AI Copyright Lawsuits Hit 144 as Third Circuit Rejects Fair Use

A docket that keeps growing

One tracker now counts 144 copyright lawsuits against AI companies. They make up 56% of the 257 AI-related cases it monitors. Of those 144, 113 are still active or pending and 31 have been resolved or otherwise ended.22 The pace has not slowed this autumn. USA Today Co. sued OpenAI on October 8. Emmerich Newspapers and other plaintiffs sued Microsoft and OpenAI in Mississippi on October 2. A group of 26 local publishers led by Times Publishing filed in mid-September and asked to be coordinated with the consolidated OpenAI copyright proceeding.21 The Seattle Times and Newsday, wikiHow, and Sony Music Publishing (suing Anthropic over lyrics) all filed between late August and early September. Universal Music Group filed a second suit against Suno covering 60,202 recordings.21

The exact size of the docket depends on who is counting. One rival tracker shows 45 curated entries and flags 17 of them as still awaiting docket verification.23 In June, the same publisher put the number of active or recently resolved cases at more than 70.24 Another outlet says global trackers count anywhere from 126 to 188 cases, depending on how narrowly they define the category.28 The spread reflects different counting rules more than disagreement about the facts. Every count, though, points the same way: news publishers, music companies, authors and image licensors keep filing.

Most of these cases won't reach a merits ruling soon. A growing share are on hold while one proceeding plays out. Cases including Richner Communications v. Microsoft, Carreyrou v. OpenAI, and Britannica and Merriam-Webster v. OpenAI have been stayed until summary-judgment rulings come down in the OpenAI multidistrict litigation in Manhattan.26 Summary-judgment briefing in that MDL began on September 4.2629 Discovery there has been aggressive. In January the court ordered OpenAI to produce 20 million ChatGPT output logs.27

The first appellate answer on AI training

The biggest development of the season came from a court that most generative-AI litigants are not in. On September 29, the U.S. Court of Appeals for the Third Circuit affirmed Judge Stephanos Bibas's ruling for Thomson Reuters against ROSS Intelligence. The court held that Westlaw's editorial headnotes are protected by copyright and that ROSS's use of them to train a competing legal research tool was not fair use.36 Several law-firm analyses describe it as the first federal appellate decision to address fair use in AI training.310

The court's reasoning is narrow, and the details matter. It found that 2,243 copied headnotes each showed enough editorial judgment to be protected.79 It called ROSS's use "minimally transformative, at best": both products existed to help users find relevant legal material, and putting a machine-learning step in the middle did not change that purpose.6 It also rejected the argument that the copying was a necessary intermediate step. ROSS could have built its training data from public court opinions and used the headnotes because that was easier. The court said ease does not justify copying.39 Three of the four statutory fair-use factors went against ROSS.6

The most consequential part may be the fourth factor, market harm. The court found harm in two markets. One was the existing legal-research market. The other was a developing market for licensing headnotes as AI training data, which Thomson Reuters was already entering by training its own AI products on that content.310 This logic reaches well beyond legal databases. Publishers, journal operators and music rights groups that have signed training-data licenses now have appellate support for the argument that a protectable licensing market exists.2 The music industry was watching closely. The RIAA and NMPA filed in support of Thomson Reuters and pointed to licensing deals between major labels and AI music companies as evidence that such a market is real.5

A loss that both sides are calling a win

Commentary on the ruling splits less on what it held than on how far it reaches. The court said in a footnote that ROSS's system was not generative. It could only return passages from existing opinions, not create new expression. That separated it from In re OpenAI, Bartz v. Anthropic and Kadrey v. Meta.6 As a result, Ballard Spahr and others conclude that the central generative-AI question is still open.34 The AI industry has embraced that reading. A Chamber of Progress executive argued that the ruling implicitly supports treating highly transformative, general-purpose model training as fair use.1 One coalition that backed ROSS in an amicus brief was more sober, saying this first appellate attempt at AI fair use "has hit a dead end."2

The music industry reads the decision more broadly. Its amicus position was that training a model to compete with and substitute for the works it learned from can never be fair use.5 The federal government is on the other side. In a September 1 filing in the OpenAI litigation, the Justice Department urged the court to treat training on copyrighted text, apart from output-side use, as highly transformative fair use. The Third Circuit opinion discusses that position.24

On balance, the ruling is narrow on its holding and broad in its effect. It binds only federal courts in Delaware, New Jersey and Pennsylvania. No Ninth Circuit court, where much of the generative-AI litigation sits, is bound by it.2 Still, its factor-four reasoning about licensing markets gives plaintiffs a ready template. Each new licensing deal makes that argument a little stronger.

District courts remain split

The Third Circuit ruling also arrives without a settled trial-court consensus to build on. In Bartz, Judge Alsup called LLM training "spectacularly" transformative fair use but found that acquiring and keeping more than seven million pirated books was not protected.8 In Kadrey, a judge in the same district went further, holding that training was fair use regardless of whether the source material was obtained legitimately. That judge also criticized the plaintiffs' market-harm arguments as weak.8 One summary describes this split between training and acquisition as the framework most other cases now rely on.28

The financial stakes are concentrated in the acquisition question. Bartz reached final judgment on July 20, 2026, with a $1.5 billion settlement.23 That works out to roughly $3,113 per book across about 482,000 works.24 Anthropic still faces active music-publisher claims, including a suit valued at $3.1 billion.23 Meta's remaining exposure in Kadrey centers on claims about "seeding" pirated files through torrenting.8

Settlements also keep important questions from being decided. Warner settled with Suno and Universal settled with Udio, which created licensing templates without producing a ruling on the merits. That gives Sony's continuing cases against both companies outsized importance.25 Outside the U.S., Suno lost a first-instance case in Germany on July 31.23 In the UK, by contrast, Stability AI defeated Getty's copyright claim in November 2025, though Getty is appealing.25

The legal-tech angle: a lawsuit about lawyers' tools

The first appellate AI-training precedent comes from legal technology, and the timing matters. Courts are dealing with AI from two directions: as a defendant in copyright cases and as a tool lawyers use inside their own filings.

The second problem is getting worse. A database maintained by researcher Damien Charlotin had recorded more than 1,148 U.S. cases of lawyer hallucinations as of a June update.11 One July count put the global total at 1,490.16 Appellate courts have stepped in. On June 3, the Ninth Circuit issued its first ruling on lawyers' responsibility for AI errors. It sanctioned two attorneys $2,500 each, suspended them from practicing before that court for six months, and required them to disclose any use of AI in their filings for two years.12 The Sixth Circuit ordered two lawyers to pay $15,000 each.19 In England, the Solicitors Disciplinary Tribunal issued its first strike-off in an AI case in September.19 Courts across jurisdictions give the same message: penalties apply to unverified citations, not to using AI. Lawyers who are open about mistakes are treated more leniently than those who hide them.1115

Adoption has not slowed. Harvey reportedly reached $300 million in annual recurring revenue at an $11 billion valuation, even as hallucination cases piled up.16 Purpose-built tools still make mistakes, though. Stanford testing found error rates above 17% for Lexis+ AI and above 34% for Westlaw's AI-assisted research.17

What to watch

The next decisive moment is summary judgment in the OpenAI MDL. Several stayed cases are waiting on it, and it will put the generative-AI question that the Third Circuit set aside to a direct test.26 Until then, the ROSS decision sets a clear practical standard. AI developers that copy curated, proprietary content to build a product that competes with it, in a market where licenses are available, face the weakest fair-use position.210 That makes licensing deals more valuable, even though the courts have not yet decided whether generative-AI training is fair use.