Federal Court Rules Against ‘Fair Use’ Defense for AI Training

On Tuesday, February 11, the U.S. District Court for the District of Delaware held in Thomson Reuters Enterprise Centre GmbH et al. v. ROSS Intelligence Inc. that the defendant’s unauthorized use of the plaintiff’s copyrighted materials to train its legal research AI tool was not fair use. This ruling is the first reported case to address whether the use of copyrighted training data for AI training constitutes fair use. Notably, as set forth in Judge Stephanos Bibas’ “Caveat,”¹ this ruling pertains solely to fair use in the context of training non-generative AI. Whether training generative AI tools on copyrighted materials constitutes fair use remains to be addressed in myriad pending lawsuits.² Bibas’ decision reflects the importance courts have historically placed on the fair use prong that evaluates the effect the defendant’s conduct would have on the marketplace for the applicable copyrighted materials. As such, the decision may inevitably be cited both by defendant developers of generative AI platforms that point to the Caveat when asserting a fair use defense and by plaintiffs against developers of generative AI when the usage could affect the market for the copyrighted materials. 

Background: 

ROSS is a legal tech startup that created an AI tool and platform to optimize case law search results based on the user’s queries. Thomson Reuters, the owner of Westlaw, a leading legal research platform, filed a lawsuit against ROSS in 2020, alleging, in part, that ROSS infringed its copyrights in Westlaw headnotes when ROSS used a dataset that copied the headnotes to train its AI tool. For context, Westlaw’s “headnotes” are brief summaries of legal opinions, each comprising a paragraph of text. 

Key Takeaways From the Ruling:

• The use of copyrighted material to train a non-generative AI tool was not considered transformative, even where that copyrighted material was not reproduced in the AI tool’s outputs. Judge Bibas cited the recent Supreme Court case, Andy Warhol Found. for the Visual Arts, Inc. v. Goldsmith, to support his approach in assessing the “broad purpose and character of Ross’s use” of the headnotes to determine whether such use was transformative. Using this approach, he concluded that the broader purpose of ROSS’ use was “to develop a competing legal research tool,” which serves the same broader purpose as Westlaw’s headnote system and is thus not transformative.
• Using copyrighted material to train AI was not seen as an “intermediary step” that qualified as fair use, unlike in past cases dealing with software interoperability. Judge Bibas differentiated ROSS’ use of the headnotes from use discussed in prior cases such as Sony Comput. Ent., Inc. v. Connectix Corp. and Sega Enters. Ltd. v. Accolade, Inc., in which copying computer code was held to be fair use, because in those cases, “[t]he copying was necessary for competitors to innovate.” In those cases, copying computer code was held to be fair use by courts because, without it, computer programmers would be precluded from making programs that could work across different technological platforms, which was widely considered desirable. In contrast, Judge Bibas, quoting Warhol, distinguished ROSS’ use of the headnotes as not “reasonably necessary” to create the AI tool— i.e., copying merely saved ROSS the added expense of manufacturing its own case summaries, rather than precluding its ability to create the AI tool.
• Whether using copyrighted materials to train generative AI tools qualifies as fair use remains an open question. As noted above, Judge Bibas made sure to caveat that this holding applies to a non-generative AI use case. Therefore, it is possible that in a different scenario, training a generative AI tool on copyrighted materials would be considered fair use. However, as it was here, it would still be important to assess whether using the copyrighted materials to train the generative AI tool is transformative in terms of its “broad purpose and character” and whether the training could be found to impinge on the copyright owner’s market for its copyrighted materials and derivative (i.e., related) markets. On this latter point, it remains to be seen whether courts agree with Judge Bibas that the mere existence of a “potential market for AI training data” and the effect copying such data could have should carry significant weight against fair use defenses in the context of generative AI.
  
  

¹ Thomson Reuters Enter. Ctr. GmbH v. ROSS Intel. Inc., No. 1:20-CV-613-SB, 2025 WL 458520 (D. Del. Feb. 11, 2025) (“Because the AI landscape is changing rapidly, I note for readers that only non-generative AI is before me today.”).
² See, e.g., In re Google Generative AI Copyright Litigation, No. 5:23-cv-03440 (N.D. Cal. filed July 11, 2023).