Written by Josh Geller
On September 29, the Third Circuit affirmed the U.S. District Court’s 2025 ruling that ROSS Intelligence infringed Thomson Reuters’s copyrights when it used Westlaw headnotes to train a competing legal research tool, and that the use was not protected as fair use. The ruling is the first federal appellate decision on whether training an AI system on copyrighted works is fair use.
The lawsuit involved training a non-generative AI tool incapable of creating expressive works. The tool directly competed with the content owner. Given those facts, ROSS is a (relatively) easy case, which is an awkward thing to say after the district court needed two tries before deciding. Even so, the decision contains important clues for how courts might approach the many ongoing infringement lawsuits involving generative AI tools.
A quick refresher
Thomson Reuters owns Westlaw, whose editors write headnotes summarizing the points of law in judicial opinions. ROSS Intelligence, a startup, built an AI legal search engine that answered plain-language questions with passages from opinions. The tool was not generative and “would only return text passages from preexisting judicial opinions” (slip op. 5). ROSS hired a company to write about 25,000 legal memos to train its model, and that company’s drafters used Westlaw headnotes to write the questions in those memos. Thomson Reuters sued.
In September 2023, Judge Bibas, a Third Circuit judge sitting by designation in Delaware, held that nearly every issue involved disputed facts that a jury should decide. Shortly before trial, he reversed himself in February 2025. After opening his opinion with, “A smart man knows when he is right; a wise man knows when he is wrong,” he granted Thomson Reuters partial summary judgment on originality, infringement, and fair use for 2,243 headnotes and certified the originality and fair use questions for appeal. The Third Circuit’s September 29 decision affirmed Judge Bibas’s second ruling.
While all this was going on, two Northern District of California judges issued the only other merits rulings on AI training. In June 2025, Judge Alsup in Bartz v. Anthropic and Judge Chhabria in Kadrey v. Meta each held that training a general-purpose large language model on books was transformative under the first fair use factor set forth in Section 107 of the Copyright Act, 17 U.S.C. §107, and, on the records before them, fair use, while disagreeing about nearly everything else. See Bartz v. Anthropic PBC, 787 F. Supp. 3d 1007, 1016 (N.D. Cal. 2025); Kadrey v. Meta Platforms, Inc., 788 F. Supp. 3d 1026, 1035 (N.D. Cal. 2025).
Three takeaways from the Third Circuit’s opinion
This was the straightforward case. ROSS built a narrow, non-generative search tool, trained it on Westlaw’s headnotes, priced it like Westlaw, and admitted it aimed to replace Westlaw. The panel called the appeal “no more than an ordinary copyright case” (slip op. 2), and on these facts it was. If you copy someone’s work to build a product that competes with that work, you should expect to pay for it. That is the thrust of the opinion.
The hard question is still open. Footnote 7 may be the most important passage in the opinion. It distinguishes Bartz v. Anthropic on two grounds: ROSS’s tool “cannot generate original expression,” and ROSS trained it “for the purpose of creating a commercial substitute for Westlaw.” The panel never says which ground carries the weight. The yet-to-be-decided case involves a general-purpose generative model, highly transformative, that competes with the works it was trained on but also has myriad additional uses. Kadrey v. Meta left the same question open when Judge Chhabria said market dilution is the operative theory but that the plaintiffs had not demonstrated it on the record before him.
Market substitution continues to carry the most weight after Warhol. In determining whether a use is fair under Section 107 of the Copyright Act, a court will weigh four factors: broadly, the purpose and character of the secondary use, the nature of the primary work (factual or creative), the substantiality of the use, and market harm. Andy Warhol Found. for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508 (2023). Post-Warhol, many of those factors turn on the same operative question: whether the secondary use serves the same purpose as the original and so substitutes for it in the market.
The holding
Originality. The first question certified for appeal was whether Westlaw’s headnotes were sufficiently original to earn copyright protection in the first place. As enjoyable as litigation over headnotes might be to a handful of litigators, this part of the opinion is, unsurprisingly, not particularly interesting. Thomson Reuters’s act of choosing which points of law to summarize and how to word them clears Feist’s “extremely low” bar for originality (slip op. 10–11). The panel declined to decide whether headnotes that quote an opinion word for word are protectable, because none of the 2,243 at issue on appeal did so (slip op. 10 n.5).
Factor one. Factor one examines the “purpose and character” of the use, including whether it was commercial, and whether it was transformative. The use was commercial and “minimally transformative, at best,” because both companies use headnotes “to create and optimize a legal-research platform that helps users find responsive legal material” (slip op. 16–17). The panel acknowledged that ROSS “took an intermediate step of using the headnotes to train an AI program, which arguably presents a slight degree of difference in use” (slip op. 16), but ultimately looked at the finished product to evaluate the actual purpose and character of the use. The intermediate-copying cases (Sega Enters. Ltd. v. Accolade, Inc., 977 F.2d 1510 (9th Cir. 1992); Sony Computer Ent., Inc. v. Connectix Corp., 203 F.3d 596 (9th Cir. 2000); Google LLC v. Oracle Am., Inc., 593 U.S. 1 (2021)) did not help, because copying in those cases was necessary to reach unprotected functional material, while ROSS could have built its memos from the freely available opinions. “Unlike necessity, ease is not a justification for copying” (slip op. 21). Factor one came out for Thomson Reuters and no fair use.
Factor two. The second factor evaluates the “nature of the copyrighted work,” including the extent to which the work consists primarily of factual or functional elements. Given the minimally expressive nature of the headnotes, this factor favored ROSS and a finding of fair use.
Factor three. Factor three considers the amount and substantiality of the use, including whether the use was reasonable in light of the purpose identified in factor one. Here, the panel departed from Judge Bibas, who had weighed the factor for ROSS because no headnotes appear in ROSS’s output. The panel never mentioned output. It asked only whether ROSS took more than its purpose required, and since the opinions were free, copying the headnotes was not necessary at all (slip op. 23).
Factor four. Finally, the fourth factor considers the market effects of the use. ROSS built a competing platform in Westlaw’s market. It also harmed the headnotes’ “value” as a feature that sells Westlaw subscriptions, even though nobody buys headnotes on their own (slip op. 24–25). And “the market for licensing headnotes as text to train AI is rapidly developing,” a market ROSS “usurped” (slip op. 26). This factor—regarded as the most important—therefore favored Thomson Reuters.
The AI training cases to date, and where they diverge
Ross departs from Bartz and Kadrey on three points.
On factor one, Judges Alsup and Chhabria treated training a general-purpose language model as transformative in itself—“spectacularly so” in Judge Alsup’s words. The Third Circuit judged transformativeness by what the trained product does. Training is not an end in itself, and the purpose and character of the use must be judged by what the tool is being trained to do.
On factor three, both California judges accepted that copying whole books was reasonably necessary because models improve with more good text. The Third Circuit asked whether the defendant could have used public material instead. The question about what training data is actually necessary to make a functional model remains an open technical question that will likely continue to be litigated.
The opinions also diverge sharply on the doctrinal question of the licensing market and its role in the market substitution factor. Judge Alsup assumed a market for licensing books as training data could develop. But he also held (without much explanation) that the authors were not entitled to that market. Judge Chhabria said whether such a market exists “is irrelevant,” because counting lost fees for an otherwise transformative use would make factor four “circular”—a finding of fair use always deprives a content owner of a theoretical licensing market. The Third Circuit recognized that a licensing market for headnotes as AI training data “is rapidly developing,” and did not address the circularity argument. Leaving aside the doctrinal question, licensing of training data is a commercial reality, and one that we have seen picking up at an accelerating pace over the past year plus since the Kadrey and Bartz decisions.
The intermediate step… and the next ones
Perhaps the operative paragraph from the Third Circuit opinion is this one:
ROSS took an intermediate step of using the headnotes to train an AI program, which arguably presents a slight degree of difference in use. But the undisputed evidence demonstrates that ROSS used the headnotes to train an AI program for the benefit of its legal-research platform. So both Thomson Reuters and ROSS use the headnotes to create and optimize a legal-research platform that helps users find responsive legal material. Thus, ROSS’s use of the headnotes shares the same ultimate purpose as Thomson Reuters’s use, making ROSS’s use minimally transformative, at best. (slip op. 16)
Calling training “an intermediate step” is like Carl Sagan’s observation that to bake an apple pie from scratch, you must first “invent the universe.” Every output has upstream inputs, and that does not make each input an intermediate step toward it. The panel is right that purpose controls: Warhol explains that “the same copying may be fair when used for one purpose but not another” (598 U.S. at 533). But training is not merely incident to the final product. What if the trainer does not know what the product will be? What if training and deployment are done by different companies? What if the trainer trains a general-purpose tool that serves innumerable functions, only some of which compete with and displace content owners?
The panel leaned on ROSS’s admission that it “aims to replace Westlaw.” A tool that does not aim to replace a work it copies—but does so anyway—is the more nuanced case. The practical question is what to do about general-purpose models that push into every vertical and compete with the works they trained on. Kadrey teed it up and Ross does not answer it. Bartz, of course, yielded a settlement: Anthropic agreed to pay about $1.5 billion to settle the pirated-books class claims.
Meanwhile, on September 1, 2026, the Justice Department filed a statement of interest in In re OpenAI arguing, on the strength of Bartz, that training a large language model is transformative and not substitutive. Judge Bibas’s decision was the first word on whether AI training is fair use; the Third Circuit’s ruling is not the last.
