Technology giants are reportedly acquiring millions of second-hand books, a practice that includes the destructive scanning of their contents for artificial intelligence training. This method, known as 'destructive scanning,' involves removing the book's spine, feeding the pages into industrial scanners, and then destroying the physical copies.
This practice has raised concerns among booksellers and publishers, who worry about the potential destruction of rare and valuable books. Lucy Hingston, a bookseller from Compass Books in Devon, England, expressed dismay after discovering that books purchased from her shop were being used for this purpose. "It's a moral dilemma," Hingston stated, feeling "eaten by the AI monster."
The Publishers Association CEO, Dan Conway, described the bulk purchasing and destruction of books as "shocking" and a form of "brazen theft of copyright materials." He added that these actions harm creative industries and the livelihoods of authors and journalists.
AI companies, such as Anthropic, the creator of the chatbot Claude, have been engaged in acquiring physical books. Court documents revealed that Anthropic's "Project Panama" involved spending millions of dollars to purchase and scan books since early 2024. These books are acquired to train large language models (LLMs), which require vast amounts of text data to learn and generate human-like responses.
LLMs are trained on extensive datasets that include internet content, digitized books, and other media. To broaden their data sources, AI firms are now turning to books that have not previously been digitized or made available online. Anthropic, for instance, previously trained its Claude model on over seven million digitized books sourced from websites where copyrighted material is illegally shared. A recent $1.5 billion settlement with authors over this practice led the company to pivot to acquiring physical books.
Under the U.S. "first sale doctrine," established by a Supreme Court ruling in 1908, the buyer of a physical product is legally permitted to do what they wish with it, including destroying it. This legal principle allows AI companies to scan and discard physical books without facing copyright infringement claims, although ethical concerns remain prominent.
The practice draws parallels to historical book burnings, though proponents argue the intent is dissemination rather than censorship. Nevertheless, the destruction of once-cherished volumes for the financial gain of tech companies is viewed by many as dystopian.
Companies are reportedly targeting second-hand book dealers, where books can be acquired at a low cost. Many firms appear to be interested only in books published before 2022, the year generative AI gained mainstream popularity. This is because books from this period are less likely to contain AI-written text, which can lead to "model collapse"—a phenomenon where AI models trained on their own output drift from reality, creating an "AI echo chamber."
Finn Windle, an employee at The Minster Gate Bookshop in York, described receiving a large, disparate order from a U.S.-based company called Zoom Books. Similarly, Helen Bott of Treasure Chest Books in Felixstowe reported massive orders from Zoom Books, noting the random and specific nature of the requests, often for non-fiction titles.
A spokesperson for Anthropic stated that Claude is trained on a mix of publicly available web data, commercially acquired datasets, and self-generated data. They emphasized that their data acquisition programs do not involve buying and destroying rare or antiquarian books. An Amazon spokesperson commented that the company purchases books through commercial channels to improve its products and services.
Even Elon Musk has weighed in, suggesting that AI teams should preserve rare books by scanning them manually rather than resorting to destructive methods.