Third Circuit affirms copyright in Westlaw headnotes and rejects fair use defense

Key takeaways

• The U.S. Court of Appeals for the Third Circuit has affirmed that Westlaw headnotes are original works protected by copyright. The court concluded in Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc. that there was no fair use, with three of the four statutory factors weighing against the defense.
• The court called the dispute “no more than an ordinary copyright case” and distinguished it from generative AI training cases.
• ROSS Intelligence’s use of the headnotes to train its AI legal search tool was held to be “minimally transformative, at best,” because ROSS used them for the same ultimate purpose Thomson Reuters does: building a legal research platform.
• The court also found sufficient market harm, crediting harm to the value of the headnotes and to the legal research market. It also counted harm to the market for licensing headnotes as AI training data, unlike two other district court decisions, discussed below, that addressed fair use in generative AI training.

On September 29, 2026, the Third Circuit issued its decision in Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., affirming the U.S. District Court for the District of Delaware’s February 2025 grant of partial summary judgment to Thomson Reuters. The interlocutory appeal presented two questions: whether Westlaw’s headnotes and Key Number System are original as a matter of law and whether ROSS’s use of the headnotes was fair use. Writing for a unanimous panel, Judge Tamika Montgomery-Reeves answered yes to the first and no to the second, in the first federal appellate decision to address fair use related to the training of an AI system.

Under ROSS’s framing, the court said, the case “appears to concern the future of AI legal technology,” but “appearances can be deceiving.” Instead, the court found the case to be “no more than an ordinary copyright case.” This Legal Development follows our prior post on the district court’s decision, sets the Third Circuit’s reasoning beside the generative AI training decisions from the U.S. District Court for the Northern District of California, and considers how the ruling may affect AI developers and content owners.

Background

ROSS built an AI legal search engine that answered plain-language legal questions by returning relevant passages from judicial opinions. It was not generative AI; as the court noted, it “would not create any new expression.” Rather, it would only return text passages from preexisting judicial opinions. To train the search engine, ROSS engaged LegalEase Solutions to write roughly 25,000 training memos, each pairing a legal question with graded judicial opinion passages, and LegalEase’s drafters used Westlaw headnotes to write the questions because they offered “an easy way” to frame them. ROSS marketed its platform as a substitute for Westlaw at comparable prices, and some law firms switched. On appeal, ROSS did not dispute the copying.

Headnotes are copyrightable

The court applied the Feist standard, under which originality requires only a “modicum of creativity.” This is an “extremely low” bar that is met so long as a work evidences at least “some creative spark.” All 2,243 headnotes were found to clear that bar because Westlaw’s editors decided which points of law were important enough to include and how to word them. The court found support in Callaghan v. Myers, an 1888 U.S. Supreme Court decision recognizing a reporter’s copyright in headnotes, as reaffirmed in Georgia v. Public.Resource.Org in 2020.

The court rejected ROSS’s three counterarguments. First, it disagreed that recognizing copyright in the headnotes would give Thomson Reuters “a monopoly over the law,” stating that “[h]eadnotes are not law; judicial opinions are.” It also held that the merger doctrine, which denies copyright where an idea can be expressed in only one or a few ways, does not apply, because “there are many ways to express points of law from judicial opinions.” Finally, it concluded that an earlier U.S. Court of Appeals for the Second Circuit precedent that denied copyright protection for the arrangement of party names and parallel citations does not help ROSS because the headnotes were “independently composed.”

The court did not decide whether headnotes that quote an opinion verbatim are original, because none of the 2,243 headnotes before it did, and it treated the originality of the Key Number System as forfeited because ROSS never briefed it.

Fair use

After resolving copyrightability, the court turned to ROSS’s fair use defense and analyzed the four statutory fair use factors.

Factor 1: Purpose and character of the use

The first fair use factor looks at both whether a use is commercial and whether it is transformative (i.e., “whether the new work merely supersedes the objects of the original creation . . . or instead adds something new, with a further purpose or different character.” Andy Warhol Foundation v. Goldsmith (quoting Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569, 579 (1994))). ROSS’s use was found to be commercial: It built a legal research platform priced comparably to Westlaw to compete for Westlaw’s customers. The court then considered whether the use was transformative. Thomson Reuters uses headnotes to help researchers find judicial opinions; ROSS used them to build a platform that helps users find responsive opinion passages. While the court noted that training an AI program was an intermediate step that “arguably presents a slight degree of difference in use,” it ultimately focused on the fact that both parties used the headnotes “to create and optimize a legal-research platform.” As a result, the court held that ROSS’s use “shares the same ultimate purpose” and is “minimally transformative, at best.” Because the court found that ROSS’s use was highly commercial and minimally transformative, it concluded that the first factor weighed against fair use.

The court distinguished Authors Guild v. Google, where the Second Circuit found that book scanning served a search function that was different from the original purpose of the books (which was to be read). The Third Circuit did not find such a transformative purpose in this case, because it found that both Thomson Reuters and ROSS used the headnotes to help users find judicial opinions related to their legal research. The court also distinguished three cases that involved the intermediate copying of computer code, Google v. Oracle, Sega v. Accolade, and Sony v. Connectix, because in those cases, it found the copying was necessary to reach unprotected functional elements. By contrast, it said that ROSS did not need the headnotes to reach the law; it had the opinions and could have built its memos from them. “Unlike necessity,” the court wrote, “ease is not a justification for copying.” In a footnote, the court added that, to the extent good faith remains relevant, the undisputed evidence showed ROSS “at times acted in bad faith” in how it obtained access to Westlaw.

Factor 2: Nature of the copyrighted work

Because the headnotes were published before ROSS used them and because the court concluded that the headnotes are “more factual than fictional” (because they must accurately convey the law), the court held that this factor favored fair use, though only slightly. However, the court noted that this factor has rarely played a significant role.

Factor 3: Amount and substantiality of the portion used

Here, the Third Circuit parted from the district court, which had given this factor to ROSS because ROSS’s output never showed users a headnote. It instead asked whether “no more was taken than necessary” for the copier’s purpose. The court held that ROSS copied entire headnotes without a transformative purpose that would have made such extensive copying reasonable. That ROSS copied only about 0.08% of Westlaw’s headnotes did not help, because each headnote is itself a complete copyrighted work. This factor was found to weigh against fair use.

Factor 4: Effect on the potential market

ROSS argued that the relevant market was a standalone market for headnotes, which it claimed did not exist. Relying on its earlier decision in Video Pipeline v. Buena Vista Home Entertainment, the court held that the statute directs attention to the value of the work, not only to a narrowly defined market. Thomson Reuters markets headnotes as a feature to attract users to Westlaw, and the court found that, by copying the headnotes and using them to build its own competing legal-research platform, ROSS diminished their value as a draw. The court also accepted the legal research platform market as a relevant original market that unrestricted copying of headnotes to build substitute tools would harm.

Significantly, the court accepted Thomson Reuters’s argument that ROSS’s actions harmed the market for licensing its headnotes as AI training data, noting evidence that this market “is rapidly developing” and that Thomson Reuters uses its headnotes to train its own AI search products. The fact that Thomson Reuters had not yet licensed headnotes to others did not disprove the market’s existence, according to the court. The court gave little weight to ROSS’s arguments that its copying increased public access to the law or that the ruling would impede AI development or national security. The court concluded that the fourth factor weighed against fair use.

Weighing the factors together, the court held that ROSS’s use was not fair.

How the decision compares with the generative AI rulings

The Third Circuit emphasizes that this case does not involve generative AI, and it sets its decision apart from the generative AI cases in a footnote responding to the U.S. Department of Justice’s September 1, 2026, statement of interest in In re: OpenAI. That statement of interest relied on Bartz v. Anthropic, a case involving generative AI, to argue that training a large language model that can generate original responses is transformative and does not result in “substitutive competition.” The court states that those concerns “do not apply here,” noting that “unlike the AI models in Bartz and In re: OpenAI, ROSS’s AI platform cannot generate original expression, and the evidence here supports the opposite conclusion about transformativeness.” The court also noted that ROSS trained its AI to create a commercial substitute for Westlaw. In Bartz, Judge William Alsup had already distinguished the district court’s decision in this case as involving “a competing AI tool for finding court opinions,” which “was not transformative.”

Despite that common ground, the decisions differ in ways worth watching. The Third Circuit takes a different view as to the ultimate purpose of the use. In Kadrey v. Meta, covered in our prior Legal Development on that decision, Judge Vince Chhabria judged Meta’s downloading “in light of its ultimate, highly transformative purpose” of training a tool that generates diverse text; the Third Circuit did not focus on the training step but rather looked at ROSS’s ultimate purpose of building a competing research platform.

One particularly significant distinction between this case and the generative AI decisions in Bartz and Kadrey is how the courts view the impact on the market for licensing training data. In both Bartz and Kadrey, the courts declined to treat harm to a market for licensing works for AI training as cognizable in evaluating the fourth fair use factor. In Kadrey, Judge Chhabria held that harm from “the loss of fees paid to license a work for a transformative purpose is not cognizable,” because counting it would make the fourth factor circular, and in Bartz, Judge Alsup held that a market for licensing books to train models is “not one the Copyright Act entitles Authors to exploit.” The Third Circuit, in contrast, having already found the use minimally transformative, credited a training data licensing market as a cognizable derivative market to be considered as part of the fourth fair use factor and did not address the arguments raised in the other two cases.

On necessity, the Third Circuit read the intermediate copying cases as resting on the need to copy to reach unprotected functional elements, while Bartz asked whether training copies were “reasonably necessary,” and Kadrey said fair use “doesn’t require that the secondary user make the lowest number of copies possible.”

Because this case involves a nongenerative AI tool built to compete with the copyright owner’s own product, it remains to be seen how, if at all, its reasoning will be applied in the various generative AI cases pending in other courts.