AI Copyright Lawsuits Mount After First Appellate Fair Use Ruling

A first appellate answer, and it is narrow

For most of the generative AI era, rulings on whether AI training is legal have come from trial courts. That changed on September 29, 2026. The U.S. Court of Appeals for the Third Circuit upheld Thomson Reuters' win against ROSS Intelligence, holding that Westlaw headnotes are protected by copyright and that ROSS's use of them to train a legal research tool was not fair use.3136 It is widely described as the first federal appellate decision on fair use in AI training.610

The panel's actual reasoning was modest. Judge Tamika Montgomery-Reeves, writing for the court, said that despite the AI framing, the dispute was "no more than an ordinary copyright case."7 The court found ROSS's use commercial and only minimally transformative. It also found harm in two markets: legal research, where ROSS was built to compete with Westlaw, and an emerging market for licensing headnotes as AI training data.3135 Three of the four statutory factors weighed against ROSS. Only the nature of the copied work slightly favored the startup.3110

The most important line for the broader industry may be a footnote. The court separated ROSS's system, which only returned passages from existing judicial opinions, from the generative models at issue in In re OpenAI, Bartz v. Anthropic and Kadrey v. Meta.36 Law firm analyses agree that this leaves the central question open: whether training generative models on books, music or images is fair use.3133

Both sides claim the ruling

Coverage divides less over what the court decided than over what it means. Some headlines read it broadly, saying a court has found that using copyrighted material as training data is not fair use.8 Tech Times called it the "worst end of the risk spectrum" for developers who trained on a competitor's curated database to build a substitute product.33 The tech industry read it the other way. A Chamber of Progress executive argued that the decision implicitly supports fair use for highly transformative, multipurpose generative training.34 Brandon Butler of the Re:Create coalition, which filed a brief backing ROSS, said this first trip through the "AI fair use labyrinth" had "hit a dead end."33

The better reading sits between those positions, and it favors rights holders more than the generative carve-out suggests. ROSS is unlikely to control cases against OpenAI or Anthropic. What it does give plaintiffs is appellate support for the theory that an existing or developing licensing market for training data counts as harm under the fourth factor.3140 Thomson Reuters already trains its own AI search products on headnotes, and the court treated ROSS's unlicensed copying as taking that licensing opportunity away.31 That reasoning could carry over to news archives, academic journals and music catalogs, where licensing deals already exist.33

The district courts have split on exactly this point. In Bartz, Judge William Alsup called LLM training "spectacularly" transformative. He said the fourth factor protects against replacement, not competition, and that authors are not entitled to a training-license market.32 In Kadrey, Judge Vince Chhabria ruled for Meta but said the authors had failed to prove market harm, and he warned that many training uses could be unlawful where that harm is shown.5 Chhabria also rejected the licensing-market argument as circular.40 The Third Circuit has now accepted the argument both judges were skeptical of, at least on facts involving direct competition. ROSS has said it will ask the Supreme Court to review the decision.67 No Ninth Circuit court is bound by it, and that is where most generative AI cases are being heard.33

The docket keeps growing

The filings show no sign of slowing. One public tracker counts 144 copyright cases against AI companies, 113 of them still active or pending.3 A broader count from the same database lists 248 AI lawsuits against 76 companies, with OpenAI named in 55.6 The tracker is anonymous, and its own figures are not fully consistent, so the numbers work better as a rough index than as an official count.6

October alone brought a wave of new complaints. USA Today Co. and 19 affiliated papers sued OpenAI and Microsoft on October 8, seeking more than $250 million. That case is paused until pending summary judgment motions are decided.5 A group of local newspapers filed a similar suit in Mississippi on October 2.4 On October 21, music publisher Reservoir filed two suits in the Northern District of California seeking nearly $340 million in potential statutory damages. One targets Suno over 1,050 compositions; the other targets Anthropic over lyrics from 853 songs.9 The same day, Anthropic and major music publishers argued fair use and licensing-market harm in a separate lyrics case. The judge did not rule.9

The biggest pending case is In re OpenAI, the consolidated author and news publisher litigation before Judge Sidney Stein in Manhattan. All sides filed summary judgment motions on September 4, filed oppositions on October 23, and owe replies by November 20. The Authors Guild expects a hearing in early 2027.5 Courts are also keeping alive claims that avoid fair use entirely. Judges have denied motions by ByteDance and Apple to dismiss claims brought by YouTube creators, including a DMCA anti-circumvention theory based on alleged scraping.13

So far, the clearest money has come from a settlement, not a verdict. The $1.5 billion Bartz v. Anthropic settlement, which resolved claims over pirated books, received final approval and judgment on July 20, 2026. The court awarded about $101.6 million in fees.39 That case suggests where the risk actually concentrates. Alsup found training itself transformative, but he held that acquiring and storing millions of pirated books was not protected.32

Policy: human authorship is settled, training is not

One side of AI intellectual property policy has largely settled. The Supreme Court declined on March 2, 2026, to hear Stephen Thaler's challenge, leaving in place the rule that works generated autonomously by AI cannot be registered.20 The Copyright Office's 2025 guidance says prompts alone do not make someone an author, while human selection, arrangement or modification of AI output can be protected.1119

On training, the Office's May 2025 report concluded that training models to produce content that competes with the originals in their existing markets is unlikely to be fair use.17 The federal government's position has since shifted. The Department of Justice told the Third Circuit in September 2026 that LLM training can be "extraordinarily transformative."33 That is a notable split within the same government. Outside the U.S., the EU AI Act now requires general-purpose model providers to publish training-data summaries and respect copyright opt-outs, regardless of how American courts rule on fair use.1233

The legal profession's own AI problem

It fits that the first appellate AI copyright ruling involves a legal research tool, because lawyers are adopting AI quickly while struggling to supervise it. ILTA's 2026 survey of 508 firms found 94% using or exploring generative AI, up from 80%. Only 6% said the AI features built into their document management systems were enabled and widely used.22 The 8am report found individual use jumped from 31% to 69% in a year. Yet 43% of respondents said their firms had no AI policy and no plans for one, and 54% said their firms offered no training.2425

The surveys measure different things. Thomson Reuters puts law firm use of generative AI at 41%, while Wolters Kluwer reports that more than 90% of lawyers use at least one AI tool.2327 Every dataset, however, shows the same gap: adoption is outrunning governance.

Courts are making that gap expensive. A widely cited database lists 2,202 decisions involving AI-fabricated content as of October 19, 2026, up from about 200 in mid-2025. Of those, 872 involve lawyers.43 Penalties have grown from $5,000 in Mata v. Avianca to about $110,000 in an Oregon case.4644 The Sixth Circuit fined two lawyers $15,000 each.41 A Mississippi judge removed all four lawyers from a case in which both sides filed hallucinated citations.45 In October, a Boston federal judge revoked a California lawyer's permission to appear in her court after he relied on an enterprise version of Claude. She wrote that AI use "does not diminish an attorney's professional and ethical obligations."47 New York's courts have taken a similar approach. Their new guidance does not require lawyers to disclose AI use but makes clear they are fully responsible for verifying every citation.41

The bottom line

Taken together, these developments show courts treating AI as a familiar legal problem. Copying a competitor's curated work to build a substitute is infringement. Filing citations no one checked is sanctionable. Neither result depends on the technology involved. The unsettled question is whether generative training changes the analysis. The answer will likely come first from Judge Stein's courtroom and the Northern District of California, not from the Third Circuit. Until then, the practical lessons are clear: license where a market exists, and verify the output.