AI Training Fair Use Rejected in Third Circuit Westlaw Ruling

The first appellate answer on AI training

For three years, AI copyright litigation has produced trial-court rulings that point in different directions. On September 29, 2026, a federal appeals court took up the question for the first time. The U.S. Court of Appeals for the Third Circuit affirmed partial summary judgment for Thomson Reuters against ROSS Intelligence. The panel held that Westlaw's editorial headnotes are protected by copyright and that ROSS's use of them to train a competing legal-research AI was not fair use.15 Judge Tamika Montgomery-Reeves wrote the opinion for a unanimous panel that also included Judges Restrepo and Bove.1416

The sequence of events was unusual. On September 29 the court issued a one-page order that simply said "affirmed," and its full opinion stayed under seal until the next day.13 That explains why some coverage dates the ruling to September 29 and some to September 30.4 The decision was effectively certain to end in a rights-holder win once the opinion became public. The open question was how far the reasoning would reach.

The short answer: it reaches less far than the headline suggests on generative AI, and further than many AI developers would like on licensing markets.

What ROSS did, and why it lost

The facts are now familiar. ROSS hired LegalEase Solutions to produce training data, and LegalEase and a subcontractor prepared about 25,000 "Bulk Memos" that drew on Westlaw headnotes. ROSS then turned those memos into machine-readable data to train its system.11 Users of the finished tool never saw a headnote. It answered plain-English legal questions by returning passages from judicial opinions.8 The district court found that 2,243 of the headnotes had clearly been copied, and the appeal concerned only those.512

On the four statutory fair-use factors, the panel found three against ROSS. On the first factor, the use was highly commercial and "minimally transformative, at best." Westlaw and ROSS both used the headnotes for the same end: building a platform that helps lawyers find relevant law.1316 The second factor tipped slightly toward ROSS because headnotes are largely factual.14 The third factor went the other way from the district court. Judge Stephanos Bibas, sitting by designation, had given it to ROSS, but the appellate panel treated each headnote as a complete copyrighted work and found that ROSS took more than it needed.1516 ROSS argued that it had used only about 0.08% of Westlaw's 28 million headnotes. The court rejected that point.914

The line most likely to be quoted in future briefs came in the court's response to ROSS's "intermediate copying" argument, which relied on Google v. Oracle, Sega, and Connectix. Those cases, the panel said, protected copying that was necessary to reach unprotected functional code. ROSS already had the free judicial opinions and chose the headnotes because they were convenient. In the court's words, "Unlike necessity, ease is not a justification for copying."815

ROSS's conduct also hurt it. A footnote cites evidence that ROSS tried to get into Westlaw using law-firm investor credentials, and that an employee asked about an account while posing as a solo practitioner. To the extent good faith still matters, the court said it weighed against ROSS.1416 The record also showed ROSS pricing its service in line with Westlaw, running ads that compared the two products, and executives hoping ROSS would replace Westlaw. Some law firms did switch.12

The narrow reading: generative AI left open

Nearly every practitioner analysis reaches the same conclusion: this is not the generative AI ruling the industry is waiting for. The panel described the dispute as "no more than an ordinary copyright case."16 In footnote 7, it separated ROSS's tool from the models at issue in Bartz v. Anthropic, Kadrey v. Meta, and In re: OpenAI. Unlike those models, the court said, ROSS's platform "cannot generate original expression."113 Patently-O reported that the footnote answered the Trump administration's recent statement of interest in the OpenAI litigation.8 Baker Donelson noted that the Justice Department chose not to weigh in on ROSS itself.14

Ballard Spahr argues that the opinion's reach is limited because the technology in the case was limited: ROSS retrieved existing text and created nothing new.1 Inside Tech Law reads it the same way. On its account, the case shows only that machine learning does not displace ordinary copyright analysis when protected expression is copied for a closely similar competitive purpose.12

That is a fair reading of the holding, but footnote 7 works in both directions. Several commentators note that the distinction suggests generative AI developers may do better on the first factor. Ballard Spahr reads the footnote as a signal that the generative/non-generative line may carry real weight in transformativeness analysis.1 On this reading, AI companies have reason to welcome footnote 7 even as they dislike the result.

The broad reading: a licensing market for training data

The more consequential part of the opinion is the fourth-factor analysis. The panel credited evidence that the market for licensing headnotes as AI training data "is rapidly developing." It noted that Thomson Reuters already trains its own AI search products on headnotes. It also held that the company's failure to license them to anyone else "does not disprove that a market exists."1016 Haynes Boone describes this as circuit-level recognition that licensing text for AI training is a cognizable derivative market, even when the owner has licensed to no one.5

This is where the reporting shows the sharpest doctrinal split. In the Northern District of California, Judge Alsup accepted that a market for licensing books as training data could develop but said it was not one the Copyright Act entitles authors to exploit. Judge Chhabria went further, calling the existence of such a market irrelevant. He reasoned that counting lost license fees for a transformative use would make the fourth factor circular.10 The Third Circuit did not address the circularity argument.1016

Baker Botts offers a way to reconcile the decisions: the first factor works as a gate. If training is transformative, licensing evidence may drop out under the California approach. If it is not, that evidence may decide the fourth factor.10 That framework fits the opinions as written. It also means the generative AI fight will turn on transformativeness, and the market-harm argument will matter only for whoever loses that question.

Rights holders move quickly

Publishers are not treating the ruling as narrow. The Association of American Publishers called it a milestone and "powerful precedent" for a licensing market between developers and rights holders. It highlighted the court's statement that a rights holder's decision not to license does not disprove the existence of a market.4 The AAP had argued that point in an amicus brief filed in November 2025.4

The decision is already showing up in other cases. Music publishers suing Anthropic filed a short brief in late September pointing to it as support for their position. That is predictable advocacy, but it shows the dynamic ahead: plaintiffs will cite the fourth-factor holding broadly, and defendants will rely on footnote 7.

What it means for legal tech and AI builders

For legal technology, the result is stark. ROSS shut down in 2021, citing litigation costs, but kept pursuing the appeal.3 The case now returns to Delaware for trial on whether some copyrights have expired, on claims involving the remaining headnotes, and on damages.8 Meanwhile, the incumbent's ability to train its own AI on its proprietary editorial work became part of the reason the challenger lost.510 Startups trying to compete with established research platforms now face a clear rule: build from public-domain primary sources, or pay for access.

The opinion has limits that coverage sometimes blurs. IPWatchdog's account highlights the district court's finding that Westlaw's Key Number System was original.2 Inside Tech Law and Ashurst Perkins Coie point out that the appellate panel treated that issue as forfeited because ROSS never briefed it. The Third Circuit therefore made no separate holding that the taxonomy is copyrightable.1216 The court also left open whether headnotes that quote an opinion word for word are protected.14

Commentators broadly agree on the practical lessons. Fair use is weakest when the trained product competes in the source work's own market. Internal marketing that pitches a product as a substitute will become evidence. Shortcuts in getting data can taint the whole analysis.14 Because the court attributed the contractors' copying to ROSS, developers should require provenance, warranties, and indemnities from data vendors. Baker Botts adds that license terms themselves may become evidence in fair-use cases.10

The bottom line

The Third Circuit decided a narrow case narrowly, but not all of its reasoning stays narrow. Its first-factor analysis depends on the specific facts of a non-generative, copy-to-compete product. Its fourth-factor analysis supports the training-data licensing economy that rights holders have been building. Until another appeals court weighs in, the main conflict in AI copyright law is between a Third Circuit willing to count licensing markets as harm and California trial judges who refuse to. The next generative AI appeal will probably have to choose between them.