AI Companies Are Buying Books in Bulk — and Destroying Them
404 Media has published a report raising serious questions about how modern AI companies obtain training data.
According to the report, AI companies have begun purchasing vast quantities of old physical books. The reason is simple: books are one of the largest sources of “clean” text written by humans rather than generated by neural networks.
Books published before 2022 are particularly valuable because they predate the widespread availability of AI-generated content. Using them helps prevent AI models from being trained on text that was itself generated by other AI systems.
For example, ISBNdb offers AI labs a service that can source anywhere from 1,000 to 1 million books in a single order. The books are then scanned and converted into datasets for AI training.
To digitize such enormous quantities of books quickly, the books are effectively destroyed. Hydraulic cutting equipment removes the binding, the pages are separated into individual sheets, and high-speed industrial scanners convert them into PDFs. The remaining paper is sent for recycling.
Anthropic is cited as one example. Court documents reportedly show that the company purchased large quantities of physical books, cut them apart, and scanned them to train its Claude model. The company allegedly attempted to keep the process confidential.
OpenAI, Google, xAI, and other companies are not named in the report. According to 404 Media, ISBNdb’s customers are reluctant to disclose that they are purchasing and destroying millions of books for AI training.
Elon Musk, however, responded to the report:
“I asked the SpaceX AI team to preserve all the rare books in the library and digitize them using traditional methods rather than simply cutting off their bindings and scanning them.”
Turning physical books into datasets for AI training raises an unsettling question: perhaps we are already living in the dystopian future we once imagined.
404 Media has published a report raising serious questions about how modern AI companies obtain training data.
According to the report, AI companies have begun purchasing vast quantities of old physical books. The reason is simple: books are one of the largest sources of “clean” text written by humans rather than generated by neural networks.
Books published before 2022 are particularly valuable because they predate the widespread availability of AI-generated content. Using them helps prevent AI models from being trained on text that was itself generated by other AI systems.
For example, ISBNdb offers AI labs a service that can source anywhere from 1,000 to 1 million books in a single order. The books are then scanned and converted into datasets for AI training.
To digitize such enormous quantities of books quickly, the books are effectively destroyed. Hydraulic cutting equipment removes the binding, the pages are separated into individual sheets, and high-speed industrial scanners convert them into PDFs. The remaining paper is sent for recycling.
Anthropic is cited as one example. Court documents reportedly show that the company purchased large quantities of physical books, cut them apart, and scanned them to train its Claude model. The company allegedly attempted to keep the process confidential.
OpenAI, Google, xAI, and other companies are not named in the report. According to 404 Media, ISBNdb’s customers are reluctant to disclose that they are purchasing and destroying millions of books for AI training.
Elon Musk, however, responded to the report:
“I asked the SpaceX AI team to preserve all the rare books in the library and digitize them using traditional methods rather than simply cutting off their bindings and scanning them.”
Turning physical books into datasets for AI training raises an unsettling question: perhaps we are already living in the dystopian future we once imagined.