The Third Circuit has now shown its work. On October 1, 2026, Above the Fold reported that the court affirmed the judgment for Thomson Reuters in its suit against Ross Intelligence and that it filed its opinion under seal. With the full opinion now public, practitioners can see how the first federal appellate court to address AI training and fair use got to its result. The court calls the dispute “no more than an ordinary copyright case.” Even so, its reasoning gives copyright owners new arguments to use. It also gives companies building or deploying AI systems a list of facts that may help them distinguish this ruling.
Link to The Court’s Reasoning The Court’s Reasoning
Writing for a panel that also included Judges Restrepo and Bove, Judge Montgomery-Reeves resolved both certified questions in Thomson Reuters’ favor. The court held that Thomson Reuters’ headnotes qualify for copyright protection and that Ross’s use of them was not fair use. On that basis, it affirmed partial summary judgment on 2,243 headnotes.
The court began with originality, applying Feist’s “extremely low” creativity threshold. Measured against that modest standard, the headnotes readily qualified: the court found a “creative spark” in the decisions Westlaw editors make about which points of law to highlight and how to phrase them so that each headnote stands on its own. Supreme Court precedent reinforced that conclusion, as the panel pointed to the Court’s recognition, in Callaghan v. Myers and Georgia v. Public.Resource.Org, that private reporters may own copyrights in headnotes. The court then turned to each of Ross’s three counterarguments. It rejected the claim that protecting headnotes would hand Thomson Reuters a monopoly over the law, reasoning that headnotes are not law and that judicial opinions remain free for anyone to use. It likewise found the merger doctrine inapplicable because “there are many ways to express points of law from judicial opinions.” Finally, the court distinguished the Second Circuit’s decision in Matthew Bender v. West, which addressed convention-driven elements such as parallel citations rather than independently composed headnotes.
Turning to fair use, the court applied Andy Warhol Foundation v. Goldsmith to the first factor, weighing how different Ross’s purpose was against how commercial its use was. The commercial side of that ledger was straightforward: Ross priced its platform to compete with Westlaw, so its use was commercial. The court acknowledged that training an AI program “arguably presents a slight degree of difference in use.” It concluded, however, that both companies ultimately used the headnotes to build and improve a legal research platform, which left Ross’s use “minimally transformative, at best.” For the same reason, the court found Authors Guild v. Google inapplicable because Google’s search tool could send readers to the original books, while Ross admittedly set out to replace Westlaw.
Ross fared no better with the intermediate-copying cases (Google v. Oracle, Sega, and Connectix), which the court confined to situations where copying is necessary to reach unprotected functional elements of computer code. Here, by contrast, Ross could have worked from the freely available opinions themselves but chose the headnotes because they were “easy” to use. The court’s answer may become the opinion’s best-known line: “Unlike necessity, ease is not a justification for copying.” In a footnote, the court added that, to the extent good faith still matters, the record showed Ross at times acted in bad faith, including by accessing Westlaw with credentials whose use Westlaw’s terms prohibited.
The second factor gave Ross its only win, and a modest one at that. Thomson Reuters had published the headnotes, and they are more factual than fictional. The court noted, however, that this factor rarely carries much weight.
The third factor swung back to favoring Thomson Reuters. The court treated each headnote as a separate copyrightable work, so every headnote Ross copied counted as a complete work. This framing defeated Ross’s argument that it took only 0.08% of Westlaw’s 28 million headnotes. And because Ross lacked a transformative purpose and had no need to copy, it “took more than necessary.”
The fourth factor, market effect, cut against Ross as well. Drawing on its own decision in Video Pipeline, the court held that copying can reduce a work’s value even when no standalone market exists. Here, the headnotes’ value lies in attracting Westlaw subscribers. Beyond that loss in value, the court found harm to the market for legal research platforms. More significantly for future cases, it also found harm to a potential derivative market for licensing headnotes as AI training data. The court described that market as “rapidly developing” and noted that Thomson Reuters already uses its headnotes to train its own AI tools. Thomson Reuters had never licensed headnotes to others, but the court held that this “does not disprove that a market exists.” Finally, the court gave little weight to Ross’s claimed public benefits, including access to the law, continued AI development, and national security, because Ross offered no supporting evidence.
Link to Significance of the Decision Significance of the Decision
Further review looks possible but uphill. The Third Circuit’s ruling came on interlocutory appeal, so the case will return to the District of Delaware after appellate review ends. Ross’s counsel has said that Ross plans to seek Supreme Court review. Ross could also ask the full Third Circuit for rehearing. Neither path looks easy. The panel ruled unanimously and with no separate opinion. It also built its analysis on settled Supreme Court authority, including Feist, Warhol, Campbell, and Oracle. In addition, the court expressly declined to treat the case as a referendum on AI, which may lead the Supreme Court to see it as a poor vehicle for deciding broader AI questions. Certiorari petitions often point to a circuit split, and the panel aligned its reasoning with the Second Circuit rather than departing from it.
Other circuits may borrow selectively. The Third Circuit’s opinion binds only courts within the Third Circuit, but other courts will likely cite it in pending AI cases. Plaintiffs suing Anthropic have already filed it as supplemental authority.
Two parts of the reasoning travel most easily. First, the derivative-market holding allows a rights holder to rely on an emerging AI-licensing market even when it has not yet licensed to others. That theory applies to any content owner whose works have value as training data. Second, the necessity-versus-convenience framework will likely lead courts to ask whether a developer could have used public domain, licensed, or self-created material instead. Developers of large general-purpose models will likely respond that massive and diverse datasets are technically necessary, not merely convenient. The Second Circuit, in particular, now has a sister circuit’s reading of Authors Guild and TVEyes to consider.
Defendants have clear ways to limit the decision. The opinion itself supplies several features that AI developers can use to confine the decision to its facts:
- Non-generative AI. The court stressed that Ross’s system “would not create any new expression.” In footnote 7, the court said the concerns raised in the DOJ’s statement of interest in In re: OpenAI “do not apply here.” It contrasted Ross with the large language models at issue in Bartz v. Anthropic and Kadrey v. Meta, which can generate original content. Generative AI defendants will rely heavily on that footnote.
- A direct competitor with the same purpose. Factors one and four both depended on Ross’s admitted goal of replacing Westlaw in the same market. General-purpose models with many possible uses offer a weaker match between the purpose of the training and the purpose of the original works.
- Training, not output. Ross returned only passages from public-domain opinions. The court therefore had no occasion to decide whether AI-generated output infringes. It also did not resolve the separate, contested question of whether non-infringing AI output can harm the market for training works.
- A narrow record. The court rested its decision on undisputed facts, including the free availability of the underlying opinions and Ross’s bad-faith conduct. It expressly left open whether headnotes that quote opinions word for word are copyrightable. It also declined to decide the originality of the Key Number System because Ross forfeited that issue. Finally, the court noted that the DOJ chose not to participate in this case.
The takeaway. Copyright owners now have appellate support for two arguments: that unauthorized use of their works as training data can infringe even when the works never appear in the AI’s output, and that an emerging licensing market counts under factor four. Companies developing or deploying AI should consider whether their products compete with the sources of their training data. They should also document why they need particular training content, because the Third Circuit made clear that convenience will not justify copying. Licensing remains the most defensible path where the purposes overlap. Above the Fold will continue to follow any further review in the case and the opinion’s influence on the generative AI cases now pending.