Fair Use or Free Pass? Court Backs Anthropic in AI Copyright Fight

 A decision in Bartz v. Anthropic (Case No. C 24-05417 WHA) (ND Calif.) represents one of the first decisions about "fair use" under copyright law, and how it applies to generative AI. 

Anthropic – the well-known provider of AI tools such as Claude, an AI chatbot – has developed its products, in part, by downloading millions of copyrighted books in digital form from pirate sites, without paying anything to the publishers or authors.

Anthropic also purchased millions of hard-copy books (some of which overlapped with the titles from the pirate sites), and literally tore off the bindings, scanned every page, and stored the content in digitized form in order to amass a central library to train its large language models (LLMs) to power its AI services, such as Claude.

The authors of some of these books sued Anthropic for copyright infringement, as part of a class action.  The preliminary issue was whether any of Anthropic's uses of these booked qualified as "fair use" under Section 107 of the US Copyright Act.  Remember this is a US case, about US copyright law; the analogous concept in Canadian copyright law is "fair dealing" and this court's reasoning may be considered when this issue arises in Canada.

There was no dispute that the authors who sued Anthropic were the owners of copyright in the books in question, and no argument that Anthropic copied the books without authorization of the copyright holders, and used the copyright-protected works for training purposes, as part of an LLM to provide their AI products.

The court found that Anthropic's use of the books in this way to train Claude was "exceedingly transformative" and qualified as "fair use" under Section 107 of the Copyright Act. In other words, the court did not consider it to be an infringing use for Anthropic to download pirated books, without the permission of the copyright holders, and use those pirated copies to train their AI products.

Secondly, the digitization of the books by Anthropic (the ones purchased, scanned, destroyed and discarded) was also considered by the court as "fair use" for different reasons. It qualified as a "fair use" because all Anthropic did was replace the print copies it had purchased for its central library with "more convenient space-saving and searchable digital copies" — without adding new copies, creating new works, or redistributing existing copies. Put another way, Anthropic merely converted the physical hard-copy books into digital form, without copying or creating new copies, or selling the copies. This 'format conversion' was considered a fair use of the copyright-protected works, even if done without the permission of the copyright owners.

The court did decide Anthropic had no entitlement to create a permanent, general-purpose library. This was not itself a "fair use" which excused Anthropic's piracy, and the court deferred this issue to an assessment of damages. 

In a major blow to copyright holders, the main message of this decision is that the copying and use of copyright-protected works by an AI company for the training of LLMs for AI products is not an infringement of copyright in the US, since such use qualifies as "fair use" under copyright law. Of course, this applies to the specific facts of the Anthropic case, and as other decisions are released, or decisions are appealed, a consensus will emerge of how copyright law will apply to this area.

If you have questions about copyright law and its application to AI, or require guidance on how fair use or "fair dealing" may impact your AI development or content protection strategy, please contact Richard Stobbe for tailored advice and support in this rapidly evolving area.