defeat fair use -- they must build a concrete, evidence-based case of market harm.

Still, while Meta prevailed, the decision does not close the door to future challenges. The court signaled that a plaintiff who can substantiate a theory of market substitution or economic displacement may prevail, even where use is transformative.