Anthropic is buying and digitizing thousands of printed books every day to train its artificial intelligence models as part of a project called Project Panama. According to court documents, the company cuts the covers of the books and converts them into digital format using high-speed scanners. After this process, the books cannot be restored.
The goal of the project is to collect content that is not available on the Internet or has not been previously used to train AI models. The documents state that it plans to scan 500,000 to 2 million books over the course of six months, which is about 11,000 books per day.
The main reason for this is the depletion of high-quality, human-written texts on the Internet. Researchers warn that if AI models are trained primarily on texts created by other AIs, their quality may deteriorate over time, knowledge diversity may decrease, and errors may increase.
Anthropic says it purchased the books legally and that they were digitized only for internal training purposes. A US federal court has also noted that in certain cases this approach can be assessed under the principle of "fair use". However, using pirated digital libraries is considered a violation of the law.
As AI companies' demand for paper books and other unique resources increases, new businesses have begun to form in this area. Some companies are already supplying AI labs with rare books that are not available in digital databases.
