Fair Use Defense Failed in Thomson Reuters v. Ross, Jury Still Out for Generative AI

The first substantive decision on the fair use defense in an artificial intelligence (AI) copyright case came down against the defendant, who used AI to create a competing product.

However, as the decision expressly limited its ruling to nongenerative AI, it remains unclear how courts may apply fair use to generative AI tools.

On February 11, 2025, the U.S. District Court for the District of Delaware reversed course from its 2023 ruling and granted Thomson Reuters’ motion for summary judgment on direct copyright infringement and fair use. In Thomson Reuters Enter. Ctr. GmbH v. Ross Intel. Inc., the court found that Westlaw’s headnotes exhibited enough of a “creative spark” to be considered “original,” “copyrightable” works and that Ross’ use of Westlaw headnotes to create a competing legal tool did not constitute a “transformative” fair use under the standard set forth in the recent U.S. Supreme Court decision in the Warhol case. Ross recently asked the court to pause the dispute (which had been heading for jury trial in May) and allow it to quickly appeal the fair use decision to the U.S. Court of Appeals for the Third Circuit.

Thomson Reuters alleged that Ross directly infringed its copyrights by using Westlaw’s headnotes and Key Number System to train Ross’ AI legal search tool. More specifically, after being denied a license to use Westlaw’s legal research database for AI training purposes, Ross instead licensed 25,000 “Bulk Memos” from a third party to train its AI legal search tool. These Bulk Memos were “compilations of legal questions” created by lawyers who were instructed by the third party to use (though not just copy and paste) Westlaw headnotes to generate their questions. In revising its 2023 opinion, the court now saw no genuine dispute as to whether Westlaw’s headnotes and Key Number System were “original enough” to be copyrightable or whether Ross’ action constituted fair use. In the court’s eyes, “Ross built its competing product using Bulk Memos, which in turn were built from Westlaw headnotes.”

Copyrightability of Westlaw’s Headnotes and Key Number System 

In reexamining Thomson Reuters’ direct infringement claim, the court took a closer look at the issue of copyrightability when it comes to the Westlaw headnotes and Key Number System. Distinguishing the copyrightability of judicial opinions from the copyrightability of the Westlaw headnotes, the Key Number System, and the Bulk Memo questions, the court noted that, pursuant to well-established caselaw, the text of judicial opinions was not copyrightable. However, it found that the headnotes were copyrightable, both as a “compilation” and as individual works, and that the Key Number System was copyrightable as a “compilation.” 

In explaining why he changed his mind, Judge Stephanos Bibas admitted to realizing that the originality threshold under Feist is “extremely low,” and the case is “not about how much effort went into developing” the work but about finding “some minimal degree of creativity … some creative spark.” Seeing that Thomson Reuters’ “selection and arrangement of its headnotes” easily met the “minimal degree of creativity” needed to clear the originality threshold, the court found the headnotes to be copyrightable as “compilations.”

Next, the court turned to the copyrightability of each headnote as “an individual, copyrightable work,” recognizing their “original value as individual works.” In explaining his “change of heart,” the court elevated the practice of law to an artform, comparing a judicial opinion to a “block of raw marble” and the lawyer’s editorial judgment to that of a “sculptor” who “creates a sculpture by choosing what to cut away and what to leave in place.” Judge Bibas admitted to having gotten it wrong when he saw “the degree of overlap between the headnote text and the case opinion text as dispositive of originality.” Correcting himself here, Judge Bibas went as far as characterizing “even a headnote taken verbatim from an opinion” as an “editorial expression” with “enough ‘creative spark’ to be original” because it expresses the editor’s idea about what is important from the opinion. 

On the Key Number System, Westlaw’s “taxonomy to organize its materials,” the court found it to have the “minimum ‘spark’ of originality” to be considered copyrightable, noting that using rote computer programs or tracking common law school courses was not dispositive on whether the system was original. After all, the Key Number System need not be “novel,” just “independently created” by Thomson Reuters, to be original. 

Ross’ Fair Use Defense 

After holding that the Bulk Memo questions were substantially similar to 2,243 Westlaw headnotes, the court granted summary judgment to Thomson Reuters on direct infringement as to those copyrighted works. The court then summarily rejected Ross’ other defenses of innocent infringement, copyright misuse, and merger before turning to Ross’ fair use defense and vacating its 2023 denial of summary judgment on this issue. The court evaluated each of the four factors that are considered in evaluating fair use and found for Thomson Reuters what it considers the two most important factors: the purpose and character of the use and the effect on the market for the work. As a result, it concluded that Ross’ use did not constitute fair use.

• Purpose and character of use. In analyzing the first fair use factor, the purpose and character of Ross’ use, the court looked to whether Ross’ use was commercial and whether it was transformative. In evaluating transformativeness, the court found that “Ross’s use is not transformative because it does not have a ‘further purpose or different character’ from Thomson Reuters’s,” as Ross used the headnotes to create a competitive legal research tool. The court then discounted as “inapt” the applicability of prior cases which found fair use in connection with intermediate copying (where, as here, a temporary copy was made as a step in the process and did not appear in the final product). In those cases, the intermediate copy was made in connection with reverse engineering video games, and the court distinguished those cases by saying they “(1) are computer-programming copying cases; and (2) depend in part on the need to copy to reach the underlying ideas.” The court went on to note that the fair use considerations related to computer programs—which “almost always serve functional purposes”—differ from those that apply to written words. Here, the court said the intermediate copying “fits more neatly into the newer framework advanced by Warhol.” The court looked to the broad purpose and character of Ross’ use and concluded that “Ross took the headnotes to make it easier to develop a competing legal research tool. So Ross’s use is not transformative.” Careful not to send too strong of a signal given the many pending cases involving generative AI, the court clarified that Ross’ technology here relates only to nongenerative AI.
• Nature of copyrighted work. In awarding the second fair use factor, the nature of the copyrighted work, to Ross, the court noted that the headnotes and the Key Number System are “not that creative” and “less than that of a novelist or artist drafting a work from scratch.”
• Amount and substantiality of portion used. Turning then to the third fair use factor, the amount and substantiality of the portion used, the court found this favors Ross based on the fact that the output to the user was a judicial opinion, not a Westlaw headnote, and further because “[w]hat 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 a public for which it may serve as a competing substitute.’”
• Effect on potential market. In analyzing the fourth fair use factor, the likely effect on the market for the original work, the court noted that the impact was not only the obvious market of legal research platforms, but other potential derivative markets, pointing specifically to “data to train legal AIs” as an example. The court distinguished the “public benefits” analysis in this case from cases involving software such as application programming interfaces (APIs), where copying of APIs by developers and users was allowed because they were used to using them. While allowing that there may be a public interest in accessing the law, the court distinguished the case here as judicial opinions are freely available, and “[t]he public has no right to Thomson Reuters’s parsing of the law.” Ultimately, the court concluded that “[t]here is nothing that Thomson Reuters created that Ross could not have created for itself or hired [a third party] to create for it without infringing Thomson Reuters’s copyrights.”
• Balancing of fair use factors. In balancing the four factors, the court concluded the most important factors favored Thomson Reuters and granted summary judgment to Thomson Reuters on fair use.

Effect on Generative AI 

The fair use arguments made by the parties in the Thomson Reuters case echo many of the arguments being made in the pending copyright infringement cases against generative AI tool providers, but the court was careful to note this is not a generative AI case. It is difficult to know what effect, if any, this case may have on generative AI, and opinions on this have been split.

Some have argued that the Ross case is distinguishable from generative AI in that in Ross, you have one legal tech provider copying another legal tech provider’s copyrighted material to create a competing legal tech product (which the court found was not transformative since it served the same purpose and was used to create a competing substitute). They contrast this from the copying of training data to create generative AI tools, arguing that the purpose of such copying is to create the tools, not to create competing content, and that tools themselves don’t compete with the training data. 

Others, however, particularly in the creative community, have argued that regardless of who creates the output, generative AI tools will ultimately displace them by facilitating the creation of competing works, and the relevant market to consider when looking at purpose and competitive effect is not just the market of selling the creative works, but the derivative market of licensing these works to train someone else’s AI.

It will be interesting to see the impact, if any, that the Ross case may have on the generative AI cases that are moving their way through the courts. 

Key Takeaways 

• The court finds case headnotes can be protectable under copyright law even if they copy case text verbatim because they express the editor’s idea about what information from the opinion is important.
• The decision holds that fair use does not apply to intermediate copies made to train AI tools when used to create a competing product that serves the same purpose as the training materials.
• Cases finding fair use for intermediate copying in connection with reverse engineering video games were found to be inapt here because copying computer code is different (because it is functional) and because in those cases, the copying was necessary to reach the underlying ideas and for the competitors to innovate.
• The case is limited to nongenerative AI, and it is not clear what effect, if any, this case may have on pending generative AI cases.