Thomson Reuters v. Ross

A legal search tool, not a chatbot

ROSS Intelligence set out to build a natural-language search engine for case law — type a legal question, get relevant precedent back. It needed training material built from real legal research, and it wanted a Westlaw license to get it. Thomson Reuters refused. ROSS bought an alternative instead: "Bulk Memos" from a company called LegalEase, built by paying lawyers to write questions and answers derived from Westlaw headnotes.

Thomson Reuters sued in May 2020, alleging the Bulk Memos infringed its copyrighted headnotes. ROSS shut down in early 2021, citing litigation costs, but kept fighting the case rather than settle.

The reversal that produced the first anti-fair-use ruling

In September 2023, Judge Stephanos Bibas largely denied summary judgment and sent fair use to a jury. He then reversed course. On February 11, 2025, his revised ruling granted partial summary judgment for Thomson Reuters: direct infringement of 2,243 headnotes, and no fair use as a matter of law. Factors one (purpose and character of the use) and four (market harm) both favored Thomson Reuters.

The opinion leaned on two specifics: ROSS's tool was non-generative — it returned search results, not generated text — and it competed directly with Westlaw in the same market. Bibas also found that a market for licensing AI training data could itself be harmed, folding a forward-looking licensing market into the factor-four analysis.

“No fair use as a matter of law — factors one and four favor Thomson Reuters.”— Summary judgment order, D. Del., Feb. 11, 2025 (Judge Bibas)

On appeal, with trial on hold

The Third Circuit accepted ROSS’s interlocutory appeal in mid-2025, deferring trial while the appellate court considers the fair-use ruling itself. Amicus briefs supporting ROSS’s position followed in October 2025.

Oral argument came on June 11, 2026, before Judges Restrepo, Montgomery-Reeves and Bove. ROSS invoked the Third Circuit’s own recent decision in ASTM v. UpCodes — a fair-use win for a defendant reproducing standards incorporated into law — to argue for a fair-use analysis sensitive to functional, law-adjacent works. The panel pressed both sides hard on transformativeness and market harm. As of July 2026, the decision is pending.

Why "non-generative" is the qualifier that matters

It's tempting to read this as the ruling that decided AI training is not fair use. It didn't, and overstating it is the most common mistake in coverage of this case. Bibas repeatedly anchored his reasoning to the fact that ROSS's tool was non-generative — it never produced new text resembling the original, only returned matches. Whether that reasoning extends to a large language model that generates novel text is exactly the question the district court did not answer and the Third Circuit has not yet answered either.

That distinction matters enormously for anyone drawing conclusions from this case. Until the Third Circuit rules, and depending on how narrowly or broadly it writes that ruling, Thomson Reuters v. ROSS is authority for one specific fact pattern — a competing search tool trained on a rival’s compiled reference product — not a general verdict on generative AI training.

What it means for training-data licensing

If the Third Circuit affirms, the practical lesson for data buyers is specific: buying a compiled dataset derived from a competitor’s copyrighted product is high-risk, even if your tool never reproduces the original text, and even if your output looks nothing like the source. A recognized market for licensing training data becomes a cognizable factor-four harm — which strengthens every licensor’s hand in future negotiations, well beyond legal-reference data.

If the panel reverses or narrows the ruling, expect renewed argument that transformative, non-substitutive AI training tips back toward fair use even when the underlying corpus was compiled from a rival’s work. Either way, this is the single biggest pending appellate catalyst in AI copyright law, and licensors should expect their negotiating leverage to shift the day the Third Circuit rules.