Skip to content
Legal Kitty. Stay rich. Stay famous. Stay free.Legal Kitty
Menu
Retro illustration of a turntable, a stack of leather-bound books, a brass desk lamp and a reel-to-reel tape machine on a wooden desk at night, for a post on AI training and fair use

IP & Trademarks4 min read

The AI Training Fair Use Loss Music Cheered

An appeals court said copying for AI training wasn't fair use here.

On September 29, 2026, the Third Circuit entered judgment in Thomson Reuters v. ROSS Intelligence, and the opinion hit the docket the next day. In a precedential opinion, the court held that 2,243 Westlaw headnotes are original enough for copyright and that ROSS's use of them to train its AI was not fair use. Its last line: "we will AFFIRM."

Why would a fight over legal research summaries make the music business happy? Because the Recording Industry Association of America and the National Music Publishers' Association filed a joint brief backing Thomson Reuters, according to Music Business Worldwide. Universal Music Group and Sony Music are suing the AI music platform Suno in Massachusetts federal court. ROSS itself is now a defunct startup, per MBW, so the reasoning is what travels. Everyone with a catalog is reading it for that. Here is what it actually says, and where it stops.

Takeaway #1: "Nobody licenses this yet" didn't save the copier

FROM THE RULE

"(4) the effect of the use upon the potential market for or value of the copyrighted work."

17 U.S.C. §107 (fair use, factor four)

FROM THE OPINION

"Here, the evidence shows that the market for licensing headnotes as text to train AI is rapidly developing. ... That Thomson Reuters did not license its headnotes to others does not disprove that a market exists to do so."

Thomson Reuters v. ROSS, 3d Cir. No. 25-2153, slip op. at 26

What it means for you. Factor four asks what the copying does to the market for your work, including markets you haven't entered yet. ROSS argued there was no market for headnotes at all. The court disagreed: Thomson Reuters presented evidence that it was using its own headnotes to train its own AI search products, and the AI-training license market was "rapidly developing." By copying without permission, the court said, ROSS "usurped Thomson Reuters's opportunity to enter that derivative market." The court's own scorecard: "the second factor weighs slightly in favor of fair use, but the first, third, and fourth factors weigh against it." Result: "we hold that ROSS's use was not fair."

For a songwriter, an illustrator or a podcaster, that is the useful part. Not having signed an AI-training deal yet doesn't mean your work has no training market.

Your move. Creators: keep a paper trail. Licensing inquiries, rate cards, deals you turned down and any AI-training terms you have signed all show a market exists. Rights holders: say in writing what you do and don't license for training.

Takeaway #2: The limit sits right in the opinion

FROM THE OPINION

"ROSS's AI was not a generative AI, meaning it would not create any new expression; it would only return text passages from preexisting judicial opinions."

Thomson Reuters v. ROSS, 3d Cir. No. 25-2153, slip op. at 5

What it means for you. ROSS built a search tool. You typed a legal question, and it handed back passages from existing court opinions. It didn't make music or images, and it didn't clone anyone's voice. The holding is about that kind of system. Music Business Worldwide put it plainly: the judgment "does not settle the fair use question in the Suno and Udio cases." Generative AI defendants will argue that making new output is a different purpose. Labels will point to this market-harm reasoning. No appeals court has decided the generative question in those cases.

So the headline "AI training isn't fair use" is too big. The devil is in the details, and the detail here is a non-generative search engine that competed head-to-head with the product it copied.

Your move. Founders and producers: if you are building on licensed data, keep the licenses tidy. If you are licensing your catalog out, spell out training, fine-tuning and output rights separately. Then ask your label, distributor or publisher what AI-training language already sits in your deal, and get the answer in writing.

Sample wording (our illustration): "No right is granted to use the Work, in whole or in part, to train, fine-tune or evaluate any machine-learning model without a separate written license."

And always remember, don't sign it till you read it, my friends.

General information, not legal advice. Reading this doesn't create an attorney-client relationship. Talk to a lawyer about your own catalog and contracts.

Sources

  1. Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153 (3d Cir.), Opinion filed Sept 29, 2026 (docketed Sept 30, 2026)
  2. 17 U.S.C. §107 (Office of the Law Revision Counsel)
  3. Music Business Worldwide, 'US appeals court rejects fair use defense over AI training', Sept 30, 2026

← Blog