Appeals Court Says Rival AI Trained on Westlaw Headnotes Was Not Fair Use vs Thomson Reuters

Copying a competitor's work to train an AI that competes with it is not fair use, at least when the competitor is Westlaw. A US appeals court has upheld Thomson Reuters' copyright win against Ross Intelligence, the defunct startup that built an AI legal search engine on material derived from Thomson Reuters' flagship research platform.

The ruling, reported by Reuters on September 29, rejects Ross's fair-use defense. The appellate reasoning remains under seal, so the full logic is not yet public. But the outcome alone matters: coverage describes it as the first US appellate decision in the wave of lawsuits over training AI on copyrighted work.

How a startup ended up training on Westlaw

The fight goes back to 2020, when Thomson Reuters sued Ross for using Westlaw content to build a rival product. Ross had asked for a license. Thomson Reuters refused, because Ross was a competitor.

So Ross went around the wall. It obtained roughly 25,000 "Bulk Memos" from a company called LegalEase, and those memos were built from Westlaw headnotes, the short summaries Thomson Reuters' editors write to capture a court's key points of law. Ross used them to teach its system how to match legal questions to relevant passages.

In February 2025, Judge Stephanos Bibas, revisiting his own 2023 decision, found Ross liable for infringing 2,243 Westlaw headnotes. His reasoning was blunt. Ross took the headnotes to make it easier to develop a competing legal research tool, so the use was not transformative. It served the same purpose as the original, in the same market.

The appeals court has now let that conclusion stand.

A win that arrives years too late for Ross

Ross is not around to feel much of it. The company shut down its platform in January 2021, blaming the cost of the litigation. Its attempt to fight back on antitrust grounds also failed: a federal judge dismissed Ross's antitrust counterclaims in September 2024.

For Thomson Reuters, the decision is a validation of a strategy that leans heavily on owning its content. A company spokesperson said respecting copyright is essential for building "fiduciary-grade" AI solutions, which is a pointed message as the company sells its own AI tools to lawyers on top of the same Westlaw archive.

Why this is not the generative AI ruling everyone is waiting for

The detail that gets lost in the headlines is that Ross was not a chatbot. It was a non-generative search tool that returned existing judicial opinions. It did not write new text, and it did not compete by producing summaries of its own.

That limits how far the ruling travels. The big cases against OpenAI, Meta, Anthropic and others turn on whether training large language models on books, news and code transforms that material into something new. Courts handling those cases have so far treated fair use as highly fact-specific, and nothing in a sealed opinion about a legal search engine settles them.

Still, the logic Bibas used, and the appeals court let survive, is the one generative AI defendants fear most: if your model was built to substitute for the source in its own market, transformation arguments get weak fast. Plaintiffs in the news and publishing suits will cite this case for exactly that point.

It also lands in a season when courts on both sides of the Atlantic are reshaping how tech giants can use what they control, from Google's appeal of EU penalties over search and Android to fights over who pays for the infrastructure AI needs, like Meta's push for tax credits on its AI data centers.

For AI companies that trained on a rival's content without a license, the message from the first appellate court to weigh in is simple. Competing with the thing you copied is the fastest way to lose.