Thomson Reuters v. Ross Intelligence is the first U.S. court decision ruling that copying copyrighted content to train an AI system is not fair use. The court found Ross Intelligence infringed 2,243 Westlaw headnotes to build a competing legal research tool. Every attorney advising AI companies needs to understand this ruling because it's the first concrete answer to the AI training data question.
Background
Thomson Reuters owns Westlaw, the dominant legal research platform. Its editorial staff writes headnotes, short summaries of legal principles from court opinions, that help attorneys find relevant case law. These headnotes are copyrighted works. Ross Intelligence was a startup building an AI-powered legal research tool designed to compete with Westlaw.
Ross needed training data to teach its AI what constitutes a good legal research result. It contracted with a third party, LegalEase Solutions, to provide that data. LegalEase's method was straightforward: it used Westlaw headnotes as the basis for the training content. Ross's AI learned what results to return by studying what Westlaw's editors had already identified as legally significant.
Thomson Reuters filed suit in the District of Delaware in May 2020, under Case No. 1:20-cv-00613, alleging Ross copied 2,243 copyrighted headnotes to train its competing product. The case was assigned to Judge Stephanos Bibas, a Third Circuit judge sitting by designation.
What Happened
The central question was whether Ross's use of Westlaw headnotes constituted fair use under copyright law. Ross argued its use was transformative: it wasn't republishing the headnotes but using them to train an AI system, a fundamentally different purpose. Thomson Reuters countered that Ross used the headnotes for exactly the same purpose Thomson Reuters created them, to help users find relevant legal information.
The court conducted a detailed fair use analysis across all four statutory factors. On the critical first factor (purpose and character of the use), the court found Ross's use was not transformative. Ross used the headnotes to build a product that served the same function as Westlaw: legal research. Taking someone's work to build a competing product that does the same thing isn't transformation. It's substitution.
On February 11, 2025, the court granted partial summary judgment to Thomson Reuters, finding that Ross's copying of 2,243 headnotes was not fair use. Ross is appealing to the Third Circuit, but the district court's reasoning has already sent shockwaves through the AI industry.
The Ruling
Judge Bibas held that Ross Intelligence's copying failed the fair use test. The first factor (purpose and character) weighed against Ross because its AI product served the same purpose as the original headnotes: helping users find relevant case law. The use wasn't transformative. The second factor (nature of the work) also weighed against Ross because headnotes involve creative editorial judgment, not mere factual compilation.
The third factor (amount used) was neutral to negative: Ross copied 2,243 complete headnotes, taking the entirety of each work. The fourth factor (market effect) strongly favored Thomson Reuters because Ross built a direct competitor using the copyrighted content of the company it was competing against.
The court was clear: using copyrighted material to train AI that directly competes with the copyright holder is not fair use. This wasn't a close call on the facts. Ross took Westlaw's product to build a Westlaw competitor.
Outcome: The court granted partial summary judgment to Thomson Reuters, finding that Ross's copying of Westlaw headnotes was not fair use. Ross's use was not transformative because it served the same purpose as Thomson Reuters's original work.
Why This Case Matters
This is the first U.S. court ruling directly addressing fair use in AI training, and the answer was no. While the case involved a relatively narrow AI application (search result ranking, not generative AI), the reasoning applies broadly. The court's framework for analyzing whether AI training constitutes fair use will be cited in every subsequent case.
The ruling's most important principle: if your AI product serves the same purpose as the copyrighted content you trained it on, that's not transformative use. This distinction matters enormously for the bigger cases. NYT v. OpenAI involves AI that can reproduce news articles. Andersen v. Stability AI involves AI that generates images in artists' styles. In both cases, the AI outputs compete with the originals.
The interlocutory appeal to the Third Circuit adds another layer. If the appellate court affirms, it becomes circuit precedent. If it reverses, it creates a split that other circuits and the Supreme Court will eventually need to resolve. Either way, Thomson Reuters v. Ross is now the starting point for every AI training data fair use analysis.
Lessons for Attorneys
For attorneys advising AI companies: this ruling means 'we used it for training, not republishing' is not enough to claim fair use. The court looked at whether the end product competes with the original. If your client's AI tool does the same thing the copyrighted content was designed to do, the fair use argument is weak. Audit your clients' training data sources and competitive positioning now.
For attorneys advising content creators: this ruling gives your clients real leverage. If an AI company trained on your client's copyrighted content to build a competing product, Thomson Reuters v. Ross is your precedent. The key is showing the competitive relationship between the original content and the AI product.
For legal research specifically: this case confirms that Westlaw's headnotes are copyrightable and that using them to train competing AI tools is infringement. Any legal tech startup building AI-powered research tools needs to ensure its training data is properly licensed or independently created. The days of treating legal databases as free training data are over.
The Bottom Line
Thomson Reuters v. Ross is the first U.S. ruling that copying copyrighted content to train AI is not fair use when the AI product competes with the original. It's now the baseline precedent for every AI training data copyright dispute, and attorneys on both sides of the AI industry need to understand its framework.
AI-Assisted Research. This piece was researched and written with AI assistance, reviewed and edited by Manu Ayala. For deeper takes and the perspective behind the research, follow me on LinkedIn or email me directly.