AI Copyright Cases: Third Circuit Ruling Reshapes 2026 Fight

A first appellate answer, with a carve-out

For three years, courts handled AI copyright disputes almost entirely at the trial level, and judges often disagreed. That changed on September 29, 2026. The U.S. Court of Appeals for the Third Circuit affirmed Judge Stephanos Bibas's summary-judgment ruling for Thomson Reuters against ROSS Intelligence. The court held that Westlaw headnotes are protected by copyright and that ROSS's use of them to train a legal-research tool was not fair use.36 Several analyses describe it as the first federal appellate decision on fair use in AI training.37

The headlines claiming that AI training is not fair use go further than the court did. The panel called the dispute "no more than an ordinary copyright case."6 In a footnote, it separated ROSS from the generative-AI cases, including In re OpenAI, Bartz v. Anthropic and Kadrey v. Meta. The court said ROSS's product created no new expression and only returned passages from existing judicial opinions.6 Law-firm commentary agrees that the core generative-AI question is still unresolved.39

The reactions show how much is left open. The RIAA and NMPA backed Thomson Reuters on appeal. They argued that the lower court's reasoning means training a model to compete with and substitute for copyrighted works can never be fair use.5 An AI-industry trade group also called the ruling a win, saying it implicitly confirmed that highly transformative generative training is likely fair use.1 Both sides can claim the same opinion because the court decided a narrow set of facts.

What the court actually weighed

The reasoning matters more than the result for the many cases still pending. The court found that 2,243 headnotes had been copied and that each one showed enough editorial judgment to be protected.9 On the first fair-use factor, it called ROSS's use "minimally transformative, at best." Both products served the same end, helping users find responsive legal material, and the machine-learning step in between did not change that.6 The court also rejected convenience as a defense. ROSS could have built its training data from public court opinions and chose headnotes because they were easy to use.9

The fourth factor, market harm, is where the ruling reaches furthest. The panel found harm in the existing legal-research market and in a developing market for licensing headnotes as AI training data. As evidence of that second market, it pointed to Thomson Reuters's own use of headnotes to train its AI search products.3 One analysis argues that a developer's fair-use risk now grows as licensing markets develop for the kind of content it trains on.2 On that view, news publishers, academic journals and music rights groups that have signed AI licensing deals have new support for their claims.2 The music industry already presented its deals with Stability AI, Udio, KLAY and ElevenLabs to the court as proof that a training-data market exists.5

The reasoning cuts both ways. If licensing deals help prove market harm, every deal a rights holder signs makes its own lawsuit stronger. That gives rights holders a reason to license widely and AI developers a reason to fear that licensing some content will weaken their fair-use defense for everything else.

The district-court split remains

The Third Circuit ruling joins an unsettled set of trial-court decisions. In Bartz v. Anthropic, Judge William Alsup called training on books "spectacularly" transformative. He also held that acquiring and storing more than seven million pirated books was not protected.8 In Kadrey v. Meta, Judge Vince Chhabria ruled for Meta on training even though the books came from shadow libraries. He criticized the plaintiffs' market-harm arguments as "half-hearted."810 Accounts differ on the procedure: one summary describes a partial motion to dismiss and another describes partial summary judgment. Either way, claims tied to torrent "seeding" are still active.810

Bartz became the reference point for settlements. Anthropic paid $1.5 billion, roughly $3,000 per work.8 Trackers disagree on when final approval came. One reports July 20, 2026, with a list of 482,460 works and all 53 objections overruled. Another lists May 2026.1317 Music publishers are using the Bartz figures as a minimum, not a maximum, and are seeking $3.1 billion from Anthropic over more than 20,000 shadow-library files.17

The case data supports the view that Bartz's split between training and acquisition is changing how complaints are written. One review found 47 dated U.S. filings in 2026, up from 32 in 2025 and eight in 2023. Fourteen of the 2026 filings include piracy or unauthorized-acquisition allegations, compared with six in 2025.16 Newer complaints name specific shadow libraries, torrent sites and paywalled sources.16 Plaintiffs appear to be building cases around how material was obtained, the theory Bartz allowed, instead of attacking training directly.

Other claim types are weakening. On September 16, the Ninth Circuit affirmed dismissal of the DMCA § 1202(b) claim in Doe v. GitHub. It reasoned that a system producing new code without copyright notices does not remove or alter anything.12 Raw Story's copyright-management-information case failed on standing.16 Counts of the docket also vary: one tracker lists 144 cases, the Copyright Alliance counted more than 110 by May, and global estimates range from 126 to 188 depending on definitions.121617

The New York Times case is the next big test

The case most likely to answer the generative-AI question is moving toward a decision. The news plaintiffs, OpenAI and Microsoft filed cross-motions for summary judgment on September 4 in the consolidated case before Judge Sidney Stein.2224 The publishers argue that infringement happened at five stages: acquiring the articles, training on them, Microsoft's Bing-based grounding of chatbot answers, Microsoft's outputs, and the companies supplying copies to each other.24 OpenAI asks the court to rule that pretraining is fair use. According to its brief, its expert searched 20 million ChatGPT logs and found 24 instances of verbatim reproduction.24

The federal government has taken a side. In a statement of interest filed September 1, the Justice Department argued that training is transformative and that the benefits of AI "far outweigh any competitive harm." It tied that position to national security.2124 It was the department's first statement on AI companies' use of copyrighted material.21 The publishers asked the court to give the filing no weight.24 It is advocacy, not a ruling, and it addresses training separately from acquisition and outputs.28 Those are the issues the Third Circuit and Bartz treated as the most dangerous for AI companies.

Reports on the schedule differ. One tracker lists replies due November 6. A detailed case summary says a September 24 scheduling order set oppositions for October 23 and replies for November 20.2224 An undecided sanctions motion is also pending. The Times and other plaintiffs accuse OpenAI of misstating whether it could search its data and of failing to preserve logs. They want the court to bar OpenAI from relying on the 20-million-log sample. OpenAI denies the accusations.24 The Supreme Court's March 2026 Cox v. Sony decision has already narrowed the secondary-liability claims. Judge Stein refused to let the Times add new contributory-infringement theories against Microsoft.2415

The docket keeps growing. Since August, USA Today Co., the Seattle Times and Newsday, a group of 26 local publishers, and wikiHow have all filed suits that point back to the OpenAI litigation.11 Several earlier related cases are on hold until summary judgment is decided.12

Legal tech's own reckoning

The first appellate ruling came from inside the legal industry: a legal publisher sued a legal-research startup. Thomson Reuters said the decision confirms that AI companies cannot free-ride on human ingenuity.2 The ruling also strengthens the position of established legal-research companies that own the editorial material AI challengers want to train on.

Lawyers' own use of AI is creating a separate liability problem. One public database counted 2,046 court cases involving AI-fabricated citations by September 21, 2026, up from about 200 a year earlier.35 Penalties are rising. The Sixth Circuit fined two lawyers $15,000 each.31 A Mississippi judge cancelled a trial and barred two lawyers from the district for two years.35 Nebraska issued the first indefinite bar suspension tied to AI filings.31 England's Solicitors Disciplinary Tribunal struck a lawyer off the register for the first time over AI misuse.32 A 2026 survey found that 69 percent of legal professionals use general-purpose AI, while only 9 percent of firms have a written, enforced AI policy.38

The copyright cases and the sanctions cases point to the same lesson. Courts are not creating special rules for AI. The Third Circuit said courts "regularly apply" long-standing copyright principles to new technology.9 Judges in the citation cases sanction lawyers for failing to verify AI output, not for using AI.33

For AI developers, the outlook after September is clearer about risk than about outcomes. Training a product that competes with its source on curated, licensable material is now the riskiest position, and courts treat how content was acquired as a separate source of liability. Whether general-purpose generative training is fair use will probably be decided first by Judge Stein, and later by the Second or Ninth Circuit.