On September 29, 2026, the Third Circuit issued what appears to be the first federal appellate decision on whether training AI on copyrighted material is fair use. In Thomson Reuters v. ROSS Intelligence, the court affirmed that case-law headnotes are copyrightable and that using them to train a non-generative legal-research tool was not fair use. Early coverage has emphasized footnote 7, which distinguishes AI that "cannot generate original expression" from generative models, and the court's finding that "the evidence shows that the market for licensing headnotes as text to train AI is rapidly developing." Less attention has gone to how courts will test evidence of that market.

Under the fourth factor, the court found harm to the original market and to the value of the headnotes, and it also recognized a potential derivative market for licensing them as training data. It cited evidence that the copyright owner was using its headnotes "as training data for its own AI search products" and held that the absence of licenses to third parties "does not disprove that a market exists to do so." Consistent with Warhol, the burden remained on the party asserting fair use.

This reasoning is difficult to square with two June 2025 decisions from the Northern District of California. Judge Alsup assumed that a market for licensing books as training data could develop but held that it "is not one the Copyright Act entitles Authors to exploit." Judge Chhabria held that whether such a market "exists or is likely to develop is irrelevant," because counting lost licensing fees for a transformative use as market harm would make the fourth factor circular. The Third Circuit did not address circularity, although it had already concluded that the use was "minimally transformative, at best." Read together, the decisions may suggest that the first factor operates as a threshold. If a court finds the training use transformative, licensing evidence may be legally irrelevant under the California approach; if it does not, that evidence may decide the fourth factor.

Where courts do weigh licensing evidence, the analysis may resemble a reasonable royalty dispute in a patent case. Courts may have to decide whether a proffered license is comparable to the use at issue, whether a bundled deal reflects a price for training at all, and how much weight to give a license negotiated to settle litigation. The Federal Circuit has held that licenses used to set a royalty must be "commensurate with what the defendant has appropriated" and has cautioned that "litigation itself can skew the results of the hypothetical negotiation" (ResQNet). In EcoFactor, the court held under Rule 702 that a damages expert should not have been allowed to testify that prior licenses reflected a per-unit royalty the licenses themselves did not support. The Supreme Court used similar reasoning in Google v. Oracle, giving little weight under the fourth factor to licensing negotiations that "concerned much more than 37 packages of declaring code." AI fair use cases may see the same comparability disputes and Rule 702 challenges to licensing experts. The analogy has limits: the burden of proof runs the other way, and patent law has no equivalent to the circularity concern.

For companies on either side of a training-data deal, the practical takeaway is that license terms may become evidence in fair use litigation. Companies should watch whether further review takes up the circularity question, whether the next district court follows the Third Circuit, and whether courts begin testing these licenses the way patent courts test comparable licenses.