Third Circuit rules against ROSS for copying headnotes it had no need to copy

What the opinion says about itself

On Sept. 29, 2026, the U.S. Court of Appeals for the Third Circuit affirmed a ruling that ROSS Intelligence infringed Thomson Reuters’s copyright when it used Westlaw headnotes to build a competing legal-research tool. The court also held that the use was not a fair use.¹

Coverage compressed that into a line about AI. Music Business Worldwide headlined it as a rejected fair-use defense over AI training, and Reuters’s wire report called it the first appeals-court ruling of its kind.³⁴ The opinion makes neither claim. It states no rule about training in general and never says it is the first. It calls itself “no more than an ordinary copyright case,” in which the court must decide “whether Thomson Reuters’s materials are sufficiently original for copyright protection and whether ROSS’s use of those materials was fair.”¹ Music Business Worldwide’s own text adds that the judgment does not settle fair use in the Suno and Udio music cases, because ROSS did not run a generative system.³

Fair use is the defense that lets some copying of a protected work go ahead without the owner’s permission. A court weighs four factors from the statute. They are the purpose and character of the use, the nature of the copyrighted work, the amount and substantiality of the portion used, and the effect of the use on the potential market for the work. The party claiming fair use carries the burden, which here was ROSS.¹

The panel of Judges Restrepo, Montgomery-Reeves and Bove was unanimous. Judge Montgomery-Reeves wrote the opinion, and there is no concurrence or dissent.¹ The case came up as an interlocutory appeal, an appeal taken before the case is over, which the court may hear only because the District Court certified two questions at the parties’ request. The District Court had granted Thomson Reuters partial summary judgment, a ruling that needs no trial on the issues it decides, holding that ROSS’s use of 2,243 headnotes infringed and was not fair use. The certified questions were whether the headnotes are original enough to be copyrighted and whether the use was fair.¹ The Third Circuit answered both for Thomson Reuters. It found that every one of the 2,243 headnotes has “some creative spark.”¹ Coverage dates the news Sept. 30 because the court first filed the opinion under seal on the 29th.⁵

Summaries of opinions, copied into training data

Westlaw is Thomson Reuters’s legal-research platform. A headnote is a short note, written by Westlaw’s editors, on a point of law in a judicial opinion. The editors’ guideline is 800 characters “where possible.”¹ Thomson Reuters uses headnotes “to help legal researchers find and understand judicial opinions containing points of law.”¹

ROSS built a tool for the same job. A user types a legal question and the tool returns passages from judicial opinions. The opinion is explicit about what it does not do: “ROSS’s AI was not a generative AI, meaning it would not create any new expression; it would only return text passages from preexisting judicial opinions.”¹ Its answers come from a fixed bank of about ten million uncopyrighted judicial opinions.¹

The headnotes entered at the training step. ROSS bought about 25,000 legal memos from LegalEase and its subcontractor Morae Global, each holding four to six passages from judicial opinions, and trained its search tool on them.¹ The memos were written from Westlaw headnotes. ROSS’s own expert analyzed a batch of 2,830 headnotes. For 2,243 of them the District Court found the memo question was “so similar to the headnote text but so dissimilar to the opinion text” that no reasonable juror could find it was not copied from the headnote.¹ The headnotes never appear in ROSS’s final product.¹ ROSS also ran ads comparing its platform to Westlaw, at prices in line with Westlaw’s.¹

The free opinions made the headnotes unnecessary

The first factor asks about the purpose and character of the use. The court found ROSS’s use commercial. ROSS “aimed to create a legal-research platform charging comparable prices to Westlaw to compete with Westlaw for customers.”¹ The harder question is whether the use was transformative, meaning whether it serves a different purpose from the original’s. The court compared the two. Thomson Reuters uses headnotes to help researchers find and understand opinions, and ROSS uses them to build a platform that helps users find responsive opinion passages. Training an AI “arguably presents a slight degree of difference,” but both parties use the headnotes to create and optimize a platform that helps users find responsive legal material, so the use is “minimally transformative, at best.”¹

ROSS’s answer was a doctrine about intermediate copying, which means copying a work as a step toward making something else, where the copy does not appear in what comes out. ROSS cited three software cases, Sega v. Accolade, Sony v. Connectix and Google v. Oracle, as dictating a different result.¹ In each, computer code was copied so that users could make their software work with an existing computer system, and the copying was held transformative.¹ ROSS’s briefs are not available here, so its position comes from the opinion’s account and from an amicus brief by CCIA, Chamber of Progress and NetChoice. That brief argues that the headnotes “do not appear as part of the final product” and that Sony shows a transformative use can compete with the original.⁶

On this view, the headnotes were raw material, and what a user receives is text from opinions nobody owns. The use is a different one from Westlaw’s, and the cases that excuse copying code to reach unprotected material apply. An outside critic makes the strongest version. Dave Hansen of Authors Alliance argues that the panel treated two competing businesses as two competing uses, when users never see the headnotes.⁷

The court’s answer is a distinction drawn from those same cases. “In each of these cases, computer code was copied to enable users to make their software operable with an existing computer system. These uses were transformative because, in each case, copying was necessary to access the unprotected functional aspects of computer code.” ROSS, by contrast, “does not need to copy” the headnotes to reach the underlying unprotected information. It “had access to the underlying judicial opinions and could freely copy them,” and it chose the headnotes because they were an “easy” route.¹ The court summarized the point this way: “Unlike necessity, ease is not a justification for copying.”¹

The reasoning never mentions generation. Its operative fact is that the opinions were free. A footnote adds a second, code-specific ground. Oracle turned in part on declaring code being “inherently bound together with uncopyrightable ideas” and on Java’s value coming from programmers adopting it. By contrast, “Thomson Reuters’s headnotes do not derive value from others copying them.”¹ The District Court had called these “computer-programming copying cases” that “depend in part on the need to copy.” The panel adopted the need-to-copy point and the code-specific point. It did not adopt the label as a limit.⁸

The Third Circuit moved factor three against ROSS

Factor three looks at the amount and substantiality of the portion of the work used. The District Court gave it to ROSS. It reasoned from Authors Guild v. Google that what matters is not the portion copied but “the amount and substantiality of what is thereby made accessible to a public,” and ROSS made no headnotes accessible to anyone.⁸

The Third Circuit counted the factor against ROSS on different reasoning. ROSS copied the headnote text in full into its memos, and because its purpose was “highly similar” to Thomson Reuters’s, no transformative purpose made extensive copying reasonable. ROSS argued that it took “only 0.08% of Thomson Reuters’s 28 million headnotes.” That fails because copying even an insubstantial portion “is not fair if, qualitatively, an important part of the work is copied.” Each headnote is also a whole copyrightable work. Necessity returns as well: “copying the headnotes was also not necessary to train ROSS’s AI, as the underlying judicial opinions were freely available.”¹ The same fact now carries factor one and factor three.

The opinion says “we will AFFIRM” and does not mention that the vote on this factor changed.¹ Under the District Court’s own weighing, the score was two factors each, with factor four counted for more. The panel’s count is three to one, and it does not rank the factors.¹ A description of the panel as agreeing with the District Court factor by factor would be wrong on factor three.

The opinion also contains a number it does not reconcile. In the factor-three paragraph it says ROSS copied “the entire text of the 25,000 Westlaw-written headnotes.” Liability covers 2,243 headnotes, and 0.08% of 28 million is 22,400. The 25,000 figure is also the memo count from the facts.¹ The opinion does not say which it means.

Market harm rested on three findings, and the licensing one is thin

Factor four asks how the use affects the market for the original and for works derived from it. The Third Circuit found harm in three places.¹

First, a standalone market for headnotes. ROSS argued that the relevant original market is the market for headnotes as a product, and that none exists. The court assumed it did not exist and turned instead to the effect on the value of the work: headnotes draw users to a Westlaw subscription, and ROSS “diminished their value as a draw.”¹ Second, the legal-research platform market. ROSS “copied the headnotes to create a competing legal-research platform in the same market as Westlaw,” and “offers no evidence to rebut the alleged harm.”¹

Third, a derivative market, which is a market for new uses of a work that its owner could license. This is the finding most often read as being about AI training. The court wrote that “the evidence shows that the market for licensing headnotes as text to train AI is rapidly developing,” and that Thomson Reuters uses its own headnotes as training data for its AI search products. “That Thomson Reuters did not license its headnotes to others does not disprove that a market exists.” ROSS “usurped Thomson Reuters’s opportunity to enter that derivative market.”¹

The evidence the opinion gives for that market is those two sentences. It cites no license, no price and no figure, and nothing from the record beyond Thomson Reuters’s own use. The District Court had said it “does not matter whether Thomson Reuters has used the data to train its own legal search tools”; the Third Circuit instead treats that use as evidence the market exists.⁸ The finding concerns headnotes, and the opinion does not say whether such a market would count for a generative model. It also does not rank the three findings or say that factor four depends on the licensing one.¹

The court then took up ROSS’s claims of public benefit. The access claim was “unclear,” since the opinions were freely available and ROSS’s prices were comparable to Westlaw’s. “ROSS presents no evidence to support” its claims that the ruling would halt AI development or harm national security. The court added that “this does not give ROSS carte blanche to violate copyright law merely because it incorporates AI.”¹

Generative AI gets one footnote

The opinion uses the word “generative” twice. One use is the description of ROSS’s tool quoted above. The other is footnote 7, attached to the sentence calling ROSS’s use “minimally transformative, at best.”¹

The footnote notes a statement of interest the Department of Justice filed on Sept. 1, 2026 in In re OpenAI, Inc. Copyright Infringement Litigation, No. 1:25-md-3143, in the Southern District of New York. The DOJ relied on Bartz v. Anthropic for the view that training a large language model that can “generate original responses” is transformative. It also argued that OpenAI’s training did not cause “substitutive competition.” The court answered: “The concerns raised in that separate case do not apply here.” It gave two grounds. “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.” And “Here, ROSS trained its AI for the purpose of creating a commercial substitute for Westlaw.”¹

The footnote does not use the words reserve, decline or leave open. “Do not apply here” is a statement about this record, resting on two facts about ROSS: it generates no expression, and it was built to substitute for Westlaw. The footnote does not say how a court should treat a system that differs on the first fact and not the second, or on both. Law-firm alerts describe the opinion as leaving generative-AI questions unresolved, which is their reading of the footnote and not its words.⁹ The footnote also stands apart from the necessity reasoning of the earlier section. Neither is tied to the other in the opinion.¹

Two other readings come from outside the opinion. District Judge Bibas wrote in 2025 that “only non-generative AI is before me today,” and that sentence is his. The panel’s opinion does not adopt it.⁸ Hansen of Authors Alliance reads the footnote’s emphasis on substitution as an invitation to plaintiffs in generative-AI cases to argue that competing models belong in ROSS’s category. That is his forecast, and the opinion draws no such category.⁷

What a court facing a generative system would still have to decide

The opinion establishes one result. ROSS’s use of Westlaw headnotes to train a search tool that returns judicial passages and competes with Westlaw was not a fair use. It reaches that result on three factors against ROSS and one slightly for it. The software intermediate-copying cases did not help ROSS because it did not need the headnotes to reach the unprotected opinions. On factor three the panel changed the District Court’s answer without saying so.¹

It does not establish a rule about AI training. It does not say whether the intermediate-copying cases would come out differently for a generative system, or whether the licensing-market finding would carry over to one. On generative systems it says only what footnote 7 says. The case now returns to the District Court in Delaware for trial on the remaining issues.²

The Third Circuit decided that ease is not necessity for a company that had about ten million free judicial opinions in reach. On generative AI, it wrote only that the OpenAI case’s concerns “do not apply here.”¹