The Last Clean Ore: AI’s Great Book Harvest
Updated
Something strange has been transpiring in secondhand bookshops around the world. Over the past several months, bookstore owners from Northumberland to Galway to Wellington have received enormous orders from buyers that they have never heard of, for large quantities of books that have nothing in common. The orders are not for a themed collection, and they aren’t from a scholar filling in gaps in a field of study. Instead, they are wild, haphazard lots of titles, the likes of an Estonian translation of a le Carré spy novel, a 1983 issue of a naval magazine, and a manual on eighteenth-century African farm tools. Adding to the strangeness, the buyers hide behind unremarkable aliases and ship their orders to anonymous freight warehouses. And most tellingly, they are just fine paying full retail, never asking for the bulk discount that any real book dealer would expect and demand. As an example of what is happening, a shop in England has quietly moved 6,000 books to these clandestine buyers since January. To date, the essentially anonymous orders keep coming.
So who are these mysterious buyers purchasing random books in bulk? Booksellers around the globe have a hunch, and, no, it is not a mysterious collector wanting to preserve books. They suspect the buyers are artificial intelligence (AI) companies that are gobbling up physical books to feed into their learning machines. Of course, suspicion is not proof, and for many of the hundreds of orders, there is none. Some of the randomly requested titles are freely available online, which would make buying a battered used copy seem pointless if text were all anyone was after. Does every strange order trace back to an AI lab? It is impossible to say with certainty. However, the suspicion is entirely reasonable, because we do know, for a fact, that this exact scenario—AI companies buying large quantities of books—has already been done on a massive scale.
Yes, indeed. The proof that AI is buying books in bulk was recently revealed in a courtroom. Last year, in a copyright lawsuit, internal documents forced into public record revealed that one leading AI company, Anthropic, had spent tens of millions of dollars buying physical books for the sole purpose of destroying them. Make no mistake: the means behind Anthropic’s madness are industrial and brutal in their efficiency. First, a heavy-duty machine slices the spine clean off of each book. It happens exactly as one would imagine, given the explanation. Once the book’s spine has been removed, the bound volume is reduced to a stack of pages ready to run at high speed through an automatic scanner, at a speed much faster than any intact book could ever be photographed. Software then converts the scanned images into raw digital text, which is poured into the vast pool of writing that AI is trained on. And the book, now a gutted pile of loose paper of no further use to anyone, is thrown away and pulped. Yes, ground down into cardboard and toilet paper.
The sequence at play here is significant. Why? Because the book is not scanned and returned to its place on a bookstore shelf. It is consumed by AI. The physical object, sometimes a book so sacred as a rare or out-of-print edition that will never be printed again, is treated as disposable paper to be destroyed once the words are extracted. And the destruction is not an unfortunate side effect of AI wanting the material for its own learning. The destruction is the point. Shockingly, in Bartz v. Anthropic (June 2025), a federal judge ruled that Anthropic’s buy-and-destroy method was legal precisely because the original books were destroyed. Essentially, the court stated that if a book is bought and scanned, and both are then kept, then one copy suddenly becomes two copies, which looks like illegal copying. But if a book is purchased, scanned, and then destroyed, there is still just one copy, simply in a new format. In other words, the law, as it now stands, doesn’t just allow destroying the book. It rewards it because keeping the original hardbound book would have been the legal risk. Wow, what great incentive to shred a physical record rather than preserve it.
There is a reason that old books in particular are now such a prized target. AI models are trained on human writing, and until recently, that was easy to come by. However, the internet is now flooded with AI-generated text, and training an AI machine on the output of other machines degrades the content (which in itself is telling). Apparently, it is akin to making a photocopy of a photocopy until the image becomes unrecognizable. Thus, Big Tech’s AI companies are ravenous for writing that they are certain was produced by actual human beings. That means older content in books published roughly before 2022, before machines began contaminating our world with its own words. These older books are genuine human thought, unpoisoned by the rapidly expanding imitation. The pre-AI human record is rapidly becoming the last clean ore in the mine, and Big Tech is destroying it as fast as they can buy it.
When the anonymity and aliases of the mysterious book orders are stripped away, we are faced with an industry poised to rule humanity, treating the written inheritance of the human race as cheap feedstock. Centuries of authors, the careful and the obscure alike, reduced to mere food for AI consumption. Think about it. Books that survived house fires, moves, and decades on a dusty shelf are being purchased in bulk not to be read and enjoyed, but to be sliced apart and fed to a machine that will then produce its own cheap imitation of the very thing it consumed. Will it alter history along the way? As human history is reduced to material to be harvested, that frightening idea seems quite plausible.