On February 11, 2025, the U.S. District Court for the District of Delaware became the first to rule on whether the use of copyrighted materials to train an AI system qualifies as copyright infringement. In Thomson-Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., the court granted summary judgment on the plaintiff’s infringement claim, and denied the defendant’s fair-use defense.

Key Points:

• • Use of the plaintiff’s copyrighted material to train the defendant’s AI system, which competed directly with the plaintiff’s legal research tool, was not fair use, because it was not “transformative”.
  • The defendant’s legal research AI tool, which competed directly with the plaintiff’s legal research tool, was not a generative AI model, since searches returned case cites, as well as language from actual case opinions.
  • The court’s opinion did not consider or rule on the use of copyrighted material to train generative AI models, which produce new content in response to user prompts.

Key Remaining Question:

• • Does unauthorized use of copyrighted material to train generative AI models constitute transformative fair use?

Background: The court found that Thomson-Reuters owns Westlaw, a legal research platform. The platform features editorial content and annotations, such as headnotes, which summarize case holdings and key points of law according to Westlaw. The parties agreed that Ross Intelligence is a competitor of Thomson-Reuters, that Ross made a deal with a third party (LegalEase) for data that Ross would use to train an AI system that would compete with Westlaw, and that the third party used Westlaw’s headnotes to prepare the AI training data for Ross. Thomson-Reuters established that it owns copyrights on many facets of the platform, including the headnotes. Thomson-Reuters sued Ross for copyright infringement. In addition to denying infringement, Ross raised several defenses, including fair use.

Court’s Analysis and Holding: After finding that LegalEase had copied many of Westlaw’s headnotes to prepare the training data, the court found that Ross committed copyright infringement because it used that data. It dismissed Ross’s innocent infringement, copyright misuse, merger, and scènes à faire defenses; then focused on Ross’s fair-use defense. Although two of the four relevant fair-use factors (nature of the copyrighted work, and amount and substantiality of the work copied by the defendant) favored Ross, the court ruled the remaining two factors favored Thomson-Reuters, and thus rejected the fair-use defense.

“Transformative” Analysis. The first fair-use factor, and most extensively analyzed by the court, considers the purpose and character of the defendant’s use of the copyrighted works. The court discussed whether copying of Westlaw headnotes to prepare AI training data was commercial, and whether that use was “transformative.” Because Ross used the headnotes to develop a competing legal research product, its use was clearly commercial.

According to the Supreme Court’s recent decision in the Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 143 S. Ct. 1258 (2023) case, a use is transformative if it has “a further purpose or different character” from the copied work. Ross highlighted the fact that no headnotes appeared in results produced by its AI research tool, and added that previous cases had found transformative fair use based on such “intermediate” copying. Although the court acknowledged the relevance of intermediate copying, it distinguished those cases on the ground that they all related to copying of computer software, which is different from other literary works in that “such programs almost always serve functional purposes.” The court also noted a related point; that, in the software context, intermediate copying is often necessary to gain access to the functional elements of the computer program, such as code required for compatibility with an add-on software tool. As a result, the court was not persuaded by Ross’s intermediate copying arguments.

Because Ross was using the headnotes simply to create a competitive legal research tool, the court decided the use had no further purpose or different character from Westlaw’s headnotes, and therefore was not transformative. However, the court carefully limited its holding by adding that “[b]ecause the AI landscape is changing rapidly, I note for readers that only non-generative AI is before me today.”

Commercial Impact Analysis

The final fair-use factor, often treated as the most important, considers the impact of uncontrolled copying of the type made by the defendant on actual and potential commercial markets for the plaintiff’s copyrighted works. Here, the court found that unauthorized use of Westlaw’s headnotes would have a negative impact on both existing and potential derivative markets for the headnotes. The current market identified by the court was as a competitive legal research tool, and a clear derivative market was licensing of the headnotes to train legal AI tools.

Conclusions

Since this case involved the use of copyrighted material to train an AI system, it could have ramifications for currently pending cases considering other AI systems. Those involve major content creators and groups that hold rights on their behalf, on the one hand; and AI system developers on the other. However, because those generative AI systems all create new content, rather than simply return existing legal case names and opinion text, this opinion provides little guidance for the central issue in those disputes. Specifically, the question of the transformative nature of intermediate copying of copyrighted material to train a generative AI model remains unanswered.

This article summarizes aspects of the law and does not constitute legal advice. For legal advice with regard to your situation, you should contact an attorney.