IP Hot Topic: Third Circuit Probes Fair Use in Thomson Reuters v. Ross Intelligence

The Third Circuit heard oral arguments in one of the first AI-related fair-use cases to reach the courts and among the first to be reviewed on appeal: Thomson Reuters v. Ross Intelligence. The case was among the earliest to grapple with the novel AI issues facing district courts across the country last year: whether AI training using copyrighted materials creates something meaningfully new, how to define the relevant market, and how to assess market harm. The Third Circuit’s questioning on these issues suggests meaningful skepticism that a system serving the same end function as the original, such as Ross’s, can satisfy fair use.

Background

This appeal was from Judge Bibas’s revised summary-judgment ruling that Ross’s use of Westlaw headnotes to train a non-generative legal-research tool was not fair use. Below, Judge Bibas awarded factors one (purpose) and four (market) to Thomson Reuters and factors two (nature of the work) and three (amount used) to Ross, then held the balance defeated fair use. More details on the lower court’s decision can be found in our prior alert here.

The Court Focused on Purpose and Market Harm

The panel pressed Ross to articulate, concretely, what made its product transformative, with one judge frustrated that counsel was “dancing” around the question instead of answering it. Ross emphasized that it wrote the Bulk Memos, used only 0.08% of 28 million Westlaw’s headnotes, returns only quotations from public opinions, and trained its system to “think like a lawyer.” But the judges seemed unpersuaded by that this adequately described a transformation. The recurring theme from the bench was whether Ross’s use can be transformative when it serves the same end purpose as Westlaw. Ross conceded that, “at some level,” its product and Westlaw’s are the same to a user but struggled to explain why that similarity does not defeat fair use.

Ross leaned heavily on the Third Circuit’s ASTM v. UpCodes (Apr. 7, 2026) decision that held that UpCodes (an online research platform that offers a searchable database of building codes) likely showed fair use in republishing copyrighted ASTM standards incorporated by reference into building codes. There, the court found transformative use despite verbatim copying because it served the distinct purpose of disseminating “only what the law is.” But UpCodes turned on ASTM’s failure to prove usurpation; and Thomson Reuters appeared better positioned on concrete market harm during arguments. Thus, this decision may be a double-edged sword for Ross.

Market harm and the definition of the relevant market also drew the attention of the court. Thomson Reuters identified three market harms: direct substitution; harm to its exclusivity in training its own AI tools (noting it trained on headnotes years before Ross existed); and a potential licensing market for AI training data, evidenced by Ross paying a third party after being refused a license. Thomson Reuters emphasized that Factor 4 was the most important. On the other hand, Ross argued that the Factor 4 analysis is narrowed because its AI was “so transformational.” Ross also maintained that there is no market for individual headnotes and no market for AI training as Ross was allegedly “the only one[] who used headnotes for AI.” While Thomson Reuters leaned heavily into who has the burden to show market harm (or a lack thereof), the court notably questioned whether the burden of proving (or disproving) market harm should drive the analysis at all. The court also drew an analogy to the Third Circuit’s Video Pipeline v. Buena Vista Home Entertainment case, where “there was no market for Disney trailers standing alone either, but the website that stole the trailers was found to have caused diminished value to Disney’s website, because Disney used that to attract people to their website to do other things on their website.”

Was the Intermediate Copying “Necessary”?

A substantial colloquy addressed the software intermediate-copying cases—Sony, Sega, and Google v. Oracle. Ross argued its copying occurred at an intermediate step (turning headnotes into numerical training data) and should be excused like the reverse-engineering in those cases. The panel was skeptical, stressing that those cases involved object code humans cannot read, which had to be copied to reach unprotected functional elements—whereas Ross’s employees could simply read the public opinions directly, so the copying was not necessary. Ross urged there is no strict-necessity requirement, only a legitimate purpose, and pointed to the time pressure on a startup; a judge noted that rationale “sounds very commercial.” This exchange is nominally about purpose, but its real thrust is the same substitution-and-necessity concern that drives factors one and four.

The Open Questions: Factor-Weighing and Bad Faith

The court flagged two questions of interest. First, the arithmetic of the factors: if Thomson Reuters takes one and four and Ross takes two and three, who wins? Ross maintained it would still prevail, calling fair use a holistic test and noting that Google began its analysis with factor two—but the panel’s repeated “who wins?” signaled doubt that a win on the two lesser factors can carry the defense once the two most important factors break the other way. Second, whether bad faith belongs in the fair-use inquiry at all. The panel raised Ross’s communications with Dentons about obtaining the materials; Ross argued bad faith is irrelevant, citing Campbell (denial of permission does not weigh against fair use) and Sega (copyright is not a privilege reserved for the well-behaved), and read Google as strongly suggesting—without holding—that it does not matter. Thomson Reuters countered that if bad faith ever mattered, this is the case, and it favors Thomson Reuters.

Takeaways

We will continue to monitor for the court’s decision. But the court’s questions during oral arguments highlight the following takeaways:

• Uses that deliver the same end function as the source product risk being treated as non-transformative.
• Some courts, like the Third Circuit, may recognize emerging licensing markets for training data even if no established market exists today. But as last year’s Northern District of California decision in Bartz et al. v. Anthropic PBC suggests, uneven treatment of this issue may still persist nationally.