The Third Circuit Just Redrew the Rules for AI Copyright

The case that changes everything

A three-judge panel of the Third Circuit Court of Appeals just handed down what may be the most consequential copyright ruling in the history of artificial intelligence. The case name is Thomson Reuters v. ROSS Intelligence. The outcome matters to every company that has ever fed copyrighted text into a machine-learning pipeline — OpenAI, Google, Anthropic, Meta, and hundreds of smaller players who have been operating on the assumption that training an AI on scraped or licensed data was defensible under fair use.

It is not, at least not anymore, not in the Third Circuit and potentially not anywhere else.

The ruling came down on a Wednesday in June 2026. Judge Tamika Montgomery-Reeves, a Biden appointee, wrote for the panel. She was joined by L. Felipe Restrepo, appointed by Barack Obama, and Emil J. Bove III, appointed by Donald Trump. The bipartisan composition of the panel is worth noting because this was not a close call along ideological lines. All three judges agreed: the headnotes in Westlaw’s legal database were copyrightable, and ROSS Intelligence had not used them fairly.

How the case got here

The story begins in 2020, when Thomson Reuters filed suit against ROSS, a startup that had built a competitive legal-research AI product. Westlaw is one of the most expensive and widely used legal-database platforms in the world. Its headnotes — short summaries that precede judicial opinions and flag the key legal issues — are among its most valuable assets. ROSS did not get access to those headnotes through Westlaw. Instead, it routed through a third-party service called LegalEase Solutions after Westlaw refused to deal with it.

ROSS then used the headnotes for one purpose: training its own AI to produce a competing product. That distinction proved fatal to its fair-use defense.

In 2025, a federal judge in Philadelphia granted partial summary judgment to Thomson Reuters and denied it to ROSS. The lower court found that Westlaw’s headnotes carried the requisite “creative spark” — the editors had to select, arrange, and summarize legal reasoning in ways that went beyond the raw judicial opinions themselves. The court compared the editorial process to a sculptor chiseling a block of marble. ROSS appealed.

What the Third Circuit actually decided

The Third Circuit’s opinion is notable for how plainly it states its reasoning. Montgomery-Reeves wrote that the case “is no more than an ordinary copyright case.” That framing is deliberate. The court declined to treat this as a novel question about the future of AI technology and instead applied traditional copyright doctrine to a modern input.

On the question of originality, the panel agreed with the lower court. Each headnote had to stand on its own — conveying enough factual background and reasoning to make sense without reference to the full opinion — while still accurately reflecting the passage it summarized. That required judgment. That judgment qualified as a minimal creative spark. The headnotes were copyrightable.

On fair use, the panel found ROSS’s argument thin. The court acknowledged that training an AI on text is “arguably” a slight departure from simply reprinting the text. But the difference was only in the means, not the end. ROSS was building a direct competitor in the same market. The court also found that ROSS’s use harmed Westlaw’s value in the legal-research platform market and in the derivative market for licensing headnotes to train AI — a point that opens the door to future damages claims beyond this case.

The holding is narrow in its facts but expansive in its logic: using copyrighted material to build a competing product through AI training does not qualify as fair use.

Why this goes far beyond legal tech

ROSS is now defunct. The immediate parties to this case will likely never appear in court again. But the precedent lives.

Every AI company that has trained models on copyrighted books, news articles, software code, or other protected text can now point to this decision and understand that the fair-use shield it once relied on has cracked. The Third Circuit applied four-factor fair-use analysis — purpose and character of the use, nature of the copyrighted work, amount and substantiality of the portion used, and effect on the potential market — and found ROSS lost on at least two of them, possibly three.

The purpose-and-character factor is the one most AI companies will contest. Many argue that training an AI is transformative because the model learns patterns rather than reproducing the original expression. The Third Circuit did not reject that argument outright. It called the transformation “minimal at best” and noted the evidence showed ROSS used the headnotes for the same purpose Westlaw intended — legal research. That factual finding may not bind other courts facing different use cases. A model trained to generate creative writing from copyrighted novels sits in a different position than one trained to replicate legal analysis.

But the market-harm factor is where the ruling bites hardest across the industry. The Third Circuit explicitly recognized a potential derivative market for licensing copyrighted material to AI trainers. That is a marketplace that did not meaningfully exist before this litigation, and the court’s acknowledgment of it changes the calculus for every company that trained without a license.

Who wins, who loses

Thomson Reuters wins a narrow victory and a broad precedent. Its Westlaw business is now protected from direct copycat AI products trained on its proprietary headnotes. The company may also pursue damages from ROSS, though the startup is defunct, which limits recovery. The more significant prize is the precedent itself — a binding appellate ruling that any other AI company suing over similar conduct will cite.

ROSS loses, definitively. The company is already dead, so the loss is reputational and financial, not operational.

The broader AI industry loses something subtler: the assumption that fair use would absorb the cost of training data. For years, companies have operated on the belief that even if their training practices were legally questionable, the doctrine would protect them. The Third Circuit’s ruling undermines that assumption in the most visible way possible. It does not settle every question — the boundaries of transformative use remain unclear — but it draws a line that no AI trainer can safely cross without a license or a compelling factual distinction.

What happens next

The immediate aftermath will involve damage claims and potentially a settlement from Thomson Reuters against ROSS’s successors or insurers. The longer-term aftermath will reshape the economics of AI development.

Companies that built products on unlicensed copyrighted data now face uncertainty about their exposure. Some may negotiate retroactive licenses. Others may bet that courts outside the Third Circuit will draw the line differently, particularly on the transformative-use question. A few may simply absorb the risk and hope Congress steps in — a legislative solution that remains unlikely in the near term.

The ruling also signals that the judiciary, not just lawmakers, is willing to apply traditional copyright doctrine to AI without creating special exemptions. That is a posture that favors rights holders and raises the cost of training data for everyone else.

For now, the Third Circuit has spoken. AI companies that ignored copyright in their training pipelines should listen.