AI Copyright Lawsuits Top 140 After First Appellate Fair Use Loss

The appeals court ruling and the case count

A federal appeals court has now ruled on whether training AI on copyrighted work is fair use, and it ruled against the AI developer. That ruling matters more than any single number in the dockets. On September 29, 2026, the Third Circuit affirmed Judge Stephanos Bibas's summary judgment in Thomson Reuters v. Ross Intelligence. The panel held that Westlaw's headnotes are protected by copyright and that ROSS's use of them to train a competing legal-research tool was not fair use.19 No U.S. court of appeals had addressed fair use in AI training before.75

The ruling arrives while lawsuits keep piling up. One public tracker counted 144 AI copyright cases in late September, which was 59% of the 246 AI cases it followed, and 114 of those copyright cases were still active or pending.22 By its October 8 update the same database listed 248 AI lawsuits against 76 companies in 14 jurisdictions, with 145 of them about copyright.5 Santa Clara Law professor Edward Lee's separate list puts the count above 100 copyright suits against AI companies.12 The exact totals depend on who is counting, but every count shows the same trend: new cases are being filed faster than old ones are being resolved.

What the Third Circuit decided

The court first issued a one-page judgment that simply affirmed. Coverage of what came next differs slightly. One early analysis said the opinion, written by Judge Montgomery-Reeves, was under seal while the parties proposed redactions.7 Later reporting said it became public the next day.917 The panel treated the dispute as an ordinary copyright case. It found that the headnotes met the low bar for creativity and that three of the four fair use factors weighed against ROSS.9

The court's main point was about purpose. Westlaw uses headnotes to help lawyers find relevant authority, and ROSS copied them so its system could learn to do the same job. The processing method was different, but the commercial function was the same.17 The panel was also stricter than the trial court on the amount copied. It treated each headnote as a separate work, so copying any one headnote meant copying all of that work.17 It also rejected the argument that copying was justified because it was the easiest route. Unlike earlier software cases, ROSS did not need the headnotes to reach the uncopyrightable court opinions underneath them.4

The district court had already found harm in two markets: legal research, and a possible market for licensing AI training data. That second finding applied whether or not Thomson Reuters had entered that market yet.7 Rights holders are likely to rely heavily on it. The RIAA and NMPA filed a brief supporting Thomson Reuters and argued that new AI licensing deals point to a real market that AI defendants ignore in court while pursuing those deals commercially.10

Why it is also a legal tech story

The first appellate ruling on AI training came from legal technology, not chatbots or image generators. ROSS was a startup building an AI legal search product to compete with the dominant research platform, and it is now defunct.910 The facts were about as bad as they get for a fair use defense: a non-generative system trained on a direct competitor's curated product to do the same commercial task.1617

For legal tech companies building AI research and drafting tools, the practical point is clear. Proprietary editorial content such as summaries, annotations and classifications can be protected even when it describes public-domain material like court opinions. Copying that content to build a competing product is now the highest-risk way to source training data. In one analyst's reading, the ruling targets developers who trained on curated proprietary collections such as legal databases, news archives and academic journals, especially where the owner already licenses AI training access.17 Thomson Reuters said it believes respecting copyright is essential to innovation, and ROSS has said it will ask the Supreme Court to review the case.5

How far the ruling reaches

The sources disagree mainly about how far Ross extends. Music-industry coverage reads it broadly. One column argued that AI "lost" and that Suno, which trained on millions of recordings without licenses, should be worried.8 An advertising-law analysis said the decision reaches well beyond legal research, because unauthorized training can infringe even when the copied material never appears in the AI's output.7

Other analysts read it narrowly, and the opinion itself supports them. The panel said generative AI may produce original expression, so the cases against generative AI companies are left "for another day."9 The court did not hold that training is always infringing.17 Commentators also note that ROSS copied before generative AI was widespread and built a non-generative system.5 The ruling binds only courts within the Third Circuit, though it will likely be cited elsewhere.7

The narrow reading is the more accurate one. Ross does not decide New York Times v. OpenAI or the book cases against Meta and Anthropic. What it does is make market substitution and licensing markets central to how future courts will analyze these cases. The line between generative and non-generative systems is now the main fault line in U.S. AI copyright law.17 That raises the stakes for plaintiffs who can show that a generative model copies their works or competes directly with them.

Washington takes a side

The executive branch weighed in a month earlier. On September 1, the Justice Department filed a 20-page statement of interest in the consolidated OpenAI case in New York. It argued that training large language models is "extraordinarily transformative" and therefore fair use.1918 The filing split model development into three stages: acquiring data, training, and producing outputs. It addressed only training. It did not take a position on acquisition and left output questions to be argued separately.1811 The government also argued that requiring licenses could leave AI development to the largest firms, which are the only ones able to pay.19

The New York Times answered that the administration was siding with a few very large AI companies against American creators.19 The filing is advocacy, not a ruling, and it binds no one.1119 Still, the timing matters. The parties filed cross-motions for summary judgment on fair use before Judge Sidney Stein on September 4, and replies are due November 6.14 The Times has made regurgitation evidence central to its case, meaning outputs that reproduced articles nearly word for word, including paywalled content.16

The U.S. split now looks like this. Bartz, Kadrey and the DOJ treat training as transformative. The Copyright Office says the answer depends on the use. Courts and the government are themselves divided over "market dilution," the idea that floods of AI-generated content harm the market for originals.18 Ross is on the rights holders' side of these debates, but on the narrowest facts.

Piracy, settlements and the money

Much of the litigation is now about how data was obtained, not about training itself. In Bartz v. Anthropic, Judge Araceli Martínez-Olguín gave final approval to a $1.5 billion settlement on July 20, 2026. She overruled all 53 objections, and the settlement covers 482,460 books.4 Earlier rulings had found training transformative while treating pirated copies as a separate liability.16 The DOJ brief notably does not address that acquisition stage.18

That gap is shaping new lawsuits:

• Authors who opted out of the Bartz class are suing Anthropic individually for statutory damages.3
• Plaintiffs keep focusing on shadow libraries such as Library Genesis and Anna's Archive.6
• A default judgment against Anna's Archive imposed $19.5 million in statutory damages.2
• Anthropic has filed motions to dismiss in Carreyrou and Chicken Soup for the Soul, with a hearing set for December 17.1

News publishers are also filing new cases. USA Today sued OpenAI on October 8. Emmerich Newspapers sued Microsoft and OpenAI on October 2. A group of 26 local publishers is seeking to have its September case coordinated with the OpenAI multidistrict litigation.1 Several related suits, including Britannica's, are paused until the summary judgment rulings come out.

Abroad and in music

European courts are focusing more on what models output and memorize. A Munich court found GEMA's case against OpenAI successful at first instance, and OpenAI has appealed.5 In July the same court ruled against Suno, finding that its model could reproduce copyrighted songs.824 One report says the German court even applied U.S. fair use law to copying done in America.8 In the U.S., the next major music decision will come from Chief Judge F. Dennis Saylor IV in Massachusetts.8 Labels keep filing there anyway: UMG filed a second suit against Suno in September covering 60,202 recordings.1

The outlook

The overall picture is consistent across the coverage. Copying to build a product that competes with the source now has appellate authority against it. Training generative models remains contested, with the executive branch backing AI developers. Acquiring pirated data has become the most expensive exposure. Judge Stein's ruling in the OpenAI case is likely to be the next decision that shapes the field, because it will be the first full fair use analysis of generative training where every factor is genuinely disputed.16 Until it arrives, licensing deals and documented data sources are the AI industry's best protection.