Thomson Reuters v. Ross: The First Refusal of Fair Use in the AI Era

Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 1:20-cv-00613-SB (D. Del. Feb. 11, 2025), occupies a peculiar place in the emerging law of artificial intelligence and copyright: it is the first federal decision to reject a fair-use defense for the use of copyrighted material to build an AI system, and it is also the decision most often over-read. Judge Stephanos Bibas — a Third Circuit judge sitting by designation on the district court — granted partial summary judgment to Thomson Reuters on direct infringement and rejected Ross’s fair-use defense as a matter of law. That the same judge had reached a more defendant-friendly conclusion in an earlier 2023 ruling, before revising course, only underscores how unsettled this terrain was.

At a glance

• Case: Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 1:20-cv-00613-SB (D. Del.)
• Decided: February 11, 2025 (revised memorandum opinion), Judge Stephanos Bibas (sitting by designation)
• Holding: Westlaw headnotes are copyrightable and were copied; Ross’s use to build a competing legal-research tool was not fair use
• Status: Not final — certified for interlocutory appeal to the Third Circuit, which accepted the appeal

The conduct and the copyrighted work

The dispute concerned Westlaw’s headnotes and its Key Number System — the editorial apparatus by which West summarizes and classifies points of law drawn from judicial opinions. Headnotes are not the opinions themselves (those are uncopyrightable government works) but West’s original distillations of them, and the Key Number System is West’s proprietary taxonomy of legal topics.

Ross, a legal-research competitor, set out to build a search tool. Lacking a Westlaw license, it obtained “bulk memos” from an intermediary, LegalEase, that had been derived from Westlaw headnotes, and used them to train its system. The court found that the headnotes cleared copyright’s modest originality threshold and that a substantial number — 2,243 of the 2,830 at issue — had in fact been copied and were substantially similar to the material used to train Ross’s tool. Establishing actionable copying of protected expression was the predicate; the contest was over fair use.

The fair-use analysis

On the four factors of 17 U.S.C. § 107, the court treated the first (purpose and character of the use) and the fourth (effect on the market) as the most important, and both favored Thomson Reuters.

The first factor turned on transformation and commerciality. The court found the use commercial and, critically, non-transformative: Ross was not repurposing the headnotes toward some new expressive end but building a product designed to occupy the same market as Westlaw — a legal-research substitute. Where the secondary use shares the purpose of the original and competes with it, the first factor cuts against fair use, a conclusion squarely consistent with Warhol’s emphasis on comparing purposes. The fourth factor followed almost mechanically: a market substitute is the paradigm of cognizable market harm, and the court credited the threat to Westlaw’s market, including the potential market for AI training data derived from its content.

The passage that defines the decision

The most doctrinally important portion is the court’s treatment of the software intermediate-copying cases — Sega Enterprises v. Accolade, Sony Computer Entertainment v. Connectix, and Google LLC v. Oracle America. Ross had invoked that line for the proposition that copying undertaken to reach unprotected functional elements can be fair use. Judge Bibas distinguished those authorities on the ground that they involved copying that was necessary to access uncopyrightable functional material — interface specifications, the means of interoperability — whereas Ross’s copying served to appropriate protected expression itself, the very editorial work that gives headnotes their value.

That distinction is the analytical core of the opinion, and it is also what separates Ross from the generative-AI cases that would follow. The intermediate-copying defense depends on the copying being a means to a functional, non-expressive end. Ross could not characterize its use that way, because the value it extracted was the expressive judgment embedded in West’s headnotes.

The limit that defines its reach

Ross is frequently cited as authority against AI training writ large. It is more precise — and more limited — than that. Ross’s product was a non-generative AI: a search tool that returned relevant judicial opinions in response to queries. It did not generate new expressive output, and the court’s reasoning was tethered to a use that competed directly and non-transformatively with the very work copied.

The generative-AI decisions handed down months later drew the contrast explicitly. In Bartz v. Anthropic and Kadrey v. Meta, courts found training to be highly transformative precisely because the output was a new tool rather than a substitute for the ingested works. Ross is therefore best understood not as a verdict on machine learning generally but as a conventional application of fair-use doctrine to a competitor who copied protected expression to build a rival product in the same market.

Its procedural posture reinforces the caution. The district court certified the question for interlocutory appeal, the Third Circuit accepted it, and the matter remained pending — with amicus participation — through late 2025. The holding is thus not even final within its own circuit.

Open questions

The Third Circuit’s resolution will be the first federal appellate word on fair use in an AI context, and it could either entrench or unsettle the district court’s non-transformative finding. Also unresolved is how the “potential market for AI training data” should figure in the fourth-factor analysis — a theory that, if generalized, could reshape the calculus well beyond non-generative tools.

Implications

• Cite it for what it decided. Ross is sound authority that copying protected expression to build a directly competing product is not fair use. It is weak authority for any categorical claim about generative-AI training.
• The non-generative/generative line matters. The transformative-use analysis depends heavily on whether the system produces new expressive output or merely substitutes for the copied work.
• Watch the Third Circuit. Until the appeal is decided, the holding is provisional even in Delaware.

Frequently asked questions

Was this a generative-AI case? No. Ross built a non-generative legal-research search tool, which is why courts have been careful not to read it as controlling the generative-AI training cases.

Is the decision final? No. It was certified for interlocutory appeal and the Third Circuit accepted review; the appeal was pending as of late 2025.

Why did the software fair-use cases not help Ross? Because those cases excuse copying needed to reach uncopyrightable functional elements. The court found Ross copied protected expression — West’s editorial headnotes — not mere functional material.

Authorities and sources

• Revised memorandum opinion, Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 1:20-cv-00613-SB (D. Del. Feb. 11, 2025): opinion PDF (D. Del.).
• Analysis: Davis Wright Tremaine; Loeb & Loeb.
• Appeal: IPWatchdog on the Third Circuit appeal.