Third Circuit rules AI training on Westlaw headnotes wasn't fair use — Coral Wire
Legal · Thu, Oct 1 2026 · Breaking · Coral Index 62/100

Third Circuit rules AI training on Westlaw headnotes wasn't fair use

The panel found ROSS Intelligence's use "minimally transformative at best" and upheld summary judgment for Thomson Reuters

Third Circuit rules AI training on Westlaw headnotes wasn't fair use
Coral’s read

This is the ruling authors have been waiting for without knowing its name. A federal appeals court just said, in plain terms, that taking someone's creative work to train a competing AI product is copyright infringement, not fair use. ROSS tried to dress this up as a case about the future of AI — the judges weren't having it. "An ordinary copyright case," they called it. That framing is the whole ballgame.

Here's the sentence every AI company's legal team is rereading this week: "This is no more than an ordinary copyright case."

That's U.S. Circuit Judge Tamika Montgomery-Reeves, writing for a Third Circuit panel that just rejected ROSS Intelligence's fair use defense — and in doing so, handed authors and publishers their clearest win yet in the fight over AI training data.

The backstory: Thomson Reuters sued ROSS in 2020, claiming the now-defunct startup used a third party, LegalEase Solutions, to get access to Westlaw's headnotes after Westlaw refused ROSS direct access. ROSS then trained its AI on those headnotes to build a competing legal research product. A federal judge in Philadelphia granted Reuters partial summary judgment in 2025. ROSS appealed. The Third Circuit just said no, twice over.

First, on originality: the panel agreed that Westlaw's editors showed a "creative spark" in selecting and arranging the headnotes — enough to make them copyrightable, even though the underlying judicial opinions aren't. Judge Montgomery-Reeves compared it to a sculptor chiseling marble. The raw material is public domain. What Westlaw's editors did with it is not.

Second, and this is the part that matters far beyond legal publishing: the panel found ROSS's use "minimally transformative at best." ROSS argued that training an AI was a meaningfully different use than Westlaw's own research platform. The court didn't buy it. Training the AI was just "an intermediate step" toward the same end — a competing research tool. And because ROSS's product would hurt Westlaw's value in both its existing market and the "potential derivative market of licensing headnotes to train AI," fair use failed on that count too.

Here's what this means for you. If you've been watching the pile of AI copyright suits from writers, publishers, and news organizations with a sinking feeling that "fair use" is a magic word that makes any training data free for the taking — this ruling says otherwise, at least in the Third Circuit. Two of the three judges on this panel were appointed by presidents from different parties, and both signed on. That's not a partisan split opinion. That's a bench reading copyright law the same way regardless of politics.

Nothing here puts money in your pocket today. But precedent is the currency those other cases run on. If you're a party to — or watching — any AI training litigation, read this ruling's fair use section now, not when your own case cites it. And if your agent or publisher hasn't flagged this to you yet, ask them this week what it means for your own licensing leverage.

Why it matters to you: A federal appeals court has now rejected the argument — used by AI companies across many pending cases — that training on copyrighted material is automatically fair use.

This story ran in the Coral Wire edition of Thu, Oct 1 2026, when the Coral Index stood at 62/100 (Tailwinds building).

Coverage — 2 sources

Every outlet found reporting this story.

Also in this edition

Spot something wrong?

Every story here links its sources so you can check them. If we have got a fact, a figure or a date wrong, tell us and we will correct the page and say so on it.

Prefer email? info@coralwire.ai

Coral Wire selects stories by a Claude-graded relevance score and writes each headline and one-line summary with an Anthropic Claude model from the linked source. Story gathering is not AI: it is a fixed set of feeds and dockets. The linked original is always the authority.