Thomson Reuters v. Ross Intelligence — Third Circuit Rejects Fair Use for AI Legal-Research Training

Background

Thomson Reuters and West Publishing operate Westlaw, whose lawyer-editors write headnotes summarizing points of law from judicial opinions and organize them through the West Key Number System. Ross Intelligence sought to build a competing AI-assisted legal-research platform. It hired a contractor to create about 25,000 training memoranda using Westlaw headnotes as questions and passages from public judicial opinions as labeled answers.

Ross’s system was not generative AI: it returned existing passages from judicial opinions in response to natural-language questions. The Delaware federal court held that 2,243 copied headnotes were sufficiently original for copyright protection and that Ross’s use was not fair. The Third Circuit accepted an interlocutory appeal. LexSummary previously covered the district court’s 2025 ruling.

The Court’s Holding

The Third Circuit affirmed. Each headnote cleared copyright’s low originality threshold because editors made choices about which legal points mattered, how to frame them, and how much factual context to include. Editorial guidelines constrained the work but did not reduce it to a mechanical process, and the court distinguished uncopyrightable judicial opinions from West’s independently written summaries.

Ross also failed to establish fair use. Its copying was commercial and only minimally transformative because both companies used the headnotes to help legal researchers find responsive judicial material. The intermediate step of training a model did not create a sufficiently different ultimate purpose, and copying the headnotes was convenient rather than necessary because the underlying opinions were freely available.

Although the factual nature of the headnotes slightly favored Ross, the other factors favored Thomson Reuters. Ross copied complete headnotes on a large scale, sought to create a Westlaw substitute, diminished the headnotes’ value as a feature attracting subscribers, and usurped a developing market for licensing editorial content as AI training data. The court stressed that its record involved a non-generative search tool built as a direct commercial substitute; it did not purport to resolve every question about generative-AI training.

Key Takeaways

• Legal headnotes can be copyrightable even though they summarize uncopyrightable judicial opinions.
• Using copied material at an intermediate AI-training stage is not automatically transformative.
• Courts will compare the ultimate commercial purposes of the original and challenged uses, especially when the products compete directly.
• A developing licensing market for AI training data can count in the market-harm analysis.
• The opinion distinguishes non-generative retrieval technology from models that generate new expression.

Why It Matters

This is a major federal appellate decision on copyrighted material used to train an AI system. Its practical lesson is narrower than a blanket rule about AI: copying is particularly vulnerable when the source material is used to build a close commercial substitute, the material could have been licensed, and the copier could have trained from public-domain inputs instead.

The decision also sharpens the role of purpose after the Supreme Court’s Warhol ruling. A technologically different process does not by itself establish a different copyright purpose. Developers and content owners will likely focus on whether model outputs substitute for the source product, whether protected expression was needed, and whether a functioning licensing market exists.

Surfaced via Law360 IP.