A groundbreaking judgment in the Bartz versus Anthropic case: US judge William Alsup qualifies the training of AI systems with the help of copyright-protected works as fair use.
In August 2024, American authors Andrea Bartz, Charles Graerber and Kirk Wallace Johnson brought a lawsuit against Anthropic, the company behind the AI chatbot Claude. According to the authors, Anthropic was infringing their copyrights by including copies of their work in a general data library. This library consists of a dataset the company uses to train its Large Language Models (LLMs).
The then Head of Partnerships was instructed by Anthropic to collect ‘all the books in the world’ in order to compose the dataset. The library consists of texts from both legally purchased and illegally downloaded books (or copies of books). Anthropic bought millions of physical, mostly second-hand, books, which it scanned by hand and converted into PDF files. In addition, it downloaded millions of copies of books from the websites Book3, Library Genenis and Pirate Library Mirror.
Then Anthropic selected the text most suitable for training specific LLMs. These texts were merged into subsets. The texts in these subsets were tokenised and then used to train its LLMs.
According to Anthropic, (illegal) copying of books was justified, because the company considered this necessary to train its LLMs.
It was assessed on the basis of the American fair use exception whether or not Anthropic could copy the books. The court looked at several factors, such as the purpose and nature of the use, the nature of the protected work, and the quantity and proportionality of the materials used.
The judgment of judge Alsup is clear. Training AI with the help of texts from books is allowed under the American fair use exception.
Alsup has a more nuanced opinion on composing the general library. According to Anthropic, it was possible that it wanted to use the library for other purposes than training LLMs. However, this use of the copyright-protected works also falls under the fair use exception. The decisive factor here is that the physical books were discarded after scanning and that the digital versions were not disseminated. The digital files replaced the physical copies. This use falls under transformative use, or use that serves a new purpose or has a new meaning.
Alsup’s judgment on the illegal copies is different. Anthropic downloaded over seven million books without paying for it. According to Alsup there is no justification for illegally downloading books that are also for sale legally or can be obtained in other permitted ways. Moreover, Anthropic retained these illegal copies not only for training LLMs. As mentioned above, these copies too were kept in the library for other potential purposes.
In other pending proceedings too, the fair use exception has been raised as a defence against accusations of copyright infringement. It remains to be seen whether this ruling will be guiding to the opinions of judges in those lawsuits.
