Thomson Reuters v. ROSS on Appeal — Training an AI on Westlaw Headnotes Was Not Fair Use

On 29 September 2026 the US Court of Appeals for the Third Circuit, sitting in Philadelphia, decided Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc. (No. 25-2153). It affirmed. The headnotes of the Westlaw legal-research service are protected by copyright, and ROSS's copying of them to train its own AI was not fair use. The panel was Judges Restrepo, Montgomery-Reeves and Bove; Judge Montgomery-Reeves wrote the opinion of the court.

Until then every substantive ruling on AI training and fair use, Bartz v. Anthropic included, had come from a trial court. The law firm Jackson Walker called this the first federal appellate decision directly addressing copyright fair use in the AI-training context, and the legal news service MLex described it as the first US appellate ruling on fair use for AI training — albeit, it added, outside the context of generative AI.

What was at stake

Westlaw organises judicial opinions with headnotes: annotations that pull out a point of law, sit above the opinion and link to the passage they summarise. According to the court, Thomson Reuters's editors decide which points matter, write each headnote so that it makes sense without the opinion, and keep it to 800 characters where possible. Thomson Reuters claims no copyright in the opinions themselves, but its registered copyright covers the headnotes.

ROSS was founded by three University of Toronto computer-science students after a competition built around IBM's Watson. Their product was a search engine that took plain-language legal questions and answered with relevant passages from a bank of about ten million judicial opinions. The court was explicit that ROSS's AI was not generative: it created no new expression and only returned passages from existing opinions.

To train it, ROSS commissioned a company called LegalEase Solutions to write training memos — about 25,000 of them, each a legal question paired with four to six passages from opinions, labelled great, good, topical or irrelevant. A footnote in the opinion describes the AI as a program that learns patterns from examples of inputs paired with outputs, the classic shape of supervised learning. The trouble was that the memo writers framed their questions from Westlaw headnotes, which offered, in the words of the record, an easy way to frame them.

Thomson Reuters sued in the District of Delaware in 2020. Judge Stephanos Bibas, a Third Circuit judge sitting by designation, granted partial summary judgment on 11 February 2025, holding that ROSS's copying of 2,243 headnotes infringed and was not fair use. At the parties' request he certified two questions for interlocutory appeal: whether the headnotes were original, and whether ROSS's use was fair. The appeal was argued on 11 June 2026.

The holding

On originality, the bar is "extremely low": under Feist, a work needs only "some creative spark." The court found that spark in the editors' choices about which points of law to include and how to word them, and it relied on Callaghan v. Myers (1888), in which the Supreme Court recognised a case reporter's copyright in headnotes. The originality of the Key Number System, Westlaw's classification scheme, was not decided, because ROSS never briefed it.

On fair use, the four statutory factors came out like this:

Under the first factor, ROSS invoked Authors Guild v. Google, the 2015 decision upholding Google's scanning of whole books for search, and the "intermediate copying" cases — Google v. Oracle, Sega v. Accolade and Sony v. Connectix. The court distinguished them: in each, copying was necessary to reach the unprotected functional elements of code, whereas ROSS had free access to the opinions and chose the headnotes because they were convenient. "Unlike necessity, ease is not a justification for copying," the opinion says.

Under the third factor, ROSS argued that it took only 0.08% of Thomson Reuters's 28 million headnotes. The court answered that each headnote is a copyrightable work of its own, so for every headnote taken ROSS copied an entire work.

Under the fourth, the court noted that Thomson Reuters uses its headnotes to train its own AI search products, including WestSearch Plus, and said the market for licensing headnotes as AI training text is rapidly developing; that Thomson Reuters had not licensed them to others did not prove the market absent. ROSS's argument that AI is critical to national security failed for lack of any evidence linking a legal-research platform to national security.

Set apart from generative AI

The scope of the decision turns on footnote 7. On 1 September 2026 the Justice Department had filed a statement of interest in the consolidated copyright litigation against OpenAI in the Southern District of New York (In re: OpenAI, Inc. Copyright Infringement Litigation), relying on Bartz to argue that training a large language model that can generate original responses is transformative. The Third Circuit said those concerns did not apply: "Unlike the AI models in Bartz and In re: OpenAI, ROSS's AI platform cannot generate original expression." It added that ROSS had trained its AI to create a commercial substitute for Westlaw.

The ruling therefore does not announce that AI training is never fair use. It denies fair use on these facts — a non-generative search tool competing head-on in the original work's own market. Jackson Walker cautioned against reading it either as a categorical bar on AI training or as a general safe harbor. The district-court rulings that found generative-AI training fair use — Bartz and, two days later, Kadrey v. Meta — stand alongside it, separated by their facts. The newspaper suits against OpenAI are traced in NYT v. OpenAI.

The opinion also recorded in a footnote that ROSS at times acted in bad faith, including attempts to reach Westlaw with a law-firm investor's credentials, while noting the Supreme Court's own skepticism about whether good faith still matters to fair use.

Afterwards

According to Courthouse News, a Thomson Reuters spokesperson welcomed the decision, and Yar Chaikovsky of White & Case, ROSS's counsel, said ROSS intends to seek Supreme Court review. Both Courthouse News and Jackson Walker describe ROSS as now defunct.

Separating training-data questions by whether a system generates and whether it competes in the original market is now written into an appellate opinion. The wider story is on the AI and copyright timeline.