One Word, Under Seal: Third Circuit Affirms the First Appellate AI-Copyright Ruling in Thomson Reuters v. Ross

One Word: AFFIRMED

On September 29, 2026, the Philadelphia-based Third Circuit Court of Appeals delivered the most consequential copyright decision of the AI era — and it did so in a single word. In Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., a three-judge panel rejected ROSS Intelligence’s argument that its AI-powered legal search engine made fair use of material from Thomson Reuters’ Westlaw platform, affirming a lower-court ruling for the information-services giant. It is the first time a US appeals court has ruled on whether training an AI system on copyrighted works can qualify as fair use.

The twist is almost poetic. The case concerns Westlaw’s “headnotes” — proprietary summaries of points of law in court opinions that sit behind a paywall. The Third Circuit’s actual reasoning now sits behind a seal. The opinion, written by Judge Tamika Montgomery-Reeves, will not be released until the parties propose redactions and the panel decides what, if anything, must be blacked out. What the public received was a one-page judgment holding that boils down to: “AFFIRMED.” As one copyright commentator quipped, it may be the shortest headnote ever written — and, being a government work, freely reproducible without a Westlaw license.

The Long Road to One Word

The case began in 2020, when Thomson Reuters sued ROSS Intelligence, a legal-research startup, alleging it copied thousands of Westlaw headnotes to train a competing AI-based legal search engine. ROSS didn’t copy the headnotes directly into its product; instead, it used them as training data — so-called “Bulk Memos” derived from the headnotes — to teach its engine to identify relevant court opinions in response to user queries. The litigation was punishing enough that ROSS shut down its platform entirely in 2021, citing the costs of defending the suit.

In February 2025, Judge Stephanos Bibas — sitting in the Delaware federal court by designation — delivered a ruling that stunned many copyright observers. He found that at least 2,243 of the copied headnotes were original and copyrightable as a matter of law, that even headnotes quoting court opinions word for word could be protected (via a sculptor analogy that divided commentators), and that ROSS’s use was not transformative because ROSS was building a direct Westlaw competitor. “Ross took the headnotes to make it easier to develop a competing legal research tool,” Bibas wrote. “So Ross’s use is not transformative.” He then certified the case for interlocutory appeal to the Third Circuit, where he himself normally sits.

The June 2026 oral argument previewed the outcome. Judge Emil Bove pressed ROSS’s counsel on what was so “earth-shattering” about the product compared to his own morning Westlaw login, accusing the lawyer of dancing around what made the tool different. ROSS’s attorney, Mark Davies, argued that headnotes simply track the language of public-domain court opinions — “It’s not poetry,” he told the panel. Thomson Reuters’ Dale Cendali leaned on the district court’s findings and urged the court not to reopen the question of whether headnotes are copyrightable at all.

Why One Word Matters So Much

An affirmance with sealed reasoning might seem anticlimactic. It isn’t. Dozens of generative-AI copyright cases are pending around the country — brought by authors, news outlets, music labels, and visual artists against essentially every major AI developer — and Thomson Reuters v. Ross is the first appellate data point any of them have. Trial judges deciding those cases will read this judgment closely, as will the parties. The sword now points, at minimum, in the direction of copyright holders on the core question of whether AI training can be fair use.

There are important caveats. The Third Circuit itself, and Judge Bibas before it, stressed that ROSS’s tool was not generative AI — it did not create new content, it searched and retrieved. The case also involved a head-on competitor using a rival’s proprietary database, which is about the least sympathetic defendant posture imaginable. Generative-AI defendants will argue their facts are different on both counts: their models transform training material into something new, and the finished models do not reproduce the works they learned from.

That is precisely why the sealed opinion’s eventual contents matter more than the judgment. When it is unsealed, close observers will be watching for at least three things. Did the panel embrace Judge Bibas’ sculptor analogy for headnote originality, or did it take the narrower road Judge Montgomery-Reeves hinted at during argument, focusing on headnotes that were not direct quotations? Did it endorse the idea that a merely potential market for AI training data is enough to tip fair use’s fourth factor on market harm — a theory ROSS attacked as circular, since any copyright owner can claim it would have licensed the very use it is suing over? And did the panel confine the “intermediate copying” doctrine to computer code, where it originated? That last answer may be the one generative-AI defendants read most anxiously: their fair-use defenses rest heavily on the argument that training copies are merely a step along the way to a transformative output.

The Licensing Signal

Whatever the reasoning, the market heard the result. Thomson Reuters said it was “pleased” with the ruling and “firmly believes that respecting copyright is essential for fostering innovation while protecting the intellectual property that fuels fiduciary-grade AI solutions.” For an industry that has signed a growing stack of licensing deals with news organizations, publishers, and image libraries, the affirmance strengthens the copyright side of every negotiation still open. Content owners now hold an appellate citation; AI companies hold a reminder that “we only used it as training data” is not, by itself, a legal shield.

The irony of the seal is not lost on anyone following the case. In a dispute about who gets to summarize judicial opinions for profit, the appellate court’s own summary of the law is temporarily withheld from the public it binds. Until the unredacted opinion drops, the entire American AI-copyright landscape is reading one word — and inferring a great deal from it.

ROSS Intelligence no longer exists to feel the loss. The dozens of generative-AI cases that very much do exist are the real audience now, and their judges will care far more about how the Third Circuit got to “AFFIRMED” than about the destination itself. That answer is coming, one redaction at a time.