The Emerging Law of AI Training and Fair Use
The proliferation of AI tools among businesses, professionals, and other consumers is a recent and rapidly developing phenomenon. Training these tools on copyrighted material has raised essentially novel questions of copyright law. Before 2025, courts had little precedent addressing whether copying copyrighted works to train an AI model constitutes infringement. Recently, multiple different federal courts have been asked to determine if the answer to this question is yes, or if fair use shields the AI developer from infringement liability. Together, the cases of Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence (D. Del. 2025), Bartz v. Anthropic (N.D. Cal. 2025), and Kadrey v. Meta (N.D. Cal. 2025) sketch the early contours of a doctrine on this topic that is still very much under construction.
The Thomson Reuters case involved Ross Intelligence's use of the plaintiff’s Westlaw headnotes—obtained through "Bulk Memos" prepared by a third-party contractor—to train its AI-powered legal research tool. Ross Intelligence argued that its copying was transformative because the headnotes were used only behind the scenes to train the tool, and were never shown to end users. Applying the U.S. Supreme Court's Warhol framework, the District of Delaware rejected the defendant's fair use arguments. The court pointed out that the defendant had copied the plaintiff's headnotes to build a competing legal research tool. The court concluded that this purpose was commercial, and aimed squarely at creating a market substitute for the plaintiff’s product, which is not a transformative use. Notably, the court emphasized that the fourth fair use factor—the effect of the use upon the potential market for or value of the copyrighted works—is the most important factor, and it weighed decisively against fair use. The decision is now on appeal to the Third Circuit.
In contrast, later in 2025, two Northern District of California opinions pointed to copying protected works to train AI models being fair use—with caveats. The Bartz case involved authors suing Anthropic for copying their books—some lawfully purchased and digitized, and many others downloaded from pirate sites—to train its Claude large language models. The Northern District held that using lawfully acquired books to train large language models was "quintessentially" transformative. The training taught the model to generate new text, not to replicate or supplant the originals, which was akin, in the court's view, to how a human writer learns from reading. Nevertheless, the court also found that downloading pirated copies and retaining them in a permanent, general-purpose digital library was not fair use—exposure that ultimately drove Anthropic to a massive class settlement.
The Kadrey case involved a group of authors suing Meta for training its Llama models on their books obtained from online "shadow libraries." In that case, the Northern District found that using copyrighted books to train a large language model is "highly transformative." The court still stressed that market harm, including the risk that a flood of AI-generated works could dilute the market for the originals, could undermine fair use. But here, the court believed the plaintiffs failed to build a record of such harm. The Northern District concluded that the plaintiffs’ mere speculation that there could be market harm was insufficient to raise a genuine issue of fact and defeat summary judgment.
As of today, cases addressing this topic remain pending in multiple federal courts. Overall, it appears that at least some courts are receptive to viewing the use of copyrighted works to train AI models as transformative under the fair-use doctrine. At the same time, the manner in which copyrighted works are acquired and used remains relevant, and evidence of market harm to copyright holders may weigh against a fair-use defense.
If you have questions regarding copyright fair use, we encourage you to contact our firm.
