Bartz v. Anthropic is a copyright class action by authors against Anthropic PBC, the developer of the Claude models. The authors alleged that Anthropic copied their books without permission to build a central library and to train its models. On 2025-06-23 Judge William Alsup held that training on the books was fair use and that converting purchased print books to digital files was fair use, but denied summary judgment on the copies downloaded from pirate libraries, which he set for trial.[1]
Background
According to the court, Anthropic downloaded millions of books from pirate sources, including Books3, Library Genesis and the Pirate Library Mirror, and separately bought millions of print books, removed the bindings and scanned them. It kept the copies in a central library intended to be retained “forever” and drew on it to train its models.[1][2] The court noted that in spring 2024 a co-founder had emailed major publishers about licensing books for AI training but let the conversations wither, and that another technology company soon licensed copies from one major publisher.[1] How Anthropic licenses its models is described in Anthropic licensing.
The dispute
The authors argued that Anthropic’s copying infringed their copyrights and that training displaced an emerging market for licensing their works for training large language models. Anthropic argued fair use and said that transaction costs would exceed its expected benefit from licensing deals.[1]
Decision or outcome
On the licensing argument the court assumed the authors were right that a market could develop, but held that “such a market for that use is not one the Copyright Act entitles Authors to exploit”.[1] It granted summary judgment for Anthropic on training as fair use and on the print-to-digital format change, but denied it for the pirated library copies, stating that the later purchase of a copy of a book earlier taken would not absolve liability and might affect statutory damages.[1] On 2025-07-17 the court certified a class of copyright owners of registered books in the versions of Library Genesis and the Pirate Library Mirror that Anthropic downloaded, and denied certification of a Books3 class and a scanned books class.[2]
On 2025-10-17 the court’s memorandum opinion described the preliminary approval of a class settlement. Class members give up claims about past use of their works up to the point of any AI output, but not claims about past outputs or future conduct, and receive an award of about USD 3,000 per work, less costs and fees, plus an injunction requiring destruction of copies derived from two pirate libraries. The order records 482,460 works identified.[3]
Significance for software licensing and SAM practice
The ruling is a copyright fair use decision rather than a contract case, and it is a district court ruling that did not go to trial or appeal. For organisations that buy or build AI tools, it shows that the source of training data and the existence of licence options were part of the court’s analysis. See also Kadrey v. Meta Platforms for a ruling by another judge in the same district and Thomson Reuters v. Ross Intelligence for a different result.
Lessons learned
- Source matters. Training on lawfully acquired copies was treated as fair use, while the pirated library copies were set for trial.[1]
- A licensing market was not protected. The court held that authors are not entitled to the market for licensing works for AI training under the Copyright Act.[1]
- Licensing discussions are evidence. The court cited Anthropic’s early licensing inquiries and the fact that another company obtained a licence from a major publisher.[1]
- Settlements can be limited in scope. The settlement released past training conduct but not future conduct or past outputs.[3]