Thomson Reuters v Ross Intelligence: What the Ruling Decided

On 11 February 2025 the District of Delaware granted partial summary judgment to Thomson Reuters against Ross Intelligence, holding that Ross’s use of Westlaw headnotes to build training data infringed and that fair use did not excuse it. It is described as the first major fair-use loss for AI training in the United States. It is also a case about a product that did not generate anything.

What Ross actually built

Thomson Reuters filed suit in 2020 over Ross’s legal research tool. Ross wanted to build a search engine that answered legal questions by returning relevant judicial opinions. Judicial opinions are not copyrightable, so the opinions themselves were never the problem. What Ross needed was training data pairing legal questions with relevant material, and it obtained that from a third party which produced “bulk memos” — question-and-answer pairs whose questions were derived from Westlaw headnotes.

Westlaw headnotes are editorially written summaries of points of law drawn from opinions, organised under the West Key Number System. Thomson Reuters claimed copyright in them as original editorial work. The critical product fact is that Ross’s output was a list of relevant judicial opinions. It did not reproduce headnotes to users and it did not generate text.

What the court held

Judge Stephanos Bibas, a Third Circuit judge sitting by designation in the district court, had previously denied summary judgment in 2023 and set the case for trial. He then revisited the issues, vacated the earlier opinion, and issued a revised memorandum opinion in February 2025 that granted Thomson Reuters partial summary judgment.

Two holdings sit at its centre. First, on copyrightability and infringement: a defined set of headnotes — 2,243 of them — were held to be original enough to be protected and to have been actually copied, with the court finding direct infringement as to those. Headnotes were treated as original because they involve editorial judgment in selecting and expressing a point of law, even where they closely track uncopyrightable judicial language. Other asserted headnotes were left for trial. Second, on defences: fair use was rejected on this record, as were the other defences Ross raised. The opinion is available through CourtListener’s docket archive.

The fair-use reasoning

Bibas worked the four statutory factors and treated the first and fourth as the most important, following the Supreme Court’s recent emphasis.

• Purpose and character: against Ross. The use was commercial, and the court held it was not transformative, because Ross used the headnotes to build a product serving the same purpose as Westlaw — helping lawyers find relevant law. The intermediate copying cases from software, where copying was necessary to reach unprotectable functional elements and reach a different market, were distinguished as computer-code cases with no application here.
• Nature of the work: for Ross. Headnotes sit near the factual end of the spectrum, with thin originality.
• Amount and substantiality: for Ross. Weighed in Ross’s favour largely because the headnotes were not shown to users in the output.
• Market effect: against Ross. The court identified harm to the market for Westlaw itself and, importantly, to a potential derivative market for licensing legal research data for AI training. The existence of that second market as a cognisable interest is the piece with the widest implications.

The derivative-market reasoning is what connects this case to the wider debate. If the market for licensing training data is a market rightsholders are entitled to exploit, then unlicensed training harms it by definition, and the fourth factor tilts before anything is said about outputs. That argument also appears in the Copyright Office’s analysis, described in the Part 3 report page, and it is contested precisely because it can be made circular.

How far the ruling reaches

The opinion itself contains the caveat that matters: the court noted the case did not involve generative AI. Several features limit its transfer to the cases everyone is actually watching.

• No generation. Ross produced a ranked list of public documents, not new text. Arguments about transformation that rest on a model producing something categorically different from its inputs were not available to it.
• A direct competitor. Ross was building a rival to the very product whose material it copied. That is the least favourable possible posture on both the first and fourth factors, and it is not the posture of a general-purpose model trained on a broad corpus.
• A narrow, identified set of works. 2,243 specific headnotes with a traceable copying path, rather than a web-scale corpus whose contents are disputed.
• Third Circuit law. The decision is a district court ruling applying Third Circuit precedent, and the generative AI cases are concentrated in other circuits.

What does transfer is the framing: that a use is not transformative merely because it is technical, and that the loss of a licensing market for training data can count as market harm. Both of those propositions are now in the briefing of nearly every AI training case, including the New York Times proceedings.

Posture, and why it is not final

This is a district court decision on partial summary judgment, and it is not the last word even in its own case. The court certified its fair-use holding for interlocutory appeal, and the Third Circuit granted permission to hear it, so the fair-use analysis has been placed before an appellate court that can affirm it, narrow it, or vacate it.

That posture is exactly why the case should be cited carefully. If the Third Circuit affirms, this becomes the first appellate authority on fair use for AI training data and its weight increases considerably. If it reverses, a great deal of commentary written in 2025 becomes wrong. Either way, the version of the reasoning that survives may not be the version described here.

One last distinction worth holding on to. This case decided that copying particular editorial works to build training data, for a product competing with their source, was not fair use on that record. It did not decide that training is infringement, that fair use is unavailable to model developers, or anything about outputs. The gap between that holding and the headline it generated is the reason to read the opinion.