At 3rd Circuit, Judges Press ROSS and Thomson Reuters on Fair Use, AI Training and Market Harm | LawSites

The ROSS v. Thomson Reuters Appeal Is About to Set the Rules for Training AI on Someone Else's Documents

On June 11, 2026, a three-judge panel of the 3rd U.S. Circuit Court of Appeals — Judges L. Felipe Restrepo, Tamika R. Montgomery-Reeves and Emil J. Bove — heard oral argument in Philadelphia in Thomson Reuters and West Publishing's copyright fight with ROSS Intelligence, the legal research startup that used Westlaw headnotes to train its system back in 2015. The dispute centers on 2,834 headnotes ROSS drew on, of which the district court found 2,430 infringing and outside fair use. Westlaw's headnote library runs to roughly 28 million entries, so the disputed set amounts to about 0.08% of the total corpus — a number both sides will now argue means something very different. ROSS's counsel, Mark S. Davies of White & Case, pushed for reversal on fair-use grounds; Thomson Reuters' counsel, Dale M. Cendali of Kirkland & Ellis, defended the district court's finding that ROSS built a competing product from protected editorial content.

What exactly is the 3rd Circuit being asked to decide?

Whether training an AI system on a publisher's proprietary text is "transformative" under fair use, and whether that use damages a real or potential licensing market.

Those are the first and fourth fair-use factors, and the panel spent most of argument on them. Davies argued the copyrighted material was the headnote as written, not the underlying idea, and that no market exists for licensing a single headnote as AI training data — arguing the reverse would let any publisher claim a "lost market" for whatever use later gets litigated. Bove and Restrepo weren't satisfied, repeatedly pushing Davies to explain how ROSS's output differed, from a user's perspective, from Westlaw itself.

Why should e-discovery and legal AI vendors care about a 2015-era case?

Because the ruling will set the appellate standard for training AI on copyrighted or proprietary text, regardless of which generation of model did it.

ROSS's own lawyer leaned into this, framing 2015-era ROSS as an "early example" of the AI transformation now underway, and Davies noted the product returned quotations pulled directly from judicial opinions rather than generated prose — meaning, he said, "no risk of hallucinations." That distinction matters for any vendor whose pitch rests on grounding answers in source documents rather than model-generated text, since it's exactly the argument a court may soon bless or reject.

Is Thomson Reuters' market-harm argument as solid as it sounds?

Not obviously — the harm is a claim about a hypothetical licensing market that Thomson Reuters itself is now defining after the fact.

Judge Bove's pointed question about whether a market for AI training "could" exist, separate from whether one did exist in 2015, is the tell. Buyers should notice that "market harm" in this case is being argued forward from litigation, not backward from an actual lost deal — the same pattern likely to recur whenever a publisher discovers its content trained a competitor's system.

What should buyers ask their own AI vendors now?

Ask exactly what content trained the model, whether any of it was licensed, and whether outputs are grounded in your own documents or the vendor's training corpus.

A platform that answers questions by citing back to a client's own collection — rather than paraphrasing content absorbed during training — sits on much steadier ground than one relying on unlicensed third-party text. This case is the clearest signal yet that "how was this trained" is becoming a due-diligence question, not a technical footnote.

Frequently asked questions

When will the 3rd Circuit rule?

No ruling date was set at argument; the court had only just released the June 11, 2026 hearing transcript when this was reported.

Does this case involve generative AI?

No. ROSS's system, built in 2015, returned quotations from judicial opinions rather than generated text — but the fair-use reasoning applies to any AI training dispute.

What's actually at stake numerically?

The district court found 2,430 of the 2,834 headnotes ROSS used to be infringing, against a Westlaw headnote library of roughly 28 million — about 0.08% of the total.

Source: LawSites, reported by Bob Ambrogi.