US court rejects fair use in Thomson Reuters v Ross Intelligence

Background

Westlaw is Thomson Reuters' legal research service. Beyond the text of court decisions, it publishes editorial material built on top of them: headnotes, which summarise the points of law a case decides, and a Key Numbering System, which sorts legal topics into a numbered classification so that lawyers can trace related cases across jurisdictions. It was this editorial layer, rather than the judgments themselves, that the dispute centred on.

Ross Intelligence was building an AI legal research tool meant to compete with existing services. To train and validate that tool, Ross used Westlaw headnotes^[1][2].

The ruling

On 11 February 2025 a US federal judge in Delaware granted partial summary judgment to Thomson Reuters^[1][2]. The court found that Ross's use of Westlaw headnotes to train its AI legal research tool constituted direct copyright infringement^[1][2]. It rejected all of Ross's defences, including its fair use defence^[1][2].

The result was not reached at the first attempt. The judge had earlier declined to grant summary judgment to either side and had been preparing for trial before reconsidering the parties' cross-motions^[1].

Key details

A claim of direct infringement turns on two questions: whether the material was protected by copyright at all, and whether the defendant copied it.

On the first, the court held that the Westlaw headnotes and the Key Numbering System were original enough to be protected. In doing so it applied the low originality threshold set by Feist Publications v Rural Telephone Service, a standard that asks for only a modest degree of independent creation rather than novelty or artistic merit^[1][2]. Editorial material of this kind, in other words, is not a free resource simply because it is assembled from public court decisions.

On the second, the court found that Ross actually copied protected content from the headnotes in order to train and validate its tool^[1].

Much of the decision dealt with fair use, the doctrine that allows limited reuse of copyrighted material without permission. Considering the first of the fair use factors, the court found that Ross's use served a commercial purpose and was not transformative^[2]. Using the content to develop a competing AI product, the court held, took Ross outside the scope of fair use^[2]. Because that factor weighed so heavily, the remaining factors could not rescue the defence, and the court rejected it along with Ross's other arguments^[1][2].

Why it matters

The judgment was described as the first substantive judicial ruling on direct copyright infringement and fair use in the AI context^[1]. That description is what gives the case its weight: for years the central legal question around generative and other AI systems has been whether ingesting copyrighted text to build a model counts as fair use, and here a court answered that question against the developer on the specific material in front of it.

The decision is narrow in what it covers. It concerns a legal publisher's editorial headnotes and classification system, and it came as a partial summary judgment rather than the end of a full trial^[1][2]. But the reasoning — that copying protected editorial content to build a product that competes with the original is commercial and non-transformative — sets out a framework that later disputes over training data will have to argue with.