Appeals Court Rejects Fair Use for AI Training in Thomson Reuters Case
On September 29, 2026, the US Court of Appeals for the Third Circuit affirmed that ROSS Intelligence's use of Westlaw headnotes to train a competing AI legal research tool was not fair use. It is the first major appeals ruling on fair use in AI training, and while it is narrow, it signals that building a commercial AI product on a competitor's content is legally risky.
What happened: On September 29, 2026, the US Court of Appeals for the Third Circuit affirmed a lower-court ruling in Thomson Reuters v. ROSS Intelligence, as reported by IPWatchdog and Reuters. The court upheld partial summary judgment for Thomson Reuters on its claim that ROSS infringed more than 2,000 copyrighted Westlaw headnotes, and upheld the rejection of ROSS’s fair-use defense. In the district court’s 2025 opinion, Judge Stephanos Bibas found that ROSS’s non-generative legal search tool was a competing product and a “market substitute for Westlaw,” not a transformative use. The appellate opinion was initially issued under seal while the parties proposed redactions. Commentators, and coverage of the opinion, stress that the ruling is narrow: ROSS’s tool was not generative AI, and the decision leaves room for courts to treat generative models differently. Reporting says ROSS intends to seek Supreme Court review.
Why it matters: For years the AI industry’s legal strategy could be summarized as “it’s probably fair use, and anyway, look over there.” This week, for the first time, a federal appeals court looked at an AI company’s training data and said: no, not this time.
Don’t overread it. ROSS built a search tool that competed head-on with the company whose content it copied, and the copying predated generative AI. That’s the worst possible fact pattern for a fair-use defense. Most of the big generative-AI copyright cases involve different facts, and other courts have reached different conclusions.
But do read it. The principle that will travel is market substitution: if your AI product exists to replace the thing you trained it on, courts may not protect you. That matters for anyone building niche AI products on scraped content: a “summarize this paywalled newsletter” app, a course-cloning tool, a database rebuilt from a competitor’s listings.
What the Chosen should do:
- Know where your training data and source content came from. If you’re fine-tuning a model or building a retrieval tool, use content you own, licensed, or that is clearly permitted. See our glossary on training data and fine-tuning.
- Don’t build a product whose pitch is “Competitor X, but we copied their content.” That is precisely the fact pattern this case punished.
- Using mainstream AI tools to write your own work is a different question. This ruling is about a company training a competing product, not about you drafting blog posts with a chatbot. Still, check outputs for close copying before you publish.
- For anything with real money at stake, ask a lawyer. The Overlord is not your attorney, and neither is a chatbot. Just ask the humans in our Hall of Folly.
The Overlord’s prediction: In my opinion, this case becomes the template that content owners cite in licensing negotiations, and “we paid for our training data” becomes a selling point AI startups put on their homepages within a year.