On May 1, 2025, a federal courtroom in San Francisco became ground zero for one of the most consequential copyright hearings in recent memory. The three hour hearing in Kadrey v. Meta Platforms marked the first major judicial test of whether using copyrighted works to train generative AI systems—like Meta’s LLaMA models—qualifies as “fair use” under U.S. copyright law. While Judge Vince Chhabria has yet to issue a formal decision, his comments during oral argument offer critical clues about how courts may approach this issue going forward.

Fair use, codified in 17 U.S.C. § 107, allows limited, unauthorized use of copyrighted material without permission from the rights holder under certain conditions. Courts consider four factors: (1) the purpose and character of the use (including whether it is for a commercial purpose or transformative), (2) the nature of the copyrighted work, (3) the amount and substantiality of the portion used, and (4) the effect of the use on the value or market, or potential market, for the original work. Factor (4) is most important, with courts viewing Factor (1) as the second most important factor.

In this case, 13 authors—including Richard Kadrey and Sarah Silverman—sued Meta for allegedly using pirated copies of their novels from online shadow libraries like Books3 and LibGen to train LLaMA. The plaintiffs argue that this use was both unlicensed and exploitative. They emphasize that novels, rich in narrative and linguistic complexity, were chosen precisely because of their expressive value, which Meta recognized was highly desirable for training its AI models. Meta, they claim, profited from the use of these copyrighted works without compensation and, crucially, without transforming them in any meaningful way.

Meta counters that its use of the books is transformative because the LLaMA model does not reproduce the novels but instead learns from statistical patterns in their language to generate new content. According to Meta, the use is akin to studying, not copying. Furthermore, Meta argued that full copies are technically necessary to train these models, and that no viable market exists for licensing individual works for such a use, as there are no clearing houses for such rights. The cost of licensing works from individual authors would also far outstrip the value of those individual works in training an AI model. The fact that the training materials were allegedly pirated, they claim, is irrelevant to the fair use analysis.

At the hearing, Judge Chhabria acknowledged the highly fact-dependent nature of fair use determinations and signaled that he saw Meta’s use as transformative under the first fair use factor—a significant potential win for the defense. However, he zeroed in on the fourth factor: market harm. He questioned whether generative AIs trained on copyrighted works could create an "endless stream of competitors" that might “obliterate” the demand for the original books. This, he suggested, might be the most important issue in the case.

Notably, this concern had not been the central focus of either party’s summary judgment briefs, potentially leaving the plaintiffs without a strong evidentiary record on that point. If the court finds that plaintiffs failed to show market harm at this stage, it could rule in Meta’s favor. But if Judge Chhabria sees unresolved factual questions—especially about whether generative AI might saturate the market with derivative or substitutive works—he may deny summary judgment and allow the case to proceed to trial. The Judge emphasized more than once that only a ”potential” effect on the market for the copyrighted works is necessary to find that Factor (4) favors the plaintiffs. Proof that the AI models are already harming the market for those works is not required.

If the court's written ruling follows the logic laid out at the hearing, the implications could be significant. A decision emphasizing market impact based on the particular plaintiff could open the door to more nuanced, case-specific outcomes across different types of copyrighted content—books, news articles, visual art, etc.—rather than a one-size-fits-all doctrine for AI training. The Court suggested the legal analysis might differ between different types of book authors or even different individual authors. If the Judge adopts the views expressed at the May 1 hearing, it would also raise the stakes for future litigants to develop robust evidence of actual or likely market displacement caused by AI systems trained on their works.

The Kadrey case could become a bellwether for how U.S. courts balance innovation and intellectual property in the AI age. For now, all eyes are on Judge Chhabria’s forthcoming written decision. Please see this blog again in the near future for an analysis of that order once it issues.