Whether using copyrighted works to train generative AI models qualifies as fair use is one of the most pressing questions in contemporary copyright law. Courts are only beginning to confront the issue, and while a handful of decisions have touched on it, none have definitively answered it in the generative AI context. One of the first rulings, in Thomson Reuters v. Ross (Case No. 1:20-cv-613-SB, 2025 WL 458520 (D. Del. Feb. 11, 2025)), found infringement where an AI company copied headnotes and editorial enhancements from Westlaw to build a legal research tool. Central to that decision was the court’s conclusion that the AI product would compete directly with Thomson’s offering, thereby weighing heavily against fair use under the market harm factor.

Now, in Kadrey v. Meta, Case No. 3:23-cv-03417-VC (N.D. Cal.), we may get another piece of the puzzle. Here the plaintiffs allege that books obtained through “shadow libraries” were used to train the AI models, without authorization or compensation. The defendant does not deny using the works but contends its conduct is protected fair use.

The judge has signaled that the fourth statutory factor, i.e., whether the use affects the potential market for the original work, may be determinative, noting at one point during a recent hearing on cross-motions for summary judgment that “if you’re obliterating the market for someone’s work and you don’t even pay a license, I just don’t understand how that can be fair use.” But he also expressed doubt that the plaintiffs could present sufficient evidence to show their works would actually be displaced by AI-generated outputs, asking whether they were merely inviting the court to speculate about lost sales.

This points to a broader tension in the fair use framework. Although the burden of proving fair use ultimately lies with the defendant, courts have often required plaintiffs to show actual or likely market harm. The Supreme Court, in Harper & Row v. Nation Enterprises, 471 U.S. 539 (1985), emphasized that the fourth factor is often the most important, but it must rest on more than conjecture. In the generative AI context where models can produce novel, synthetic outputs not tied to any one work, demonstrating substitution or market erosion may be especially challenging.

At the same time, the Thomson Reuters decision shows that courts seriously consider the risk of functional or commercial displacement. If the output of an AI system competes with the original, the market effect prong may weigh heavily against fair use.

Kadrey v. Meta is likely to shape how courts assess this question moving forward. As more litigation emerges over AI training practices, the balance of proof on market effect could determine whether fair use offers a safe harbor for AI systems training on copyrighted works at scale.