Early last week, the first substantive US ruling on fair use in AI-related copyright litigation, Thomson Reuters v. ROSS Intelligence, No. 1:20-cv-00613 (D. Del.), was issued by Judge Stephanos Bibas. This landmark opinion marks a significant development in AI litigation, particularly concerning the use of copyrighted materials in training AI models.

However, while this decision focuses on the training of an AI model, it does not involve generative AI technology. Hence, it will be important for AI developers and deployers to continue to monitor future decision that address whether the subsequent steps of generating and distributing AI-generated content are considered fair use of the original works.  

Key Takeaways

• The court held that ROSS’s copying of Thomson Reuters’ content to build a competing AI-based legal platform is not fair use under the US Copyright Act.
• The court found actual copying and substantial similarity of 2,243 Westlaw headnotes.
• The court rejected ROSS’s defenses of innocent infringement, copyright misuse, merger, and scenes à faire.
• ROSS’s commercial use weighed heavily against its fair use defense.
• The court vacated its previous denial of summary judgment on the issue of fair use.

Background

Thomson Reuters owns Westlaw, one of the largest legal research platforms in the US. Through a subscription, Westlaw users are able to access a wide range of resources, including case law, state and federal statutes, state and federal regulations, practical guides, news, law review articles, legislative histories, and trial transcripts. A key feature of Westlaw is its headnotes, which summarize the key points of legal opinions. Additionally, West.aw includes the “Key Number System,” which organizes legal opinions.

ROSS Intelligence, a competitor, sought to license Westlaw’s content to develop its own legal AI-based tool. After Thomson Reuters refused, ROSS obtained “Bulk Memos,” created using Westlaw’s headnotes, through a third-party legal services vendor. Thomson Reuters discovered this and sued ROSS for copyright infringement based on its use of Westlaw content to train its AI model.

Overview of the Case

The court granted partial summary judgment to Thomson Reuters on direct copyright infringement, fair use, and other defenses, while denying summary judgment motions from Ross, analyzing the fair use factors. Under the US Copyright Act (17 § USC 107), there are four factors:

(1) the purpose and character of use, including whether the use is of a commercial nature or for nonprofit educational purposes;

(2) the nature of the copyrighted work;

(3) the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and

(4) the likely effect of the use on the potential market for the copyrighted work.

Thomson Reuters prevailed on the first factor. In examining the purpose and character of ROSS’s use, the court focused on whether the use was commercial and transformative. ROSS acknowledged that its use was commercial but argued that it was transformative, as the headnotes in question were allegedly “transformed” into numerical data representing the relationship among legal words in its AI system. The court disagreed, noting that ROSS’s use did not have a further purpose or different character from Thomson Reuters’ use. ROSS also argued that its use was permissible under the doctrine of “intermediate copying,” but again the court disagreed, noting that the cited cases were inapt because they involved copying of computer code rather than written words and the code copying was necessary for competitors to innovate. By contrast, use of the headnotes was not necessary to achieve ROSS’s desired purpose.

The court resolved factor two in ROSS’s favor, finding that although Westlaw’s material has the minimal required originality, it is not highly creative. Further, while the headnotes involve some editorial creativity, the Key Number System is a factual compilation with limited creativity.

The court also ruled in favor of ROSS on factor three, despite the number of headnotes used, because the material available to the public did not include the Westlaw headnotes. According to the court, what matters is not the amount and substantiality of the portion used in making a copy, but rather the amount and substantiality of what is thereby made accessible to the public for which it may serve as a competing substitute. It determined there was no factual dispute, as ROSS’s output did not include Westlaw headnotes.

Finally—and most important in this case—the court emphasized that because ROSS could have developed its own product without infringing Thomson Reuters’ copyrights, the fourth factor weighs in favor of Thomson Reuters. The court examined the likely effect of ROSS’s copying on the market for Westlaw’s product and, while initially considering whether ROSS’s use served a different purpose by creating a new research platform, ultimately concluded that ROSS intended to compete with Westlaw, and failed to prove otherwise. Court have differed over the years on whether the first or the fourth fair use factor is the most important; here analysis of the fourth is crucial (but both favored Thomson Reuters).

Looking Ahead

The implications of this decision for AI copyright litigation and fair use arguments are significant.

First, many practitioners have been waiting for a decision whether creating an AI model is considered transformative and fair use, particularly since AI models store their intelligence as numerical weights that are updated during the training process. But even with such advanced technology, the court in this case declined to hold that the use was transformative, based largely on its ultimate purpose of competing with the owner of the original works.

Also noteworthy is that the court declined to find fair use for an AI technology that is not generative AI. Even though the output from ROSS’s AI system was uncopyrighted verbatim quotes from court opinions (and not the original copyrighted headnotes), there was no fair use.

This could have broader consequences for large language models (LLMs) and generative AI technologies. When judges in other pending generative AI cases consider both the training step (as in ROSS) and the output generation step for a generative AI technology (e.g., an AI-generated image), it could be even less likely that fair use will apply. Here, the court emphasized that “factor four is undoubtedly the single most important element of fair use.” So, if AI-generated content, including that produced by LLMs, is substantially similar to an original work and has a detrimental effect on the market for the original work (e.g., puts an artist out of business), a finding of fair use may be less likely.

Our “fair warning” is this: AI developers and deployers should continue to monitor ongoing AI litigation, while considering the market implications of the use of copyrighted materials for training AI models or distributing AI-generated output. Because fair use is heavily dependent upon the facts, we anticipate different rulings from different courts, particularly where the commercial use of the original content is not a clear-cut as here.