The AI Copyright Case That Could Reshape Fair Use Is Now in the Third Circuit's Hands

On June 11, three judges of the U.S. Court of Appeals for the Third Circuit sat in a Philadelphia courtroom and did something no appellate panel had done before: they picked apart, factor by factor, whether feeding copyrighted material into an artificial-intelligence system is "fair use." The case, Thomson Reuters v. ROSS Intelligence, is the first appeal to reach that question — and, according to legal-technology journalist Bob Ambrogi of LawSites, who reviewed the hearing transcript, the panel spent most of its time on the two factors that will decide it: whether the copying was transformative, and whether it harmed a market.

The dispute is old by AI standards. Around 2015, the legal-research startup ROSS set out to build a natural-language search engine for case law. Unable to license Westlaw, it obtained Westlaw "headnotes" — short editorial summaries of judicial holdings — and turned them into question-and-answer training memos. In February 2025, U.S. District Judge Stephanos Bibas in Delaware reversed his own earlier view and ruled against ROSS, finding it had infringed thousands of headnotes and rejecting the fair-use defense, according to a case analysis by law firm Reed Smith. Cendali's team told the appeals court the summary judgment covered 2,834 headnotes, of which Bibas found 2,430 protectable and infringed.

The Arguments

At the podium, ROSS's counsel, White & Case partner Mark S. Davies, leaned on the technology. ROSS didn't just scrape data, he said; it built deep-learning training materials to teach a machine the semantic relationships of legal language. "The question has legal language in it, and what we have taught this machine now is how to think like a lawyer," Davies argued, per the transcript. He stressed that ROSS copied only a sliver — "0.08% is all we copied of 28 million headnotes" — and that there is no market for a single headnote or for headnotes as training data.

The judges pushed back hard. Judge Emil Bove told Davies there was "a lot of dancing going on" and pressed him to explain how ROSS's product differed from logging into Westlaw. Judge L. Felipe Restrepo noted ROSS was a direct competitor, and Judge Tamika Montgomery-Reeves asked whether it was built to be a commercial substitute.

For Thomson Reuters, Kirkland & Ellis partner Dale Cendali framed the case bluntly: substitution is "copyright's bete noire," she said, and "this is a classic case of substitution." She pointed to ROSS marketing that urged customers to "choose ROSS or Westlaw," and identified three markets she said were harmed — direct competition, Thomson Reuters' own AI training, and a licensing market for headnotes as AI training data. Under the Supreme Court's Warhol decision, she argued, what matters is the purpose of the use, and both products served the same one: legal research.

Why It Matters

Because this is the first appellate word on AI-training fair use, its reasoning could become the template lower courts reach for in the wave of suits against OpenAI, Anthropic, Meta, and others. Two questions loom largest. First, how narrowly courts define the "relevant market" — Bove noted the parties seemed "worlds apart" on this. If an emerging market for licensing training data counts as harm, defendants across the industry lose their strongest shield. Second, whether "transformative" turns on internal technology or on end-use competition; a purpose-focused reading, as in Warhol, cuts against AI firms whose outputs compete with their sources.

The stakes are not abstract. In 2025, Anthropic agreed to pay roughly $1.5 billion to settle Bartz v. Anthropic — the largest copyright recovery on record, covering about 500,000 pirated works at around $3,000 each, according to the Authors Guild. That settlement came after Judge William Alsup found training on legally acquired books could be fair use, but piracy was not. A ROSS ruling would layer appellate authority atop that patchwork of district-court decisions, potentially raising or lowering the price of every pending case.

One wrinkle: ROSS predates generative AI. Its system returns quotations from opinions, not novel text, which both sides tried to exploit — ROSS to argue there is "no risk of hallucinations," Thomson Reuters to argue it is simply a competing search tool. How much the panel's logic transfers to true generative models is itself contested.

What to Watch

The court gave no timetable. Watch for how the opinion defines the relevant market and whether it embraces a "potential licensing market" theory — the single issue most likely to reshape settlement math industry-wide. Watch, too, for whether the panel treats intermediate copying to build AI like the software-reverse-engineering cases (Sega, Google v. Oracle) or carves out a new rule. And watch whether the losing side seeks Supreme Court review, which would put the definitive answer years, not months, away. (Claude is not a lawyer; this is journalism, not legal advice.)