Transformative Means More Than “Powered by AI”

Last week, the Third Circuit issued one of the first appellate rulings on copyright and AI training in Thomson Reuters v. ROSS Intelligence. The court held Westlaw’s headnotes copyrightable and ROSS’s copying not fair use, with three of the four fair-use factors cutting against ROSS. The reasoning is less about AI than about purpose: when the end product serves substantially the same purpose as the source, the intermediate training step may add little transformative weight. And, as a warning to AI developers, the court recognized a “rapidly developing” market for licensing copyrighted content, such as the headnotes at issue, for AI-training data. This strengthens copyright plaintiffs’ future arguments of market harm.

Background

ROSS built a non-generative AI legal tool that answered plain-language legal questions with passages from public judicial opinions. To train it, ROSS used Westlaw headnotes to frame the questions and marketed itself as a Westlaw substitute at comparable prices.

In February 2025, Judge Bibas, sitting by designation in Delaware, held that 2,243 headnotes were original and infringed. He gave fair-use factors one and four to Thomson Reuters, factors two and three to ROSS, and rejected the fair-use defense. He certified for interlocutory appeal under 28 U.S.C. § 1292(b) the issues of “(1) whether the West headnotes and the West Key Number System are original as a matter of law and (2) whether ROSS’s alleged use of the headnotes was fair use.” The Third Circuit heard argument on June 11, 2026.

We covered the district court’s decision and the oral argument in more detail in prior alerts here and here.

The Holding

The Third Circuit panel affirmed the district court’s originality and fair-use rulings concerning the 2,243 headnotes but declined to address the originality of the Key Number System because ROSS forfeited that issue. According to the Third Circuit, the headnotes clear Feist’s “extremely low” originality bar. And ROSS’s use was not fair because it was commercial, minimally transformative at best, more extensive than necessary, and harmful to both Westlaw’s market and a developing market for AI-training licenses.

The Analysis

Originality. Each headnote reflects two creative choices: (1) which point of law matters and (2) how to phrase it so it stands alone yet tracks the opinion. The court found that was enough under Feist. ROSS’s three counterarguments failed. Headnotes are not law, so no monopoly results. Merger does not apply because there are many ways to express a point of law, drawing comparisons to a banana costume. And while precedent denies protection to citation formats dictated by industry convention, that does not extend to independently composed headnotes.

Factor one. The court concluded this factor weighs against fair use. First, ROSS’s use was highly commercial; it priced against Westlaw and sought its customers.

Second, it was also minimally transformative. Both parties use headnotes to help researchers find responsive legal material. Training an AI was an intermediate step that offered “a slight degree of difference,” but the ultimate purpose matched. The court distinguished Authors Guild v. Google because Google Books served a different function from reading and drove book sales, while ROSS aimed to replace Westlaw. The intermediate-copying cases ROSS cited also did not help, because each involved copying necessary to reach unprotected functional elements of code. ROSS had the public opinions and chose headnotes because they were easier. “Unlike necessity, ease is not a justification for copying.”

In a footnote, the court found ROSS “at times acted in bad faith,” citing record evidence of access to Westlaw through investor and student credentials. Although the court expressed doubt that good faith remains part of the fair-use analysis, it concluded that, to the extent the consideration remains relevant, it weighed against ROSS.

Factor two. The court noted that this factor rarely matters but slightly favored ROSS here. The headnotes were published and more factual than fictional.

Factor three. Here the panel broke from Judge Bibas, who had given this factor to ROSS because its output contained no headnotes. The Third Circuit treated each headnote as a complete work. ROSS copied each in full, and copying was unnecessary because the opinions were free. The court also rejected ROSS’s argument that it only took 0.08% of Thomson Reuters’s headnotes: a small share of a collection can still be a qualitatively important taking. The court held that the crux of the inquiry is whether “no more was taken than necessary” to achieve the copier’s purpose. And here, the purpose behind ROSS’s copying was “minimally transformative” and “not necessary to train ROSS’s AI.” Thus, this factor weighed against fair use.

Factor four. The court “consider[ed] harm to the original market, harm to the value of the copyrighted work, harm to the potential derivative market, and the alleged public benefits of the copying.” The court held that a missing standalone market for headnotes is irrelevant. Headnotes draw users to Westlaw, and ROSS appropriated that value. The legal-research platform market is also a relevant original market, and ROSS offered no evidence rebutting harm there. On derivative markets, the court found that the market for licensing headnotes as AI-training data is “rapidly developing” and that Thomson Reuters already trains its own tools on them. ROSS “usurped” that opportunity. The court dismissed ROSS’s public-benefit arguments: opinions are already free, ROSS charged Westlaw prices, and ROSS offered no evidence that the ruling would halt AI development or implicate national security.

Takeaways

1. Same function implicates both factors one and three. In the Third Circuit, factor one may weigh against a product that uses copyrighted material to serve a substantially similar function in the same market. This opinion adds a second cost: without a transformative purpose, copying whole works is unreasonable under factor three, regardless of what the end user sees. AI developers should document a purpose distinct from the source material before training begins. Copyright holders should build the substitution record early.
2. GenAI plaintiffs will likely invoke the market-harm analysis in GenAI cases. Although the court acknowledged that this is not a generative AI case, its market-harm analysis is not limited to non-generative AI. Therefore, copyright owners will likely cite the court’s finding, based on evidence before it, that a market for licensing headnotes as AI-training data is “rapidly developing,” arguing that the same reasoning supports recognition of training-data licensing markets in generative-AI cases. AI companies, in turn, will emphasize that the court’s finding rested on the particular evidentiary record concerning Westlaw headnotes.
3. Provenance and necessity matter. The court rejected ease as a justification and found bad faith on a record of credential misuse. Companies acquiring training data should consider public or licensed sources, document why copyrighted material was needed when a free alternative existed, and avoid access practices that may support allegations of bad faith. These considerations may be especially important when relying on intermediate-copying precedents grounded in necessity.
4. Factor one remains unsettled in generative AI cases. By expressly distinguishing Bartz and In re OpenAI and confining its holding to the facts before it, the Third Circuit did not resolve whether training generative models that produce new expressive outputs should be treated as transformative. Copyright owners will argue that the court rejected the notion that AI training is inherently transformative or presumptively transformative under Factor 1. AI companies, however, will emphasize that the court preserved the argument that generative AI warrants a fundamentally different transformativeness analysis from the one applied in ROSS. Together, the opinion’s treatment of transformativeness and market harm leaves several of the central questions now being litigated in generative-AI copyright cases.