The Questions AI Copyright Cases Need to Answer

C opyright law has long aimed to protect creators’ rights and incentives while also giving people access to knowledge and creative works. The U.S. Constitution laid the foundation for this legal protection: copyright exists to “promote the Progress of Science and useful Arts” by giving authors exclusive rights to their works for a limited time.

Generative artificial intelligence (AI) is testing how those principles apply to a new technology. AI models are increasingly powerful tools that can create text, images, and other content. However, they rely on existing material for training, much of which is copyrighted. As AI advances, creators and AI companies are increasingly facing legal battles over how that material can be used. Creators argue that companies are using their work without permission to build commercial AI systems, while AI companies argue that training models on copyrighted works constitutes transformative fair use under existing copyright principles.

But recent AI copyright cases are adding more complexity to these legal questions. The issue is no longer simply whether copyrighted works can be used to train AI; courts are looking more closely at how those works are acquired and used, what the models produce, and whether those products compete with — or potentially extract value from — the works that made them possible.

The legal doctrine of fair use allows certain uses of copyrighted material without permission. When deciding whether a use qualifies, courts weigh four factors: the purpose of the use and whether it adds something new or “transformative”; what kind of work it is; how much of the work was used; and whether the use affects the existing or potential market for the copyrighted work. While these factors are well established, AI is pushing them into unfamiliar territory. As AI companies use copyrighted material to train their models, courts are now applying these long-standing principles to questions that copyright law did not anticipate.

In Bartz v. Anthropic, authors sued over the company’s use of millions of copyrighted books to train Claude. The resulting case, for which the Taxpayers Protection Alliance filed an amicus brief, was a pivotal test for the future of AI. While Anthropic bought and scanned some of those books, it also downloaded millions of pirated copies from LibGen and PiLiMi and stored them in a central library. In June 2025, the court drew an important distinction for the copyright and AI conversation: Using the books to train Claude was fair use because the training was highly transformative — the books helped teach the model to generate something new — regardless of whether the copies came from purchased or pirated sources. But the court treated Anthropic’s creation and retention of a permanent library of pirated books as a separate use that was not protected by fair use.

The Third Circuit’s recent ruling in Thomson Reuters v. Ross shows what happens when AI is used to build a product that competes with the copyright holder. Ross Intelligence developed an AI-powered legal-research tool using training materials derived from thousands of Westlaw headnotes, 2,243 of which were at issue in the court’s ruling. The Third Circuit recently affirmed the district court ruling against Ross, finding that the use was highly commercial and its purpose was very close to the original purpose of the headnotes. A large amount of material was copied, and the result was a direct competitor in Westlaw’s own market. The court also found that Reuters’s headnotes had enough originality to qualify for copyright protection in the first place.

These cases point toward a broader distinction that often gets lost in the AI copyright conversation: How a model is trained is a separate question from what it produces. The first question is whether an AI company can use copyrighted material and how it obtains the material to train a model. The second is what happens when the model generates an output — whether it infringes copyright or affects the market for existing works. Although Ross did not involve generative AI, the case shows why the purpose of an AI tool and its effect on the market for copyrighted works are important. Even if training qualifies as fair use, an output can still reproduce protected expression or affect the market for an existing work. Although competition alone does not establish infringement, its effects on the market are relevant to the fair use analysis. This distinction matters because copyright law aims not only to protect creative works, but also to preserve the economic incentives that encourage people to create.

The AI copyright conversation still has more questions than answers, and courts will keep working out how fair use applies as the technology evolves. Copyright should not make it unnecessarily difficult to build new technologies, but AI companies should still be held accountable when their outputs infringe copyright. The challenge is to apply existing copyright principles without losing sight of what copyright is ultimately meant to achieve: encouraging creation while promoting access and innovation.