Even Amazon is destroying rare books to train its AI
To train their language models, AI companies are buying old books by the palette. What brings short-term profits to bookstores raises critical copyright issues and concerns about the loss of cultural heritage.
AI companies are desperately searching for material to train their language models. «Material» is anything that contains language – from social media posts to newspaper articles to YouTube videos. Companies don’t even stop at printed literature.
A few weeks ago, antique bookstores in Germany and elsewhere reported a striking surge in bulk orders. Older books, mainly unsellable shelf huggers, were suddenly in high demand – by the thousands. (Link in German). They suspected the books were going to be digitised and then destroyed. And that’s exactly what U.S. company Anthropic did. Anthropic is developing the chatbot Claude. However, it was unclear whether other AI companies were taking the same approach.
Amazon’s buying books on a large scale
Now, clear evidence points to Amazon, the online giant that started out as an online bookstore in 1994. A research team from the U.S. website 404 Media hid an AirTag in a shipment of about 1,000 books that a distributor had ordered from a bookstore. This allowed them to track the shipment as it passed through several waypoints on its journey to an Amazon fulfilment centre in Las Vegas.
You couldn’t miss the huge logo displayed at the entrance to the building where the delivery arrived. It features the branch code, VGT3, above a Tyrannosaurus rex holding an open book in its claws. According to 404 Media, the main task of the VGT3 branch is systematic digitisation of books.
After VGT3 employees were discussing working conditions and job responsibilities online, it was clear they were systematically processing books. The ISBN codes are scanned, the spines are trimmed, and the pages are then digitised. According to the employees, in some cases this is even done during day and night shifts.
Amazon confirmed the purchase of books in a statement to 404 Media. It said this helps the company «develop and improve the products and services our customers use». It’s obvious what this means, since Amazon operates its own AI models.
A thirst for human-written texts
A language model is only as good as its training. Its quality, in turn, depends on the quantity and quality of the data used to produce it. Edited texts from magazines and books are considered to be of higher quality than social media articles by randomers on X. The Washington Post reported that the Facebook-owned company Meta sent an internal e-mail in 2024 saying that books were vital in its competition with other AI firms.
AI companies have an interest in tapping into as many high-quality text sources as possible. Many publicly available digital sources have already been systematically analysed in recent years. Whether this was always done lawfully is the subject of numerous court cases. Thousands of authors and publishers are accusing AI companies of using texts for training purposes without permission.
At the same time, websites with high-quality digital content are increasingly incorporating restrictions that prevent data collectors from accessing the content automatically. In contrast, there are enormous quantities of books, journals, and magazines that were written by humans before powerful language models appeared (2022 onwards) and have never been digitised. It’s precisely this vast repository of data on paper that is becoming the focus of attention for AI companies.
The patterns in bulk orders suggest that distributors are automatically processing ISBN-based title lists. This allows them to search specifically for books that haven’t been purchased yet and order them. At any cost, too. Antiquarian booksellers report that the books sold this way are usually unsellable and therefore very cheap, while the costs for shipping and customs duties amount to many times that amount.
Scan and destroy for AI
The suspicion that AI companies are behind these orders hasn’t exactly been plucked out of thin air. Anthropic, the developer of the chatbot Claude, was in court for exactly that this year. Internal documents released during this year’s proceedings reveal that Anthropic launched Project Panama in 2024. Anthropic’s goal was to «scan all the books in the world in a destructive manner,» as the documents indicate.
To efficiently handle the enormous task of processing the millions of books purchased since then, the spines were cut off to separate the pages. High-performance scanners can quickly digitise individual pages in this way. The physical texts were then disposed of or recycled. This practice is safer from a legal standpoint than using online book archives because the books were acquired legitimately. This distinguishes this case from the use of illegally obtained copies.
Judge William Alsup ruled in the lawsuit against Anthropic that the use of text from lawfully acquired and digitised sources for training AI models could be considered a highly transformative use and therefore falls under the fair use doctrine widely recognised in the USA. Transformative use means that the scanned texts are not used like books, but serve a different purpose i.e. they’re now used to train AI.
However, some issues still need to be clarified in future court rulings, such as when authors can prove specific economic damages.
Cultural treasure or surplus stock that nobody wants any more?
The issue of book destruction, on the other hand, poses no legal problems. The owner may destroy property acquired legally.
Anthropic denied in court that it had purchased and destroyed rare or historically significant books. The report by 404 Media also confirms that the shipment in question did not contain any books of cultural or historical significance.
It sounds like a win-win situation when booksellers are happy to sell large orders of merchandise that no one else wants any more. However, these orders involve rare works – for example, because they were printed in small print runs or published in a rarely spoken language.
Librarians, historians and antiquarian book dealers’ associations see this as a problem. In some cases, low-cost destructive scanning can permanently destroy physical cultural artefacts. This presents some booksellers with a moral dilemma: do they want to be complicit in knowingly exposing rare books to destruction?
Feels just as comfortable in front of a gaming PC as she does in a hammock in the garden. Likes the Roman Empire, container ships and science fiction books. Focuses mostly on unearthing news stories about IT and smart products.
Interesting facts about products, behind-the-scenes looks at manufacturers and deep-dives on interesting people.
Show all
