# A Federal Court Just Drew a Hard Line Between AI Training and Theft By Rafiqul Islam Rabbi · AI · Published Sat, 03 Oct 2026 17:14:06 GMT Source: The Current Tribune — https://currenttribune.com/article/ai-fair-use-ruling-thomson-reuters-ross-intelligence A federal appeals court just handed the AI industry its clearest answer yet on a question every lab has been dodging for three years: does feeding a machine someone else’s work to build a rival product count as fair use? For one very specific kind of AI company, the answer is now a hard no. ## What Actually Happened The Third Circuit Court of Appeals affirmed a lower court ruling against Ross Intelligence, a now-defunct legal research startup that trained its AI tool on roughly 2,243 headnotes lifted from Westlaw, the legal database owned by Thomson Reuters. Headnotes are the short, human-written summaries that sit at the top of court opinions in Westlaw’s system, distilling dense rulings into digestible points of law. Ross never licensed them. It scraped them, fed them into its model, and used the resulting system to build a product that competed directly with Westlaw’s own search tools. Thomson Reuters sued back in 2020, years before “AI training data” became a dinner-table phrase. The district court initially ruled in Ross’s favor, then reversed itself in a widely watched 2025 decision. Ross’s appeal to the Third Circuit produced a ruling late last week, and the court’s full written opinion explaining its reasoning only became public days later, once the sealed portions were cleared for release. That unsealing is what turned a decided case back into a live news story, because it let lawyers everywhere finally see exactly how the judges got there. ### The Court’s Reasoning, Broken Down Fair use in copyright law rests on four factors, and the Third Circuit walked through each one methodically: - **Purpose and character of the use:** The court found Ross’s use was only “minimally transformative.” It wasn’t repackaging the headnotes into something new and different — it was using them to build a tool that did the same basic job as the original. - **Nature of the copyrighted work:** Westlaw headnotes involve real editorial judgment and legal expertise, not raw facts, which pushed this factor toward Thomson Reuters as well. - **Amount and substantiality:** Ross didn’t borrow a phrase here or there. It copied entire headnotes, wholesale, which weighed heavily against any fair use defense. - **Effect on the market:** This was the knockout factor. The court described Ross’s product as “substitutive competition” — a direct stand-in for Westlaw, built using Westlaw’s own content, aimed at Westlaw’s own customers. Three of four factors broke against Ross, and in fair use cases, that’s typically the whole ballgame. ## Why This Isn’t the Earthquake Some Headlines Suggest It would be easy to read this as a death blow to generative AI’s entire legal strategy. It isn’t, and that distinction matters enormously for how this plays out. The court went out of its way to separate Ross’s situation from the kind of training that companies like OpenAI, Anthropic, and Google do when they hoover up books, articles, and web text to build general-purpose language models. Ross built something narrow: a direct substitute for one specific commercial product, using that same product’s proprietary editorial content, aimed squarely at stealing that product’s customers. A general-purpose chatbot trained on a vast, diverse corpus of text to generate novel outputs across countless domains is a different animal in the eyes of this court, even if plaintiffs in other lawsuits will absolutely try to argue otherwise. ### Where the Line Actually Sits Strip away the legal jargon and the lesson for AI builders is almost mechanical: - Training on copyrighted material to build something genuinely new and transformative still has a fighting chance under fair use. - Training on copyrighted material to build a near-identical substitute for the thing you copied does not. - Copying whole, discrete, human-authored units of content — like headnotes, rather than diffuse patterns across billions of documents — makes the “amount taken” factor much harder to defend. - If your product’s entire value proposition is “the same thing as the competitor, cheaper,” courts are going to notice. ## The Ripple Effects Nobody’s Talking About Yet Legal research and specialized data products are exactly where this ruling bites hardest. Any startup building an AI tool on top of a paywalled, editorially curated database — financial data terminals, medical coding references, patent search tools, academic citation indexes — just watched a court draw a bright line around exactly that business model. Expect incumbents in those industries to start citing this opinion in demand letters within weeks, not months. It also hands ammunition to publishers and data companies currently negotiating licensing deals with AI firms. Before this ruling, AI companies could credibly argue that any use of copyrighted training data had a decent shot at fair use protection. Now there’s a published appellate opinion saying that argument collapses the moment your product looks and acts like a clone of what you trained on. That changes the leverage in every negotiation happening right now between publishers and AI labs. At the same time, the ruling is a gift to the big foundation model companies currently fighting their own copyright battles in courts across the country. They can now point to this exact opinion and say: see, the court itself drew a distinction, and we’re on the other side of it. Expect OpenAI, Anthropic, Meta, and Google’s lawyers to cite this case liberally — not as a warning, but as cover. ### What Happens to Ross Intelligence Now Ross Intelligence shut down its operations years ago, a casualty of the lawsuit’s financial strain well before this final ruling landed. There’s a certain irony in a case that will shape AI law for years being decided after the company at the center of it had already ceased to exist. The ruling doesn’t resurrect Ross or change its fate — it exists now purely as precedent, a fixed point that every AI copyright case filed from here forward will have to measure itself against. ## What This Means This ruling doesn’t settle the generative AI copyright wars raging in courtrooms from San Francisco to New York. What it does is give everyone involved — judges, plaintiffs’ lawyers, AI company counsel — a concrete, appellate-level example of what a losing fair use argument looks like. The companies that should be nervous are the ones building narrow, substitutive tools on top of someone else’s curated database. The companies that can breathe a little easier are the ones building broad, general-purpose models from diverse sources, at least for now. The bigger story is momentum. Courts are no longer treating “but it’s AI training” as a magic phrase that short-circuits copyright analysis. They’re applying the same four-factor test that’s governed fair use for decades, and increasingly asking a very old, very human question: did you build something new, or did you just take something that wasn’t yours and call it innovation? For one legal AI startup, the court already has its answer. For the rest of the industry, the clock is now running on finding out theirs.