AI Copyright Ruling Rejects Fair Use, Raising Stakes for AI Users

The first appellate ruling has arrived

AI copyright litigation has been running for more than three years, and on September 29, 2026 it produced its first answer from a federal appeals court. The Third Circuit upheld Judge Stephanos Bibas's summary judgment for Thomson Reuters. It held that Westlaw's headnotes are protected by copyright and that Ross Intelligence was not making fair use when it trained its AI legal-research tool on them.19 The panel's initial order consisted of a single word, "AFFIRMED," and the full opinion was unsealed the following day.9 Ross was a legal AI startup that has since shut down, and the court found it had set out to compete directly with Westlaw using Westlaw's own material.6

The ruling comes as the caseload keeps growing. One public tracker counts 248 AI lawsuits against 76 companies. It lists 145 of them as copyright cases and 69 as class actions, with OpenAI named in 55, Google in 25, and Meta and Microsoft in 21 each.25 A different tracker counts 144 copyright cases.2 Other outlets give lower figures, such as "almost 100" suits or more than 70.1427 The totals vary with who is counting and what qualifies, but they all point the same way. New complaints are arriving every few weeks. USA Today Co. sued OpenAI on October 8. Mississippi's Emmerich Newspapers sued Microsoft and OpenAI on October 2. A second suit by Universal Music against Suno, covering 60,202 recordings, was filed in September.1

For any business that uses, fine-tunes or buys AI tools, these cases now bear directly on how it operates.

Narrow holding, broad shadow

The key question is how far Ross reaches, and accounts differ more here than anywhere else.

The court tried hard to keep its ruling small. Judge Tamika Montgomery-Reeves wrote that the case might look like it was about the future of legal AI, but was really "no more than an ordinary copyright case."6 She noted that Ross's system could not produce original expression. She contrasted it with the generative models in Bartz v. Anthropic and indicated that her holding does not conflict with that case.6 Practitioners have mostly read the ruling the same way. One law firm's analysis says it settles a dispute about a non-generative tool and "leaves generative AI for another day."9 A commentator in The Daily Record argued that it should apply only when someone copies protected material to build a product that competes with the copyright owner. He suggested that a general-purpose system trained on the same material might well be found transformative.5 LawNext, as summarized by one tracker, adds that Ross did its copying before generative AI took off.10

Others see much broader consequences. An audio-data industry group argued that the ruling effectively ended fair use as a dependable risk-mitigation strategy for insurers, whatever the type of data.22 The recording industry also clearly sees value in the case. The RIAA and NMPA filed a brief calling the appeal the first chance for an appellate court to decide whether commercial training is fair use when the model competes with what it learned from. They cited Suno's own admission that it trained on "essentially all music files of reasonable quality" on the open internet.7

In this analysis, the narrow reading is right about the law, and the broad reading is right about what happens in practice. Strictly speaking, the opinion does not decide whether training ChatGPT or Llama is lawful. But two parts of it will be quoted in every later case. The court pointed to harm to a developing market for licensing headnotes as AI training data. It also said that, unlike necessity, convenience does not justify copying.3 Any plaintiff that already licenses its content to AI developers now has a stronger market-harm argument. Ross has said it will ask the Supreme Court to review the decision.10

The emerging rule: training may be fine, sourcing is not

When Ross is read alongside the 2025 district court rulings, a usable framework appears. In Bartz, Judge William Alsup found that training on lawfully acquired books was fair use, but that Anthropic's library of pirated copies was not. That distinction led to a $1.5 billion settlement. The court gave final approval on July 20, 2026, covering 482,460 works with a 91.3% claims rate and overruling all 53 objections.3 Meta separately won summary judgment against 13 authors because they could not show harm to the market for their books.14

Commentators broadly agree on the lesson. Training itself may be defensible as transformative, but courts examine separately how the data was obtained and kept, and that can create liability even when the training is protected.12 Plaintiffs have adjusted. The May 2026 suit by Hachette, Macmillan, McGraw Hill, Elsevier, Cengage and Scott Turow against Meta accuses it of training Llama on books from Anna's Archive, LibGen, Z-Library and Sci-Hub. It also names Mark Zuckerberg personally.11 The complaint addresses the market-harm gap that sank the earlier authors' case: it alleges that Llama wrote a ten-chapter, nearly 5,000-word sequel to Turow's Innocent.11 A similar coalition sued Google over Gemini in July. That complaint adds an allegation that Google stripped copyright management information to hide what Gemini was trained on.12

Not every theory is working, though. On September 16, the Ninth Circuit upheld dismissal of the DMCA § 1202(b) claim in Doe v. GitHub. It reasoned that a model producing new code without copyright notices does not remove anything.2 After the Supreme Court's decision in Cox Communications v. Sony, some plaintiffs dropped their contributory and vicarious infringement claims.4 The OpenAI multidistrict litigation is also shifting. The court denied OpenAI's motion against Ziff Davis's contributory claim but refused to let the Times and Daily News amend theirs.4 One industry analysis from June argued that the early cases mostly favored fair use.19 That reading now has to account for Ross.

Why companies that only use AI should care

Most businesses do not train frontier models, and the coverage agrees they are not the main targets. A privacy-focused firm notes that companies deploying third-party AI face more limited copyright exposure. That exposure rises when outputs reproduce protected material, when companies fine-tune on data they lack rights to, or when vendor contracts say nothing about where the training data came from.13 Stradling, writing about the Meta suits, laid out the downstream risks: contributory infringement claims, reputational damage, and losing a tool overnight if a court orders it shut down. The plaintiffs' request that Meta destroy unlawfully acquired copies could force a retraining that would disrupt every product built on Llama.11

The risk is also showing up in boardrooms. Three shareholder derivative suits have targeted Microsoft and Adobe executives, alleging they misled investors about AI training practices.1618 A Penn law professor told Bloomberg Law that such suits may succeed regardless of whether infringement is ever proven.18 Microsoft called the claims meritless.18 Unredacted filings in the news-publisher litigation quote a Microsoft applied-science director calling AI web scraping "the largest theft of labor in human history."21 He also warned internally of a "doom loop" that could harm both the company's models and the web.14 None of this is a ruling. Still, one business-advisory site notes that vendors' own documents now describe their training-data supply chain as a risk.21

Copyright is also only one of several legal theories. Plaintiffs are applying wiretap statutes, Illinois's biometric privacy law and employment-discrimination law to chatbots, voice agents and hiring tools.13 One suit alleges Adobe trained Firefly on narrators' voiceprints without the consent BIPA requires.17

Legal tech is where the risk lands first

It fits that the first appellate ruling concerns a legal research tool. Legal AI depends on curated editorial content, which is exactly where Ross found a protectable "creative spark."6 Legal-tech vendors and the law firms that buy from them now face a clear standard. A product built to replace a licensed database, using that database's material, carries the most risk. The practical response is the same across the coverage.

• Document provenance. Record where training and fine-tuning data came from and on what license terms. Do not keep copies acquired outside a clear license.1211
• Read the indemnity clause closely. Google and Anthropic extend indemnity to claims about their own training data. Many providers exclude data the customer supplies, and some require customers to indemnify them for it.23 An indemnity is only as good as the vendor's insurance, and insurers have begun excluding AI liability from standard commercial policies.22
• Check outputs before publishing. The Turow sequel allegation shows that output risk is concrete.11

What comes next

The next milestones are already scheduled. Anthropic's motions to dismiss several cases will be heard on December 17.1 The Suno fair-use summary judgment, which has been delayed repeatedly, is now set for January 2027 and may influence how courts treat text, image and code models.12 Cross-motions for summary judgment in New York Times v. OpenAI were filed on September 4.10 Judge Vince Chhabria declined to send the Meta question up early, so the Ninth Circuit will get it only after a final judgment.34

Until those rulings arrive, the clearest guidance is what Ross and Bartz show together. Courts may accept that training is transformative, but they are not excusing pirated data or products built to replace their sources. For companies, the lowest-risk approach is to know where their training data came from, have the paperwork to prove it, and have a vendor contract that holds up if a court finds the vendor got it wrong.