ROSS AI Decision Gives Early Indication of Strengths and Weaknesses of Fair Use Defense
On Tuesday, a Delaware federal district court granted partial summary judgment to Thomson Reuters Enterprise Centre GmbH (“Thomson Reuters”) in its copyright litigation against ROSS Intelligence (“ROSS”).1 The lawsuit, which was filed by Thomson Reuters in 2020, alleges that ROSS utilized copyrighted content from Thomson Reuters’ Westlaw database in order to create its artificial intelligence (“AI”) legal research tool. This week’s decision, which revises a prior 2023 summary judgment opinion and order from the same court, grants summary judgment on certain of Thomson Reuters’ direct infringement claims, and most notably rejects ROSS’ assertion that the fair use defense applied to ROSS’ use of the copyrighted material.2
The litigation centers around ROSS’s use of Thomson Reuters’ Westlaw headnotes, which it utilized in the process of training its AI legal research tool.3 ROSS’s AI model is not designed to provide a generative response, but is instead designed to provide a list of judicial opinions responsive to a legal-question prompt. The model was trained with a series of legal questions and answers that were based on material identical or similar to numerous Westlaw headnotes.
In evaluating whether ROSS directly infringed on Thomson Reuters’ copyright, the court held that (1) both the headnotes and the key numbering system are original; and (2) at least 2,243 of the headnotes at issue were actually copied and substantially similar. The court summarily dismissed ROSS’s innocent infringement, copyright misuse, merger, and scènes à faire defenses.
Developers of AI models will likely find most interesting the decision’s detailed analysis of each of the four fair use factors. While the court concluded the factors were evenly split between the parties—with factors one and four favoring Thomson Reuters and factors two and three favoring ROSS—the more heavily weighted factors, one and four, supported an overall finding in favor of Thomson Reuters.
The court found that the first fair use factor—which considers the purpose and character of the use, including whether the use is commercial in nature—favored Thomson Reuters. Unsurprisingly, the court concluded that ROSS’s development of a for-profit AI legal research platform was for commercial use. The court also concluded that ROSS’s use was not transformative, emphasizing the fact that the ROSS AI model is not generative AI, and that it instead “spits back relevant judicial opinions that have already been written,” a process which “resembles how Westlaw uses headnotes and key numbers to return a list of cases with fitting headnotes.”4 A question remains as to whether generative AI models that provide substantive responses based upon training data would provide a transformed output that distinguishes the analysis in this case, and the Court took pains to make this distinction.
Noting that ROSS used the copyrighted headnotes in an “intermediate” step in the end product, the court’s transformative analysis focused on the fact that the copyrighted material did not appear in the final product—a list of cases—and differentiated the instant case from cases involving the intermediate copying of source code, which, for example, courts have ruled is, or may be, necessary to use the functional aspects of the code. This intermediate step analysis—and the manner in which the court distinguished it from the existing line of computer programming cases—may be a point of focus of pending cases regarding the use of copyrighted material in the training of AI models.
The court determined that the second fair use factor, which considers the nature of the copyrighted work, favored ROSS. Although the court found that the headnotes had “more than the minimal spark of originality required for copyright validity,” it emphasized that “the material is not that creative” (emphasis in original) and that the headnotes “have creative elements but are far from the most creative works.”5 Notably, many other pending training data cases involve works likely to be deemed at the heart of copyright law, such as news articles and photographs.
The court found that the third factor, which considers how much of the copyrighted work was used and how substantial a part it is relative to the whole, also favored ROSS. The output of the AI model does not reproduce the copyrighted material. This holding may not provide insight into the analysis of an output from a generative AI model, which content owners have asserted has the potential to reproduce portions of copyrighted content.
The court found that the fourth fair use factor, which considers the effect on the value or potential market for the copyrighted work, favored Thomson Reuters, and the court weighed this factor heavily, calling it “undoubtedly the single most important element of fair use.”6 The court identified ROSS’s AI tool as potential competition in the market for legal research platforms, also noting its potential to compete in the market for licensing data for training of other AI models. While many AI models serve a clearly different purpose than the data and materials used to train them, and the model owners can assert that they will not compete directly with content owners, the latter point—regarding the market for licensing data—could be more widely applicable, as each content owner could assert value in the data used to train the AI model. If the potential market analysis articulated here is accepted as articulated by other courts, it may be challenging for AI models to distinguish this case.
While distinguishable in several key ways from many of the currently pending copyright infringement actions involving AI models, this case is the first of its kind to rule against a fair-use defense for the AI model. As such, the decision is likely to be cited frequently in pending and future actions.
1 Thomson Reuters Enterprise Centre GmbH and West Publishing Corp. v. ROSS Intelligence Inc., 1:20-cv-00613-SB, ECF 770 (“Order”).
2 Order at 2.
3 Order at 3
4 Order at 17.
5 Order at 20.
6 Order at 21.
