The copyright question AI companies have been avoiding just reached a federal appeals court.
On June 11, 2026, the Third Circuit heard oral arguments in Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., a case that’s been working through federal courts since 2020. Ross built a
legal research tool that competed with Westlaw by training on Thomson Reuters’ copyrighted headnotes. A Delaware federal district court found direct copyright infringement of more than 2,000 headnotes. Ross appealed, and the Third Circuit is now deciding whether that training constitutes fair use.
Ross’s argument is that the copying was transformative, it used legal content to train an AI that
delivered a different kind of legal service, without reproducing the headnotes in output. Thomson
Reuters argues the copying was direct, systematic, and caused measurable market harm: Ross built a
competing product using Thomson Reuters’ own intellectual property as its training substrate.
This is the more precise framing of the case’s significance: it’s the first federal appeals court
ruling on AI training copyright fair use in a legal tools context. Multiple AI copyright cases are
moving through US courts, the broader statement that this is the first AI copyright federal appeal
of any kind requires more care. What makes this case distinct is the combination of a non-generative
training use, a direct commercial competitor as plaintiff, and a factual record built from a district
court infringement finding. Those specific facts make any Third Circuit ruling highly instructive for
the rest of the field.
The fair use analysis turns on the four-factor test under 17 U.S.C. § 107. The district court found
the copying wasn’t transformative enough to qualify, Ross wasn’t creating commentary, criticism, or
a new form of expression. It was extracting structured legal reasoning to replicate a competing
commercial function. That finding put factors one and four (purpose and market effect) squarely
against Ross.
At oral argument, legal analysts have noted the possibility that the court could “cabin” its ruling –
limiting the holding to the specific facts of non-generative AI training on structured legal content –
rather than issuing broad guidance on AI training fair use generally. That framing comes from
commentary associated with IP scholars tracking the case; the FSU source providing that specific
attribution isn’t currently accessible. The analytical point itself is legally sound: appeals courts
routinely narrow holdings to the facts before them, and a cabined ruling would leave generative AI
training cases for other circuits or a future Supreme Court term.
What to Watch
The real question is whether the Third Circuit writes a ruling narrow enough to avoid settling
generative AI training disputes, or broad enough that it becomes the reference point every other
circuit uses. For compliance teams at AI companies that trained models on licensed or unlicensed
datasets, the answer determines whether their legal exposure is immediately clarified or whether
they’re waiting for another year of circuit-level litigation.
Don’t expect a ruling for three to six months after oral argument. When it comes, it’ll be the most
authoritative federal statement yet on AI training and copyright law. Every AI company with training
data provenance questions should have outside IP counsel read the opinion the day it publishes.