Former Book Bought to Train AI, Physical Copy Reportedly Destroyed

Illustration of a book used as a material for artificial intelligence training amid copyright debates. Photo: Pexels.

JAKARTA - Artificial intelligence companies are reportedly buying thousands of used books, scanning their contents to train AI models, then destroying or recycling their physical copies. The practice is fueling a debate over copyright.

According to a report by Anadolu Agency, quoted on Tuesday, July 14, antique book dealers in a number of European countries have received an unusual number of orders from the Canadian company, Zoom Books, since the beginning of the year.

The company is said to be looking for many non-fiction and academic books published in the 1970s. In one transaction, the number can reach dozens to hundreds of books from one seller.

The books were reportedly sent to a temporary warehouse in Germany before being taken to Canada and the United States.

The Washington Post said that some of the books were processed with a destructive scanning method. The volumes were removed, the pages were scanned using a high-speed industrial machine, and the rest of the physical material was recycled.

The newspaper also reported that Anthropic spent tens of millions of US dollars through "Project Panama" to buy and digitize millions of printed books.

Court documents cited by The Washington Post said Anthropic recruited experts who had previously been involved in Google's book digitization project.

Zoom Books denied participating in digitizing and destroying books for AI development.

This controversy arose because AI companies need a very large amount of material to train models, which are systems that learn from collections of text and data.

Jannis Lennartz from Humboldt University Berlin said AI developers are starting to look at physical books because digital sources for training materials are becoming more limited.

In his essay titled AI Is Eating the Book World, Lennartz said the increasing demand for used books reflected unresolved legal issues and broader cultural concerns.

He said a number of court decisions in the United States in the Anthropic and Meta cases tend to favor AI companies through the doctrine of fair use. This doctrine under certain conditions allows the use of copyrighted works without permission, but not unlimited permission.

The EU approach is different. The Digital Single Market Directive allows for text and data mining, i.e. the processing of large amounts of material to find patterns or train computer systems, under certain conditions.

The right holder can still refuse their work to be used. However, uncertainty arises in old books that were published before the rule came into force. Many editions do not contain a disclaimer, while some publishers or authors can no longer use the right.

A similar debate occurred when Google digitized more than 40 million books. A US court eventually ruled that the project was a fair use because the limited search and snippet databases were considered to provide new functions without replacing the original book market.

The rejection of the author and publisher has not receded. Earlier this year, thousands of writers published a symbolic blank book titled Don't Steal This Book to protest the use of works without permission or payment.

In December 2025, Amazon also introduced a Kindle feature that allows readers to interact with books using AI.

A number of authors are concerned that the contents of the book are used to train AI models. Amazon states that the feature does not use the contents of the book for model training and is only a development of the Kindle search function.