Appeals Court: Training AI on Westlaw Wasn't Fair Use

The James A. Byrne U.S. Courthouse in Philadelphia, headquarters of the Third Circuit Court of Appeals. The unanimous panel affirmed that ROSS Intelligence infringed Thomson Reuters' copyrights when it trained its AI on Westlaw headnotes. Photo: Beyond My Ken, CC BY-SA 4.0, via Wikimedia Commons.

No American appeals court had ever answered the question: is training AI on copyrighted material fair use? This week, the Third Circuit finally did — and the answer went against the AI company. Before Silicon Valley panics, though, here's the twist: the court carved generative AI out of the ruling in a single footnote — and the company that lost shut down five years ago. This might be the strangest landmark ruling the tech industry has ever received.

This week, the Philadelphia-based Third Circuit Court of Appeals affirmed that ROSS Intelligence, a now-defunct legal-tech startup, infringed Thomson Reuters' copyrights when it used Westlaw's headnotes to train a competing AI-powered legal search engine. The opinion had been filed under seal the day before; the unsealed 32-page version landed on September 30, and Reuters carried both the ruling and the unsealing.

Fair use is the defense that every AI lab is counting on. OpenAI, Google, Anthropic, Meta — all of them are facing lawsuits arguing that hoovering up copyrighted text to train models is perfectly legal because training is "transformative." This case was supposed to give the industry its first real answer from a federal appeals court. What it gave them instead was a unanimous panel shrugging and saying, in effect: this one isn't about you.

Update, Oct. 2: the unsealed opinion names names

The Third Circuit's opinion, unsealed September 30, was written by Circuit Judge Tamika Montgomery-Reeves for a unanimous three-judge panel. The court called the dispute "no more than an ordinary copyright case," found Westlaw's headnotes carry a "creative spark" sufficient for copyright, and held that ROSS's copying was "an intermediate step" that was "minimally transformative, at best." The killer line: ROSS copied the headnotes because it was easier — and "unlike necessity, ease is not a justification for copying." A footnote expressly setting generative AI aside answers the Trump administration's statement of interest in the OpenAI litigation. What remains: a Delaware trial on expired copyrights, the remaining headnotes, and damages. ROSS shut its platform in 2021, citing litigation costs — and never showed a headnote to users; the copies went into training memos only.

Six years, 2,243 headnotes, one sealed opinion

The story starts in 2020, when Thomson Reuters sued ROSS Intelligence, accusing it of copying thousands of Westlaw's "headnotes" — short editorial summaries of points of law that Westlaw's editors attach to judicial opinions. ROSS had a clever product: a legal search engine that answered lawyers' questions in plain English. To build it, the company wrote roughly 25,000 practice questions and used the headnotes to write AI training memos for its system, as patent-law professor Dennis Crouch documented on his Patently-O blog this week.

The litigation killed the startup before the law ever got to speak. ROSS shut down its platform in 2021, citing the cost of fighting Thomson Reuters in court. It has been litigating as a corpse ever since — five years of zombie proceedings over a product nobody uses anymore.

In February 2025, Delaware district judge Stephanos Bibas granted Thomson Reuters partial summary judgment, holding as a matter of law that ROSS had infringed 2,243 headnotes and that the copying wasn't fair use. Bibas then certified two questions for immediate appeal: whether the headnotes are original enough for copyright protection at all, and whether ROSS's use of them was fair. On September 29, the Third Circuit answered both questions in Thomson Reuters' favor, unanimously, and told the two sides they had ten days to propose redactions before the opinion went public.

Nobody fought the unsealing. The opinion came out Wednesday morning, written by Circuit Judge Tamika Montgomery-Reeves, with Judges Restrepo and Bove joining. And it contains a line that legal Twitter will be quoting for years:

"Under ROSS's framing, this case appears to concern the future of AI legal technology," Montgomery-Reeves wrote. "But appearances can be deceiving. In truth, this is no more than an ordinary copyright case; we must determine whether Thomson Reuters' materials are sufficiently original for copyright protection and whether ROSS's use of those materials was fair."

Translation, from a judge who has clearly read one too many breathless AI-law hot takes: calm down. This is a copying case.

"Ease is not a justification for copying"

The reasoning is where the industry should pay attention, because it's blunt. ROSS argued the kind of intermediate-copying defense that AI labs love: we copied the headnotes not to show them to users but as a training input, the same logic that won in Google v. Oracle. The court said no — and the distinction it drew is the one every AI lawyer in America is now underlining in red.

ROSS lost on three of the four fair-use factors. It won only the second, the nature of the copyrighted work, which Crouch dryly notes is "probably the least interesting" of the four. The panel found the use was highly commercial and minimally transformative, because ROSS and Thomson Reuters were doing the same thing with the material: helping researchers find judicial opinions. As the court put it, "Thomson Reuters' materials possess a creative spark, and ROSS aspired to be a direct competitor by using them for a highly similar purpose."

And here's the detail that sank the defense completely: ROSS didn't need the headnotes. The underlying judicial opinions are public domain — free for anyone to use. ROSS copied Westlaw's editorial summaries because, in Crouch's reading of the opinion, that was easier. "Unlike necessity, ease is not a justification for copying," he writes, and that's the sentence that should be framed on the wall of every AI lab's legal department.

Oh, and one more delicious fact from Crouch's analysis: ROSS never showed a single headnote to its users. The headnotes were copied to write training memos, and the finished product returned only passages from the judicial opinions themselves. The copying was invisible — and it still wasn't fair use. For the labs betting that "the user never sees the training data" is a magic shield, that's a genuinely uncomfortable paragraph.

The footnote that matters more than the ruling

Now the carve-out. The Third Circuit went out of its way to say this case is not about generative AI — the kind that creates new content, which is what OpenAI, Anthropic and Google actually build. In a footnote, the court said the concerns raised by the U.S. Department of Justice in its statement of interest in a copyright lawsuit against OpenAI "do not apply here." ROSS's engine was a retrieval tool, not a generator. It found existing text; it didn't make new expression.

That's why the Trump administration's DOJ filing — a statement of interest in the OpenAI litigation — gets a footnote and not a chapter. The court is telling the industry, politely but firmly: come back when your technology is actually in front of us.

So who won? Thomson Reuters, obviously, and the quote its spokesperson gave Reuters on Tuesday — before the opinion was even unsealed — is worth reading in full, because it's the entire content industry's position in one sentence: the company "firmly believes that respecting copyright is essential for fostering innovation while protecting the intellectual property that fuels fiduciary-grade AI solutions." ROSS and its lawyer, Mark Davies of White & Case, didn't respond to Reuters' requests for comment. There is, to be fair, not much left to say when you're a dead company that just lost unanimously.

(Disclosure worth making, since Reuters made it too: Thomson Reuters is the parent company of Reuters News. We checked both of Reuters' stories on this ruling against the docket summary and Crouch's analysis before writing — the dates, the panel, the 2,243 headnotes, and the ten-day redaction window all line up.)

Law books in a law school library. ROSS wrote about 25,000 practice questions and used Westlaw's headnotes to write AI training memos for its legal search engine. Photo: Tony Webster, CC BY 2.0, via Wikimedia Commons.

Why every AI lab is reading this with a highlighter

The ruling lands in the middle of the most expensive copyright fight in tech history. Authors, news outlets and music labels have filed dozens of lawsuits against AI companies over training data. The industry's best precedent so far came in June 2025, when then-presiding Judge William Alsup ruled in Bartz v. Anthropic that training Claude on lawfully obtained books qualified as fair use — transformative, in the legal sense. But Alsup also found that Anthropic had crossed a line by building a central library of more than 7 million pirated books, even ones it never trained on. That half of the ruling produced the largest copyright settlement in U.S. history. The numbers: $1.5 billion, final approval from Judge Araceli Martínez-Olguín on July 20, 2026, roughly 482,000 works at about $3,000 a book — and 91% of eligible rightsholders filed claims.

Anthropic, for its part, called the outcome a win. Deputy general counsel Aparna Sridhar told Reuters the company had "reached this settlement in 2025, after the court's landmark ruling that training AI on books is fair use under copyright law — which remains the law today."

That is still the law — but only in the Northern District of California, only at the trial-court level, and only on generative-AI facts. The ROSS ruling doesn't touch that. It's a different circuit, a different technology, a dead defendant. But it does something Alsup's ruling never did: it gives the copyright owners an appellate win to wave around, and it does it with language about competitive substitution that plaintiffs' lawyers in the New York Times' case against OpenAI are going to quote verbatim.

The pattern the court punished — copying someone else's editorial work to build a competing product — is uncomfortably close to what AI search and AI answers do to publishers every day. Which is why Google's quiet decision to start paying publishers for AI search content, revealed this week, suddenly looks less like generosity and more like risk management. When the courts start saying "creative spark" and "direct competitor" in the same sentence, licensing starts looking cheap. Meanwhile, another government evaluator just published its own incident report — Britain's AI Security Institute had to freeze cyber tests after its agents took "sustained action against real people", and has now restarted them with no more internet access.

There's also a legal-industry subplot that nobody should miss. Thomson Reuters isn't just a news company — Westlaw is the legal profession's operating system, and the same company is watching OpenAI's GPT-6 push into Big Law with the focus of a firm that just spent six years proving it will litigate. The legal-tech arms race now has a court-tested incumbent that has demonstrated, twice, that it will sue competitors into the ground. Startups building on top of other people's editorial databases just got a very expensive warning.

Florida's AG is separately trying to halt OpenAI's frontier training in court — the courts-vs-labs pressure is building on more than one front.

Our read: the industry wanted a rule and got a map with the hard part blank

Let's be honest about what this ruling is. The AI industry spent six years waiting for an appeals court to say something — anything — about fair use and AI training. What it got is a ruling against a company that no longer exists, on pre-generative technology, with an explicit footnote telling the generative-AI docket that this case is not about them. As one legal analysis put it this week: the industry wanted a rule, and got a map with the hard part left blank.

Brandon Butler, the executive director of the Re:Create coalition, which filed an amicus brief in the case, put it more colorfully: "As it stands, this first judicial walk through the AI fair use labyrinth has hit a dead end."

He's right, and it's worth saying plainly: this is a botched precedent. Not botched by the judges — the opinion is careful, unanimous, and clearly written — but botched by history. The one case that made it to an appeals court first is the one case that can't help anyone, because ROSS's facts are from another era. The questions that actually matter — is training a generative model on copyrighted text transformative? does a model that competes with publishers fail fair use? — are still sitting in trial courts in New York and California, years from an appellate answer.

What remains in Delaware is the unglamorous tail of the case: whether some of Thomson Reuters' copyrights have expired, the claims on the remaining headnotes, and damages. The money phase. Nobody will read those filings except the lawyers billing for them.

The real action is elsewhere. Watch the New York Times' suit against OpenAI and Microsoft, the music labels' cases, and whatever the next appeals panel does when a generative-AI defendant finally gets in front of it. The Third Circuit just told that panel, in a footnote, that this case doesn't decide anything for them. Somewhere in San Francisco and Redmond, general counsels are exhaling. Somewhere in a newsroom, a media lawyer is printing the "creative spark" paragraph and taping it to the wall.

See also the SpaceXAI antitrust fight over AI's biggest labs — the companion case in the courts-are-redrawing-AI's-boundaries beat.

ROSS v. Thomson Reuters: the questions everyone asks

What is a headnote? Short editorial summaries of points of law attached to court opinions — Westlaw editors write them, and the court found 2,243 of them original enough for copyright.

Why doesn't this decide the OpenAI and Anthropic cases? ROSS's tool retrieved existing text; it "cannot generate original expression." The court expressly set generative AI aside in a footnote.

What happens next? Back to Delaware: expired-copyright defenses, the remaining headnotes, and damages. En banc or certiorari petitions are possible but none have been filed.

How does this sit beside Bartz v. Anthropic and Kadrey v. Meta? Those are district-level generative-AI rulings. This is the first appellate word on AI training — and on retrieval facts, it went against the AI company.

Sources: Reuters' reporting on the September 29 sealed ruling and the September 30 unsealed opinion (Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 25-2153); Dennis Crouch's analysis at Patently-O; dig.watch; TechTimes; eWeek and Startup Fortune on the $1.5 billion Anthropic settlement and its July 20, 2026 final approval.

Related on TechPulse USA: OpenAI's GPT-6 push into Big Law — the legal industry's AI land grab, with Thomson Reuters watching from the incumbent's seat; and Google's deal paying publishers for AI search content — when courts talk about "creative spark," licensing starts looking cheap.

By Daniel Rao, TechPulse USA — covering the companies, models, and money moves shaping artificial intelligence. We read the fine print so you don't have to.