The first U.S. appellate court to weigh in on whether AI training on copyrighted material qualifies as fair use has sided with the copyright holder. On September 29, 2026, the Third Circuit affirmed the district court’s ruling in Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., holding that a legal-technology startup infringed Thomson Reuters’ copyrighted Westlaw headnotes when it used material derived from those headnotes to build a competing AI-powered legal research tool. The decision marks a watershed moment for copyright owners, AI developers, and anyone tracking the rapidly evolving intersection of intellectual property law and artificial intelligence.
District Court Decision
Thomson Reuters and its subsidiary West Publishing Corp. filed suit in 2020 in the District of Delaware, alleging that Ross Intelligence—a startup that built an AI-driven legal research engine—infringed copyrights in Westlaw’s editorially authored headnotes. Those headnotes are summaries of key legal points drawn from judicial opinions, organized through Westlaw’s proprietary Key Number System.
Ross had first sought to license Westlaw’s content to train its AI, but Thomson Reuters refused because Ross was a direct competitor. Ross then turned to a third party, LegalEase, which compiled roughly 25,000 “Bulk Memos”—documents pairing legal questions with model answers—that LegalEase’s lawyers created using Westlaw headnotes as a reference. Ross used those Bulk Memos to train its AI search tool, and Thomson Reuters sued when it discovered the arrangement.
The case first went through summary judgment in 2023, when U.S. Circuit Judge Stephanos Bibas, sitting by designation in the District of Delaware, largely denied Thomson Reuters’ motions. But as a trial date approached in August 2024, Judge Bibas studied the case materials more closely, continued the trial, and invited the parties to renew their summary judgment briefing.
On February 11, 2025, Judge Bibas issued a revised opinion granting partial summary judgment to Thomson Reuters on direct copyright infringement and granting summary judgment to Thomson Reuters on the fair use defense. After conducting a headnote-by-headnote comparison of the Bulk Memo questions, the Westlaw headnotes, and the underlying judicial opinions, Judge Bibas found that Ross had copied 2,243 headnotes and that the Bulk Memo questions were substantially similar to those headnotes.
Judge Bibas rejected each of Ross’ defenses. On the critical fair use analysis, he weighed the four statutory factors and found that factors one and four—the purpose and character of the use, and market harm—favored Thomson Reuters, while factors two and three—the nature of the copyrighted work, and the amount used—favored Ross. Crucially, on factor one, Judge Bibas held that Ross’ use was commercial and not transformative because Ross had used Thomson Reuters’ headnotes to build a competing legal research tool that served a substantially similar purpose to Westlaw. He distinguished the intermediate-copying cases involving computer code—Google v. Oracle, Sony v. Connectix, and Sega v. Accolade—on the ground that those cases involved functional computer code where copying was necessary to access unprotected elements, a circumstance not present here. On factor four, Judge Bibas held that Ross competed in the legal research platform market and that its copying affected a potential market for AI training data, regardless of whether Thomson Reuters had yet entered that market.
Oral Argument at the Third Circuit
A three-judge panel, U.S. Circuit Judges L. Felipe Restrepo, Tamika R. Montgomery-Reeves, and Emil J. Bove III, heard oral argument on June 11, 2026. The argument produced several notable exchanges that foreshadowed the panel’s ultimate resolution.
The panel pressed Ross’ counsel early and often on how its AI-powered search tool differed in practice from Westlaw. Judge Bove expressed open skepticism about Ross’ framing, noting that Ross’ counsel emphasized the technology behind the product rather than describing how the user experience differed from Westlaw’s.
The judges also probed where to aim the market-harm analysis. Ross urged the court to focus narrowly on the market for individual headnotes, while Thomson Reuters pointed to the broader legal research market and the emerging market for AI training data. Judge Bove observed that the parties stood “worlds apart” on which market the court should examine.
The panel showed significant skepticism toward Ross’ reliance on the intermediate-copying line of cases. Judge Bove questioned whether those precedents, which involved unreadable object code, applied at all, pointing out that Ross’ employees could simply read the judicial opinions without copying Westlaw’s headnotes. Finally, Judge Montgomery-Reeves challenged Ross’ contention that the headnotes lacked sufficient originality, asking why attorney-edited summaries that condensed judicial language would not satisfy copyright’s low creativity threshold.
Significance of the Decision
The Third Circuit affirmed Judge Bibas’ order in a one-page judgment issued on September 29, 2026. Judge Montgomery-Reeves authored the panel’s opinion, which remains under seal as the court gives the parties time to propose redactions within 10 days. The full reasoning and analysis will become public once the court determines what, if anything, should remain redacted.
Even without the written opinion in hand, the affirmance itself carries immediate significance for several reasons.
- First appellate ruling on AI training and fair use. This decision represents the first time a U.S. Court of Appeals has addressed whether using copyrighted material to train an AI system can qualify as fair use, providing an appellate data point where district courts have reached conflicting conclusions. Although the ruling binds only courts within the Third Circuit, commentators have noted that it will likely be cited in pending AI copyright cases—though its broader influence will depend on whether the panel resolved fair use as a matter of law or on narrower, fact-specific grounds.
- The non-generative AI distinction. Judge Bibas emphasized at the district court level that this case involved non-generative AI, a tool that returned existing judicial opinions rather than generating new content. That factual distinction may limit the ruling’s direct applicability to cases involving generative AI models, where developers argue that training produces fundamentally different outputs. The unsealed opinion will reveal whether the Third Circuit treated this distinction as material to its analysis.
- Reinforcement of Warhol and limits on intermediate copying. The district court grounded its analysis in the Supreme Court’s 2023 Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith decision, which held that when an original work and a secondary use share the same or a highly similar purpose and the secondary use is commercial, the first fair use factor likely weighs against fair use. That framework complicates the path for AI developers invoking fair use when their products compete with the copyrighted works they trained on. The affirmance also signals that courts may cabin the intermediate-copying doctrine (developed in cases involving functional computer code) to its software-specific origins rather than extending it to human-authored editorial content used as AI training data, a position both the district court and the Third Circuit panel at oral argument endorsed.
The Third Circuit’s affirmance delivers concrete guidance for both sides of the AI-copyright divide.
- For AI developers, the decision reinforces that training on copyrighted material to build a product that competes with the source of that material carries serious infringement risk, particularly after Warhol narrowed the scope of transformative use. Companies developing AI tools should audit their training pipelines, assess whether their products serve a substantially similar purpose to the works they train on, and evaluate whether licensing, rather than relying on a fair use defense, offers a more defensible path. The intermediate-copying defense, long a refuge in software disputes, now appears unlikely to extend to AI systems trained on human-authored content.
- For copyright owners, the ruling validates enforcement efforts and strengthens the position that unauthorized use of protected content as AI training data—even when the content does not appear in the AI’s output—can constitute infringement. Rights holders should take note that the court recognized harm not only to the existing market for Westlaw but also to a potential licensing market for AI training data, a theory that applies well beyond the legal-research context.
All eyes now turn to the unsealed opinion, which will reveal the Third Circuit’s full reasoning and signal how narrowly or broadly this landmark decision applies across the AI training ecosystem. More updates to come.
Recent Comments