When books become AI training data
A book takes years to write, print, sell, and find its way onto a shelf. It can take a machine only seconds to scan a page. The investigation behind this story follows a trail of unusual book orders, a hidden tracker, and a warehouse in Las Vegas. Let’s look at what that trail establishes—and where the evidence leaves questions open.
The reported process is stark: books arrive in bulk, their spines are cut off, and the loose pages move through high-speed scanners. The text becomes digital data. The paper copy cannot be put back together.
Watch the investigation to see the reporting behind this story.
Following the evidence
The investigation began with unusual orders for old, out-of-print, and less common titles. In one reported case, a seller received an order for nearly 1,200 books worth about €20,000. The journalists placed a tracker in a book from a shipment and followed it to an Amazon facility. Their reporting for 404 Media, including interviews with workers, described a site where books were disbound and scanned.
That trail is compelling, but the details matter. A large order is not proof on its own of who ultimately wants every title or how each book will be used. The report connects this facility to AI training through its tracking and interviews. Company statements and other individual cases belong in the same account, because one investigation does not establish that every bulk order follows the same path.
The investigation also lands in the middle of a wider story. Court documents described in The Washington Post’s reporting revealed Anthropic’s Project Panama, an effort to acquire physical books and scan them destructively. Anthropic has said its data acquisition programs do not buy and destroy rare or antiquarian books. Those statements sit alongside the reporting; they do not make every book order or every scanning operation the same case.
Why go back to paper?
AI models learn patterns from examples. Books can provide long, carefully edited explanations, uncommon perspectives, and material that may not be readily available in digital collections. They are one more source of text for developers building systems that can write, answer questions, and reason through language.
The web is enormous, but “everything is already online” is not quite true. Some books were never digitized, and some digital copies are difficult to find or use. That makes physical collections attractive as a source of text. It also makes the way those collections are handled worth paying attention to.
There is something unsettling about a book being treated as a temporary container: valuable while its words are trapped inside, disposable once the words have been extracted. A scan can preserve the text, but it does not automatically preserve the object, its notes, its paper, or the fact that someone could have opened that particular copy and read it.
What gets lost in a faster scan?
Cutting a spine is an efficient way to feed flat pages into a scanner. It is also irreversible for that copy. That tradeoff becomes harder to shrug off when the book is scarce, out of print, or simply part of a human collection with a history that does not fit into a training file.
The public discussion often jumps from “the book was scanned” to “the machine now knows what the book means.” Those are different things. A language model learns statistical patterns from text; it does not read with a human life behind it, connect an idea to personal experience, or decide what deserves to be remembered. More text can broaden what a system can produce. It does not make the system a person.
And more information, by itself, does not make a person wiser either. We learn when we work with an idea: connect it to what we know, ask where it came from, explain it in our own words, and see whether we can recall it later. A pile of pages can hold knowledge. Understanding takes participation.
Help people learn, too
That is the part of AI and books that matters to us at Reader++. We want to make people as capable and curious as the tools they use—not by turning a reader into a data pipeline, but by helping ideas make sense to a human mind.
Reader++ keeps an explanation close to the passage that prompted it. Ask a question about a paragraph, explore how one concept connects to another, then try to explain the idea or recall it without looking. The source stays nearby, so you can check the answer against the book instead of taking a fluent response on faith.
The machines may be learning from books at industrial speed. We can still make room for the slower, more human work: reading carefully, following a question, and making the knowledge our own.
Detective