The Third Circuit’s Turn on AI Training and Fair Use

The most consequential intellectual property question of the decade, whether training artificial intelligence on copyrighted works is fair use, is heading toward the federal appellate courts. The Third Circuit, which covers New Jersey and Pennsylvania, is near the front of the line.

The trial-level decisions have split. In Thomson Reuters v. Ross Intelligence, a Delaware federal court granted partial summary judgment to Thomson Reuters in early 2025, holding that copying Westlaw headnotes to train a competing legal-research tool was not fair use, largely because the use was commercially competitive and harmed the market for the original. That ruling is now on interlocutory appeal to the Third Circuit, positioning our circuit to deliver one of the first appellate words on AI training and fair use.

Two California decisions in mid-2025 cut differently. In Bartz v. Anthropic, the court found that training on books could be transformative fair use, while separately holding that building a library from pirated copies was not protected; the case later settled for a reported figure of roughly $1.5 billion. Days later, in Kadrey v. Meta, another judge found training “highly transformative” but signaled that a plaintiff who could prove market dilution might win on different facts. Same doctrine, different emphases, and no consensus, especially on whether AI-driven market harm defeats a fair use defense.

Why this matters if you are not an AI company. First, appellate law binds. A Third Circuit ruling would govern every federal court in New Jersey and Pennsylvania and would shape licensing leverage for anyone whose work is ingested by a model, or who builds products on model outputs. Second, the split means the answer turns heavily on facts: transformation, the use of pirated inputs, and provable market harm. That is exactly the evidence litigants should be preserving now.

For creators and rights holders, the takeaways are concrete. Register copyrights, because registration is a predicate to statutory damages and a stronger negotiating position. Add “no training” or restricted-use terms to licenses and terms of service; courts have treated the presence or absence of such terms as meaningful. And document any concrete market harm from unlicensed use, because that is the factor on which these cases increasingly turn.

For companies building with AI, the mirror image applies. Understand what data trained the tools you use, demand contractual assurances and indemnities, and avoid pipelines that depend on pirated or unlicensed inputs, which is where even AI-friendly courts have drawn the line.

The doctrine is unsettled and will stay that way through at least the next round of appeals. But unsettled is not unmanageable. The parties who fare best will be the ones who built their records and their contracts before the appellate courts drew the lines.

Legal note: This article provides general information and is not legal advice.