On 29 September 2026, the Third Circuit affirmed that ROSS Intelligence infringed Thomson Reuters' copyright when it used Westlaw headnotes to build a competing legal search tool, and that the use was not fair. Judge Tamika Montgomery-Reeves wrote for a panel that also included Judges Restrepo and Bove. Within hours the ruling was being described as a verdict on AI training in general. The opinion itself is considerably more limited, and that limit matters for innovation policy.
What the court actually held
The facts were unusually bad for ROSS. According to the opinion, ROSS hired a vendor, LegalEase, to write roughly 25,000 legal memos for training, and the memo-writers leaned on Westlaw headnotes because they were "an easy way [to] frame questions." ROSS ran ads comparing itself directly to Westlaw at prices "in line with" Westlaw's. The district court held that 2,243 headnotes were copied and granted partial summary judgment to Thomson Reuters. The Third Circuit took the case on interlocutory appeal.
On the four statutory factors, the panel found that factors one, three and four weighed against fair use and factor two weighed slightly in ROSS's favor. It called ROSS's purpose "minimally transformative, at best." The opinion says expressly that ROSS's AI "was not a generative AI, meaning it would not create any new expression; it would only return text passages from preexisting judicial opinions." The court framed the dispute as "no more than an ordinary copyright case."
The strongest case for the ruling
Thomson Reuters' argument deserves a fair statement. Headnotes are editorial work product, and the opinion describes the style rules editors must follow, including limits of about 800 characters. A newcomer that takes that labor wholesale to launch a substitute product, while knowing it was working around Westlaw's terms of service, is doing something copyright law has long treated as unfair. The court also noted ROSS had access to roughly ten million uncopyrighted judicial opinions and chose the shortcut anyway. On that record, even a pro-innovation reader should be uncomfortable defending ROSS.
Why the reasoning is narrower than it looks
Two features of the opinion keep it from being a general rule against training.
First, the transformativeness analysis rests on necessity. The court distinguished software-interface cases such as Google v. Oracle, where copying was needed to reach unprotected functionality. ROSS did not need the headnotes because "ease is not a justification for copying." A developer who trains on lawfully available, unprotected material or on licensed data is on very different ground.
Second, the market-harm analysis leans on direct substitution. The court found ROSS built a competing legal-research platform in Westlaw's own market. It also accepted that a derivative market for licensing headnotes as AI training data was "rapidly developing," and that ROSS "usurped" Thomson Reuters' chance to enter it. That second finding is the part generative AI developers should watch. Reasoning that a licensing market exists because rights-holders say it could exist can be applied to nearly any training dataset. The court did not need to go there to decide this case, since the direct-competition finding was sufficient, and ROSS offered no evidence to rebut either market.
The court also rejected ROSS's public-benefit arguments as unsupported. It found it "unclear how ROSS's copying would greatly increase public access to the law" given that the underlying opinions are free, and that ROSS "presents no evidence" that the decision would halt AI development. The lesson for defendants is evidentiary: policy claims about innovation need proof in the record.
What is not decided
The ruling does not say that training a generative model on copyrighted works is infringement. Plagiarism Today notes that ROSS was not a generative AI company and contrasts it with the Bartz litigation, which involved models that create original expression. The question of how courts treat outputs that do not substitute for the inputs remains open. So does the weight of the Copyright Office's analysis, which it has been publishing in parts; per its AI page, a pre-publication version of Part 3, on generative AI training, was released on 9 May 2025.
The proceedings also had a procedural wrinkle. IPWatchdog initially reported the opinion as sealed pending redactions, and Plagiarism Today reported it unsealed the next day. The opinion is now public.
What policymakers and developers should take from it
The sensible reading is a rule against free-riding on a direct competitor's editorial product, not against machine learning. That is a proportionate outcome, and it argues against the legislative responses that tend to follow headline rulings. A blanket training-licence mandate or a statutory carve-out would be premature when the only appellate decision involves a non-generative tool and a defendant that skipped freely available alternatives.
Three practical steps follow. Developers should document why each protected input was necessary and prefer unprotected or licensed sources where they exist. Rights-holders should recognise that a licensing market has to be real, not hypothetical, to carry factor-four weight in later cases. Courts and regulators should resist extending a competitor-substitution holding to technologies the panel took care not to address. The open internet depends on both a functioning copyright system and room for new entrants, and Ross shows the system can protect the first without foreclosing the second.