Is it legal to train AI on your work? An appeals court answered for the first time — with an asterisk

What happened

On Tuesday, Sept. 29, a three-judge panel of the Third Circuit affirmed that ROSS Intelligence — a now-defunct legal research startup — infringed Thomson Reuters' Westlaw headnotes and could not claim fair use for using material derived from them to train a competing AI search tool.

The opinion was written by Judge Tamika Montgomery-Reeves. It runs 32 pages and is marked precedential, meaning it binds district courts across the circuit. It affirms Judge Stephanos Bibas's district court order of Feb. 11, 2025, and taxes costs on appeal to Thomson Reuters.

One catch: nobody has read it yet. The opinion is sealed while the parties propose redactions, which are due within 10 days of Sept. 29. No publication date has been set.

This is the first time a US appeals court has ruled on whether training an AI system on someone else's copyrighted work is fair use. The answer was no — in this case.

The part that should change how you negotiate

The district court analysis that was affirmed did not stop at "ROSS competed with Westlaw." It accepted a second kind of harm: a potential market for licensing headnotes as AI training data. Thomson Reuters put forward evidence that such a market was developing and that it trained its own AI products on its own headnotes.

So the unlicensed taking wasn't just competition. It was the appropriation of a licensing opportunity the owner could plausibly have sold.

Read that again if you have an archive. Every licensing deal your peers sign makes the next unlicensed scrape harder to defend. The Google publisher pilot, the $5,000-per-book HarperCollins rate, the Taylor & Francis deal — those aren't only revenue line items. They are now evidence that a market exists.

The music industry saw this coming. The RIAA and the NMPA filed a joint brief telling the court that training a model on copyrighted works to build a competing service "cannot qualify as fair use." NetChoice and other tech trade groups, plus digital rights organizations, backed ROSS.

What this is not

It is not a ruling about generative AI. ROSS's system used machine learning to surface relevant judicial opinions. It did not generate new text. The decision does not hold that every use of copyrighted material to train a generative model is unlawful, and it does not remove fair use as a defense in cases with different technology, different outputs, or different markets.

It is also not the only direction the law is moving. In Bartz v. Anthropic, the court found that training on books was fair use, while storing pirated copies was not. That case settled for $1.5 billion — roughly $3,000 per work.

And we don't have the reasoning. One secondary summary reports the panel found two factors favored ROSS — that headnotes contain less creativity than conventional literary works, and that ROSS never displayed them to users — but that commercial purpose and market effect weighed heavier. We could not verify that against the opinion, because the opinion is sealed. Treat it as a report, not a holding.

Whether ROSS seeks rehearing or Supreme Court review is unknown. The company is defunct, which complicates any appeal.

What to do about it

1. Stop treating "fair use" as settled in either direction. One appellate ruling exists, it's narrow, and the text isn't public yet.
2. If you're negotiating an AI license, cite the market-harm theory. The affirmed analysis counted a developing licensing market as real. That is leverage.
3. Document that you license. A public rate card, even an unused one, is evidence the market exists.
4. Keep training, retrieval and transformation priced separately. This ruling involved a non-generative search product. The categories are legally distinct and should not share one price.
5. Train AI on your own archive and say so publicly. Thomson Reuters' use of its own headnotes to build AI products was part of what made the licensing market credible.
6. Watch for the unsealed opinion. The reasoning decides how much this travels — and whether it reaches generative cases at all.
7. Don't tell your audience the courts ruled against AI training. They ruled against one company, on one dataset, with one business model.