What The Third Circuit Decided
On Tuesday 29 September 2026 the US Court of Appeals for the Third Circuit affirmed a ruling for Thomson Reuters against ROSS Intelligence. The opinion, written by Judge Tamika Montgomery-Reeves for a panel with Judges Luis Restrepo and Emil Bove, holds that Westlaw headnotes are original enough for copyright protection and that ROSS's use of them was not fair. It upholds Judge Stephanos Bibas's partial summary judgment of 11 February 2025 and orders ROSS to pay Thomson Reuters' costs on appeal.
ROSS built a legal search engine that answers plain-language questions with passages from judicial opinions. To train it, a contractor wrote about 25,000 memos, using Westlaw headnotes to frame the questions, and ROSS ran ads comparing its product with Westlaw at similar prices. Some law firms switched. The summary judgment covers 2,243 headnotes.
This was an interlocutory appeal, argued on 11 June 2026. Music Business Worldwide reports the opinion runs to 32 pages and is marked precedential, so it will bind district courts in the circuit, and describes ROSS as a now-defunct startup.
How The Four Factors Came Out
The court weighed the four statutory fair use factors, and three went against ROSS.
| Factor | Finding | Main reason |
|---|---|---|
| 1. Purpose and character | Against fair use | Commercial, and only minimally transformative at best |
| 2. Nature of the work | Slightly for fair use | Headnotes are published and more factual than creative |
| 3. Amount used | Against fair use | Entire headnotes copied, and more than necessary |
| 4. Market effect | Against fair use | Substitute product, and a developing market for AI training licences |
On the first factor the court found that ROSS and Westlaw use headnotes for the same ultimate purpose, helping users find responsive legal material, and that training an AI was an intermediate step. It distinguished the Google Books case and the software-interface cases because ROSS could have copied the freely available judicial opinions instead. It chose the headnotes because they were an easy way to build its memos, and the court said ease is not a justification.
On the third factor the court rejected ROSS's point that it used only 0.08 percent of 28 million headnotes. On the fourth it found that the market for licensing headnotes as AI training text is rapidly developing and that ROSS usurped Thomson Reuters' chance to enter it. In a footnote it also noted evidence of bad faith, including attempts to reach Westlaw through borrowed or disguised accounts.
Why It Is Narrower Than The Headlines
Several outlets summarised the ruling as a rejection of fair use for AI training. The opinion is more specific. It says that, despite ROSS's framing, this is an ordinary copyright case about whether the materials are original and whether this use was fair.
A footnote does the limiting work. ROSS's AI was not generative and returned only existing passages, so it created no new expression. The court said the concerns raised in the pending OpenAI litigation do not apply, where the US Justice Department's 1 September statement of interest relied on the Bartz v. Anthropic ruling that training a language model can be transformative. Here, it wrote, ROSS trained its AI to build a commercial substitute for Westlaw.
The music industry bodies RIAA and NMPA filed a brief arguing that training a model to compete with the works it was trained on can never be fair use. That is their reading, and the opinion does not adopt it in those words. What it does hold is that copying protected text to build a competing product with the same purpose, when free alternatives exist, was not fair.
What Companies Training Or Buying Models Should Do
If you train models, audit the data for third-party editorial layers such as summaries, labels, annotations and curated categories, especially where the underlying source is free. The court treated the headnotes as a shortcut ROSS chose for convenience, and that choice counted against it on two factors.
If you buy models or model-built products, ask vendors for training-data warranties and indemnities, and ask whether the product competes with any content owner whose material was used. The court's market reasoning means a visible licensing market for a type of content weighs against copying it, so budget for licences rather than assuming a defence.
This is a US appeals ruling under US fair use law, so it does not bind courts in the EU or the UK, where the exceptions differ. What travels is the shape of the evidence: the purpose of the use, whether the output substitutes for the original and whether a licensing market exists. The reports we reviewed do not say whether ROSS will seek further review.
Servola Journal
We do this for everyone trying to keep up with what technology is doing to our lives. The people who build it, and the people it happens to. The Servola Journal exists so that what we learn belongs to all of them.
Nobody pays us for this. No ads, no paywall, free to everyone. We just believe that understanding what's happening to all of us shouldn't depend on who can afford to pay for it.
If it gave you something today, tell us to keep going. Follow us, leave a like, or write a positive comment. We read every one, and they are what keeps us going.
Read next: Newspapers Sued Over AI Training. The Number Was Filed by Microsoft. | AI Act Compliance Did Not Save Suno From GEMA
