An Appeals Court Just Issued the First Ruling on AI Training and Fair Use. The AI Company Lost.

On September 29, 2026, the US Court of Appeals for the Third Circuit upheld a district court's ruling in favor of Thomson Reuters against Ross Intelligence, a legal-research startup. Per Reuters's own wire report, it is the first US appeals court decision to address a copyright dispute over AI training.

The dispute goes back to 2020. Thomson Reuters accused Ross Intelligence of copying thousands of Westlaw "headnotes," the short editorial summaries that distill the key legal point out of a court opinion, and using them to train a competing AI-powered legal search tool. In February 2025, Judge Stephanos Bibas ruled against Ross. His reasoning, per the reporting: Ross's use of the headnotes was not transformative, because Ross used them for the same purpose Thomson Reuters did, to build a tool that helps lawyers find relevant case law, and aimed the result squarely at Thomson Reuters's own market. The Third Circuit's ruling leaves that result in place. The panel's detailed reasoning was not public at the time of reporting.

Ross shut its product down in 2021, citing the cost of fighting the lawsuit. The company that lost the case no longer operates. The ruling still matters, because it is the first time an appeals court has weighed in on whether training an AI system on someone else's copyrighted work counts as fair use, and the answer here was no.

It is worth being precise about what this ruling does and does not decide. Ross built a search and retrieval tool, not a generative one. It never produced new text that read like Thomson Reuters's. It used the headnotes to get better at pointing lawyers to the right case. The pending suits against OpenAI, Microsoft, Meta, Anthropic, and others mostly involve generative systems that can output text resembling what they were trained on, a materially different fact pattern. Those cases are still working through other courts, and this ruling does not resolve them. What it does establish is a data point against one specific argument AI companies have leaned on: that training a model on copyrighted material is automatically fair use just because the resulting model works differently from the original. Here, a federal appeals court said that argument fails when the resulting product competes directly with the source of the training data and serves the same purpose.

Thomson Reuters's statement, per the reporting, put it plainly: "respecting copyright is essential for fostering innovation while protecting the intellectual property that fuels fiduciary-grade AI solutions."

For anyone building a product on top of an AI model, the practical lesson isn't about Westlaw headnotes specifically. It's about where the legal risk actually sits: in training on content you don't have clear rights to, for a product that competes with or substitutes for the content's original purpose. That risk doesn't go away because the underlying case settles or a particular company shuts down. It sits in the architecture until someone builds around it.

That is the design choice behind ForIntel, Foragentis's intelligence-products line. Every claim in a ForIntel report resolves to a cited, verifiable source row, not to a model trained on an unlicensed corpus with no attribution path back to where the information came from. The question this ruling raises for the rest of the industry, whether an AI product's output can be traced back to sources it had the right to use, is the one ForIntel was built to answer from the start: https://foragentis.com