AI Training Fair Use Loses as Third Circuit Backs Data Licensing

The first appellate answer on AI training

For more than two years, the copyright fight over artificial intelligence has run almost entirely through trial courts. That changed on September 29, 2026, when the U.S. Court of Appeals for the Third Circuit became the first federal appeals court to decide whether copying protected material to train an AI system qualifies as fair use.2 It held that it does not, at least on the facts in front of it.18

In Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., Judge Tamika Montgomery-Reeves wrote for a unanimous panel that also included Judges L. Felipe Restrepo and Emil J. Bove III. The court upheld a ruling that Westlaw's headnotes are copyrightable and that ROSS infringed them by using them to train a competing legal research tool.14 The panel's makeup is worth noting. Restrepo was appointed by President Obama and Bove by President Trump, so the outcome does not fall along partisan lines.1

The headline result was widely expected. The part that will carry furthest is the court's statement that licensing text as AI training data is a real derivative market, and that Thomson Reuters had a claim to it even though it had never licensed its headnotes to anyone.16

How a defunct startup ended up making precedent

Thomson Reuters sued in 2020. It alleged that ROSS, after Westlaw refused it access, obtained headnotes through a contractor called LegalEase Solutions.1 LegalEase and a subcontractor wrote about 25,000 "Bulk Memos." Each paired a legal question built from a Westlaw headnote with passages from judicial opinions, and ROSS turned those memos into machine-readable training data.14 According to the drafters, they used headnotes because headnotes made framing questions easy. The passages they ranked "great" were most often the same passages Westlaw linked to the headnote.18

ROSS closed in 2021, worn down by litigation costs, but it kept pursuing the appeal.14 Judge Stephanos Bibas, a Third Circuit judge sitting as the trial judge, first sent most of the case toward a jury in 2023. In February 2025 he reversed himself and granted Thomson Reuters partial summary judgment on 2,243 headnotes.4 He then certified the originality and fair use questions for immediate appeal.3 The case is not over. A Delaware trial still has to address whether some of the copyrights have expired, what happens with the remaining headnotes, and damages.8

The opinion was first issued under seal while the parties proposed redactions.5 A public version came out later, and legal-tech outlets were reporting on the unsealed text by mid-October.4

Why ROSS lost on fair use

On the first factor, the court called ROSS's use commercial and "minimally transformative at best." ROSS priced its platform to compete with Westlaw for the same customers.41 The panel granted that training an AI was an intermediate step that "arguably presents a slight degree of difference in use." It then judged the use by the finished product. Both companies used the headnotes to build and improve a legal research service, so the training step did not change the purpose.13

ROSS relied on software cases such as Google v. Oracle, where copying in the middle of a process was excused. The court said those cases turned on necessity. ROSS already had the judicial opinions, which carry no copyright, and copied the headnotes because that was easier.8 The panel's line that "ease is not a justification for copying" is likely to be quoted by plaintiffs for years.15

The second factor slightly favored ROSS, because headnotes are mostly factual.4 The third factor went against it. ROSS argued that it took only about 0.08% of Westlaw's 28 million headnotes. The court treated each headnote as a separate work and found that ROSS had copied each one in full.9 On this factor the panel parted ways with Bibas, who had sided with ROSS because no headnotes appeared in its output. The appeals court did not discuss output at all.13

The court also gave little weight to ROSS's warnings that enforcing copyright would "halt AI development" or harm national security, noting there was no evidence behind those claims.9 A footnote added that ROSS's attempts to get into Westlaw using investor and student credentials counted against it, to the extent good faith still matters.1618

The licensing-market holding

The fourth factor, effect on the market, is where the decision matters most. The panel found harm in two places. The first was Westlaw's existing legal research business, which ROSS's product was built to replace.2 The second was a "rapidly developing" market for licensing headnotes as AI training data.4 The court pointed to evidence that Thomson Reuters already trains its own AI search products on headnotes. It said the company's choice not to license them to others "does not disprove that a market exists."15 By copying without permission, ROSS "usurped" Thomson Reuters's chance to enter that market.4

Publishers praised this part. The Association of American Publishers said the court rightly found that the AI training market both exists and is growing, and called the decision a basis for a "robust licensing market" between developers and rightsholders.6 Commentary written for rightsholders went further. One firm told content owners to set up and document AI licensing channels now, so they can prove a market exists if they ever have to sue.9

Critics read the same passage as the opinion's weakest point. Authors Alliance argues that the court cited nothing in the record for its claim that the market is "rapidly developing." It also argues that the court never dealt with the circularity problem: if the relevant market is defined as licensing the very use being challenged, every rightsholder loses a fee, and the fourth factor always favors the owner.3 Mitchell Silberberg & Knupp makes a related point. Judge Vince Chhabria, ruling in the case against Meta, said whether such a market exists "is irrelevant" for exactly that reason, and the Third Circuit did not respond to that argument.13

Some commentators defend the court on this point. The National Law Review argues that the opinion does not let any imagined licensing market defeat fair use. It reads the holding as resting on a record showing the market was growing and that Thomson Reuters was already using the headnotes for its own AI.15 That is the fairest reading of what the court said. The critics are still right that the reasoning is brief. In practice, a rightsholder that trains its own models on its own content now has a strong argument that a licensing market exists, whether or not it has sold a single license.

What it means for the generative AI cases

Nearly every analysis agrees that the ruling's reach is limited. ROSS's system was not generative AI. It returned passages from existing opinions and did not produce new expression.4 A footnote distinguished Bartz v. Anthropic, Kadrey v. Meta and the OpenAI multidistrict litigation. It also set aside the Justice Department's statement of interest in the OpenAI case.1216 The panel called the dispute "no more than an ordinary copyright case."8

Mitchell Silberberg & Knupp points to an ambiguity in that footnote. The court gave two reasons ROSS differs from Anthropic: ROSS's tool cannot generate original expression, and it was trained to be a commercial substitute for Westlaw. The court never said which reason does the work.13 That gap matters. If the deciding factor is that ROSS was a direct competitor, the reasoning could apply to generative products that compete with their sources, such as AI legal assistants or news summarizers. One compliance analysis makes that argument, describing the direct-competitor test as the real holding whether or not a model is generative.10 Ballard Spahr takes the narrower view. It reads the opinion as leaving the main generative AI questions open, including whether training resembles human learning and whether a flood of competing output counts as legally relevant market harm.2

On balance, the competition reading is the better one. The court decided the factors by looking at what ROSS was building and whom it was trying to take customers from. Defendants can use the generative-AI footnote to argue the case doesn't apply to them. The reasoning in the main text gives plaintiffs a template for any case where the AI product competes with the work it was trained on.

The legal-tech and policy stakes

For legal technology, the decision protects the editorial layer that companies like Thomson Reuters have built on top of public-domain law. The court rejected ROSS's argument that protecting headnotes gives Westlaw a monopoly over the law. It pointed out that headnotes are not judicial opinions, and the opinions themselves remain free.14 Startups that want to compete with established research platforms now face a clear choice. They can build training data from primary sources themselves or pay for the incumbents' annotations.

More broadly, data licensing is becoming the expected route. One industry analysis notes that the Anthropic settlement put a price per work on unlicensed training data. It describes ROSS as an appeals court rejecting fair use where training fed a competing product.17 Advisers on both sides are giving the same practical advice: document where training data comes from, ask whether copyrighted material is actually necessary, and look for licensed or public-domain alternatives.157

The ruling does not make AI training unlawful in general.7 It does mean that AI is not an automatic defense. Courts will judge the use by the product it creates, and when that product competes with the source, a licensing market can be found before anyone has ever sold a license.