Si chiama “destructive scanning” ed è il modo più rapido per allenare i modelli di intelligenza artificiale. Tantissimi i testi rari sacrifi
AI Companies Are Buying Millions of Books Only to Destroy Them
It’s called “destructive scanning,” and it’s the fastest way to train artificial intelligence models. Countless rare books are being sacrificed
Caterina Malanetto July 31, 2026 Updated at 8:06 p.m.
If Fahrenheit 451 had been set in 2026, perhaps the protagonist would have worked for an AI company. In the novel, Montag destroys books because critical thinking is not permitted in the dystopian society in which he lives. Today, in a world often as absurd as the one described by Bradbury, there is no desire to restrict culture in order to prevent people from thinking. Books, on the other hand, are transformed into fuel to feed the “digital brain”: not burned, but scanned, assimilated, and fed to machines so they can learn to write, respond, and create. And then reduced to pulp.
It all starts with massive purchases of books. All kinds: Russian novels, romance novels, fantasy, religious texts. It’s strange—who would buy so many books all at once? Pieter de Vries, an antiquarian bookseller in Haarlem, thinks so too when he sees a bizarre request in his inbox: “I want to buy 3,000 titles,” writes a certain Nataly. “I’m participating in a new project focused on collecting books in multiple languages, currently mainly in English.” In his story reported by The Telegraph, de Vries says he dismissed it, thinking, “It’s probably a scam.” But it wasn’t.
The Anthropic Case
But let’s take a step back. A few weeks ago, Anthropic (Claude’s parent company) found itself at the center of a legal battle, accused of using thousands of pirated books to train its artificial intelligence models. A $1.5 billion settlement, and the case is closed. But the most interesting point is another: according to the court, artificial intelligence companies cannot illegally copy books. However, they may purchase them in print, digitize them, and use that content to train their own models, provided that the physical copies are subsequently destroyed. From a copyright perspective, therefore, this would not constitute a new copy, but rather a “transformation” of the work.
“Destructive scanning”
That is why the antiquarian bookseller in Haarlem received that request. Training a language model requires enormous amounts of text, preferably of high quality. This process is called “destructive scanning,” and there are two methods. The first is fast: the book is taken apart, the spine is cut off, the binding is removed, and the pages are scanned. It is an efficient process, but irreversible: the book becomes unusable. The second method, on the other hand, involves scanning page by page. The book is preserved, but it takes a long time and requires human intervention. Obviously, the major AI companies have preferred the first option.
Why Artificial Intelligence Seeks “Human” Books
The goal is to teach the AI a language that is as rich, natural, and close to that of a “cultured” human as possible. At Anthropic, it’s known as Project Panama, but this isn’t the only company involved. Other major players in the industry, including Google, have also been working for some time to digitize vast collections of books.
These texts thus become valuable resources, and those published before 2022 are even more valuable. In fact, there is a risk that texts written after that date may have been even partially contaminated by artificial intelligence with what is known as “AI Slop”: texts generated by language models, if used in a sort of “cannibalism” to feed new AIs, risk progressively degrading their quality.
Great Novels Sacrificed for an Email
No one is spared—neither common books nor rare volumes. It is this last detail that provokes the most outrage. But alongside those who are outraged, there are also those who immediately sensed an opportunity. Take the ISBNdb platform, for example, which has launched a section dedicated to bulk purchases—a nod to companies interested in acquiring huge stocks of books without drawing too much attention. And so, books that have survived centuries of war, passed from hand to hand, are shredded so that an AI can learn to write a better email or the perfect essay.
