AI companies have begun heavily using books published before 2022 to train language models. The main reason for choosing these books is that they are free from AI-generated content.

Data provider ISBNdb provides developers with access to human-written texts without AI output in bulk. The systems used to digitize these books have become a major topic of debate.

The Destruction and Digitization Process of Books

During the digitization process, the spines of the books are cut, separating the pages, and this process results in the complete destruction of brand-new physical copies. This method, carried out using high-speed scanning equipment, is considered more cost-effective compared to slower methods that preserve the physical integrity of the books.

Although ISBNdb calls this process digital preservation, critics argue that it leads to a cultural loss. The company enforces strict confidentiality agreements to keep its customers’ identities confidential and attempts to manage the repercussions of these business activities.

Experts emphasize that some of the works undergoing these scanning processes are rare copies that have survived through ravages of war and fire. Unlike internet content, these physical works, once destroyed, cannot be recovered.

A recent court decision in the US regarding Anthropic deemed scanning legally purchased books for AI training as fair use under reasonable conditions. In the rulings, it was argued that destroying the physical copy during scanning replaced the existing legal copy, rather than creating new copies.

Concerns about cultural heritage and the future

In response to reactions to this method, Elon Musk stated that he told the SpaceXAI group to keep the few books in libraries and scan them using slower but more reliable methods. This approach is seen as an effort to create an alternative model to the rapid destruction technique.

This method, used to preserve the quality of artificial intelligence models, aims to prevent the loss of quality known as model collapse, which occurs when systems are trained with their own generated content. Furthermore, this process avoids data poisoning techniques used by authors to hinder AI training.

However, this silent destruction process continues to raise significant questions about the preservation of cultural heritage. The physical destruction of a few works while converting them into digital information sets risks creating an irreparable cultural void in the future.

Do you think it’s acceptable to destroy physical copies of a few works in order to digitize them?