AI Training Fair Use Fails First Appellate Test in Westlaw Case

The first appellate answer rejects fair use, but only on narrow facts

For more than two years, AI copyright fights in the United States were decided only in trial courts, and those courts disagreed. That ended on September 29, 2026. A three-judge panel of the U.S. Court of Appeals for the Third Circuit affirmed that Ross Intelligence infringed Thomson Reuters' copyrights when it used Westlaw headnotes to train an AI-powered legal research tool. It also held that the copying was not fair use.3 Commentators widely describe it as the first federal appellate decision to address fair use in AI training directly.45

The ruling is less sweeping than early headlines suggested. Ross's product did not generate text. It was a search tool that returned passages from existing judicial opinions, and the court built its reasoning around that fact.5 It took six years from Thomson Reuters' May 2020 filing to get here, and the case turned on 2,243 headnotes.8 My reading is that this is a real but limited win for rights holders. It tells generative AI developers what the courts will look at next more than it resolves their cases.

What the Third Circuit actually held

The panel opinion is credited to Judge Tamika Montgomery-Reeves, who called the dispute "no more than an ordinary copyright case."20 The panel first found that the headnotes cleared copyright's low bar for originality. Editors chose which legal points to include and how to phrase them, and that showed enough "creative spark."5 On fair use, three of the four statutory factors went against Ross. Only the nature of the copyrighted work tilted slightly its way.5

Two factors carried the most weight. Under the first, the purpose and character of the use, the court found Ross's use commercial and only minimally transformative. Westlaw's headnotes help lawyers find case law, and Ross copied them so its system could learn to do the same thing.22 Under the fourth, market effect, the court found harm in two places: the existing legal-research market, where Ross competed head-to-head with Westlaw, and an emerging market for licensing headnotes as AI training data.5 One account adds that the panel went further than the district court on the third factor. Because each headnote is its own copyrighted work, copying one in full counts as copying the whole work.22

The panel also refused to treat AI as a special case. It rejected Ross's warnings that ruling against it would stall AI development or hurt national security, and said new technology does not give anyone "carte blanche" to ignore copyright.7 Some practical lessons came out of the record too. The copying was done by a contractor and a subcontractor, yet Ross was held responsible. Evidence that Ross got Westlaw access through borrowed credentials undercut any claim that it acted in good faith.7

The procedural path drew attention as well. The panel first issued a one-word affirmance with its opinion under seal, and the opinion became public the next day.3 Accounts differ on the timing: one tech outlet described the decision as still sealed as late as October 18.21 Several detailed legal analyses quote the opinion directly, so the substance appears to be public whatever its formal status.722

The generative AI carve-out

The most important part of the opinion may be what it left undecided. The panel separated Ross's retrieval-only tool from generative systems that can produce original expression, and so did not answer the central question in the generative AI cases.35 One law firm's analysis says the court noted that the evidence in the generative AI cases pointed to more transformation and less direct competition.7 Another account says the panel discussed a Department of Justice statement of interest, filed September 1 in the consolidated OpenAI litigation, which called training large language models "extraordinarily transformative."22

That explains the split in how the ruling is being read. The music industry treated it as a precedent it can build on. In an amicus brief, the RIAA and the National Music Publishers' Association (NMPA) argued that the case would establish that training a model to compete with and substitute for copyrighted works can never be fair use.2 Defense-side reactions were more pointed. Brandon Butler of the Re:Create coalition, which supported Ross, said this "first judicial walk through the AI fair use labyrinth has hit a dead end."22 One columnist argued the decision should be applied narrowly, to cases where the copier builds a product that competes with the copyright owner.10

The narrower reading fits the opinion better. Still, plaintiffs now have an appellate citation they lacked before.26 The most portable part of the ruling is the licensing-market theory. Thomson Reuters' own AI use of its headnotes helped show that a training-data licensing market was plausible. That gives news publishers, journal operators and music rights groups that have signed AI deals a way to argue market harm.22

How it fits with the earlier trial-court rulings

The Third Circuit decision lands among trial-court rulings that point in different directions. In Bartz v. Anthropic, Judge William Alsup held that training on books was fair use. He also held that acquiring and storing more than seven million pirated books was not.1 That case settled for $1.5 billion, roughly $3,000 per work,1 and received final approval on July 20, 2026. It is described as the largest copyright settlement on record, and the order states that it sets no binding precedent.23

Kadrey v. Meta differed on two points. That court held LLM training was fair use regardless of how the materials were obtained, and it criticized the authors' market-harm arguments as "half-hearted."1 The judge stressed that the ruling bound only those thirteen authors and did not declare Meta's training lawful. He suggested that a theory of AI flooding the market with competing works could be stronger.1 Claims tied to Meta's torrent "seeding" remain active.25

Taken together, the cases show a pattern. Courts have so far been willing to treat generative training on lawfully obtained works as transformative, while how the data was acquired and whether the product substitutes for the original are where defendants lose.923 Ross lost on substitution, and Anthropic paid for acquisition. On the trend so far, the next appellate fight will again turn on market harm.8

The docket still ahead

The next wave is large. One public tracker counts 248 AI lawsuits against 76 companies, with OpenAI named in 55 and Google in 25.26 In the consolidated OpenAI litigation in New York, a January 2026 order required OpenAI to produce 20 million output logs.1 Disney and other studios are seeking up to $150,000 per work against Midjourney over character outputs.1 In July, publishers including Hachette, Cengage and Elsevier, along with author Scott Turow, filed a class action over Gemini training.23 Anthropic has also asked for summary judgment on fair use in the music publishers' lyrics case.

Reports on music generators conflict. One tracker lists the labels' claims against Suno and Udio as settled into licensing deals in late 2025.25 Other coverage describes the Suno case as active, with a fair-use summary judgment ruling repeatedly delayed and now set for January 2027.28 The likely explanation is that the settlements covered only some parties and claims, so readers should treat the music docket as unsettled.

On authorship, the law is clearer. The Supreme Court declined on March 2, 2026, to hear Stephen Thaler's appeal, leaving human authorship as a requirement for copyright.1 Protection for works made with AI assistance, as opposed to by AI alone, remains an open question.1 Transparency rules are also getting through court. A federal judge refused to block California's training-data disclosure law, rejecting xAI's trade-secret and free-speech arguments.27

Liability is moving faster for users than for builders

While the copyright questions remain open for AI developers, courts have set a clear rule for lawyers who use AI. In June, the Ninth Circuit issued its first ruling on lawyers' responsibility for AI errors. It fined two attorneys $2,500 each, suspended them from practicing before the court for six months, and required them to disclose AI use in filings for two years.11 The Sixth Circuit's Whiting v. City of Athens imposed $15,000 per attorney, plus fees and double costs.15 In Mississippi, a judge cancelled a trial after finding that lawyers on both sides had cited AI-invented cases.17

The numbers depend on who is counting. One tracker cites about 1,490 decisions worldwide as of May.12 Another counts 2,046 by late September.19 Sources also put the record penalty at either $110,000 or about $109,700.1319 The direction is the same in every count: penalties are rising, and supervisors and local counsel are increasingly held responsible.18 Paid legal tools offer no protection either. A Fifth Circuit sanction involved a lawyer who drafted with Thomson Reuters' CoCounsel and vLex.18

There is an irony here. Westlaw's owner has just won the leading case on what legal AI tools may train on, while courts sanction lawyers for trusting what those tools produce. The common thread in both lines of cases is that courts are applying existing doctrine, not writing new AI-specific rules. Copyright's four fair-use factors and lawyers' duty of candor apply as they did before, whatever technology is involved.