Court Rejects AI Training Defense

Philadelphia — A federal appeals court has ruled that a startup called ROSS Intelligence broke copyright law when it used summaries written for Thomson Reuters’s Westlaw legal-research service to train its own artificial-intelligence search tool. The decision is the first by a U.S. federal appeals court on whether using someone else’s copyrighted work to train AI is allowed under the legal doctrine called fair use.

The U.S. Court of Appeals for the Third Circuit, which hears appeals from federal courts in Pennsylvania, New Jersey and Delaware, filed its opinion on 29 September. It was first kept under seal and became public the next day. Judge Tamika Montgomery-Reeves wrote for a three-judge panel that also included Judges L. Felipe Restrepo and Emil J. Bove III.

What was copied

Westlaw is one of the main services lawyers use to find court decisions. Above many decisions it shows “headnotes,” short summaries written by Thomson Reuters editors that state each point of law the court decided. The judges’ opinions themselves are public and free for anyone to copy. The headnotes are the company’s own writing.

ROSS was founded by three University of Toronto computer-science students who wanted to build a tool that answered plain-language legal questions by pulling up the right passages from court opinions. To teach its program which passages answer which questions, ROSS hired a company called LegalEase Solutions to write about 25,000 training memos. The court found that the memo writers used thousands of Westlaw headnotes to frame the questions. The case covers 2,243 headnotes that the trial judge found had been copied.

ROSS’s tool was not “generative” AI. It did not write new text; it only returned passages from existing opinions. ROSS advertised itself as a rival to Westlaw at similar prices. The company has since shut down.

What fair use means

Copyright gives the author of an original work control over copying it. Fair use is the exception written into federal law that lets people copy without permission in some cases, such as criticism, news reporting, teaching or research. Courts weigh four things: why the material was copied and whether the new use adds something different, what kind of work was copied, how much was taken, and whether the copying hurts the original’s market.

What the court decided

First, the panel held that the headnotes are protected by copyright. The legal bar for originality is low, and the court said each headnote reflected editors’ choices about which points mattered and how to say them, which it called a “creative spark.” Headnotes are not law, the court said; the opinions are, and those remain free.

Second, the panel held that ROSS’s copying was not fair use. Both companies used the headnotes for the same end, helping people find relevant law, so ROSS’s use was “minimally transformative, at best,” Montgomery-Reeves wrote. ROSS could have trained on the free opinions but chose the headnotes because they were easier. “Unlike necessity, ease is not a justification for copying,” she wrote. The court also found that the copying hurt Westlaw’s business and a growing market for licensing such material as AI training data.

“Under ROSS’s framing, this case appears to concern the future of AI legal technology,” the opinion says. “But appearances can be deceiving. In truth, this is no more than an ordinary copyright case.”

The ruling upholds a February 2025 decision by Judge Stephanos Bibas, an appeals judge who heard the case as a trial judge in federal court in Delaware. He had ruled for Thomson Reuters on those two questions before any trial, finding the key facts were not in dispute. That decision was appealed early, before the rest of the case was finished.

How far it reaches

The ruling has limits. The court noted that generative AI systems, the kind behind chatbots, raise different questions, and it pointed to other cases where courts and the U.S. Justice Department have treated training such models differently. Bob Ambrogi of the legal-technology site LawSites wrote that the copying happened before generative AI became widespread and was done to build a direct competitor, which may keep the decision’s reach narrow.

Still, the ruling gives authors and publishers an appeals-court precedent that training data has value, and it signals that copying a rival’s material to build a substitute product will be hard to defend.

“The appellate court agreed that Westlaw’s editorial content — created and maintained by our attorney editors — is protected by copyright and that ROSS’ copying of Thomson Reuters content did not constitute fair use,” a Thomson Reuters spokesperson told Courthouse News.

ROSS’s lawyer, Yar Chaikovsky of White & Case, said the company would ask the U.S. Supreme Court to take the case. “We respectfully disagree with the Third Circuit’s decision and believe it creates continued uncertainty around the application of copyright law to AI model training,” he said.

What happens next

ROSS plans to petition the Supreme Court, which chooses which cases it hears. Unless the justices step in, the ruling stands in the Third Circuit, and other courts weighing AI training lawsuits can look to it.