Amazon is reportedly buying massive quantities of rare books, scanning their pages to train its AI models and destroying the tomes in the process — joining Anthropic, Meta and other tech giants in operating massive book scan-and-destroy operations.
The goal, according to a report by news site 404 Media, is to train AI on well-written texts as opposed to low-quality internet slop, which AI models have already for the most part devoured. Slicing machines sever the spines of the books to speed up the process and cut costs, according to reports that have detailed operations by Amazon and others playing out in non-descript warehouses.
404 Media said it uncovered Amazon’s book-buying operation by hiding a tracking device in a rare volume that it suspected an AI firm would purchase. Its movements were traced across states to its final destination – an Amazon warehouse in Las Vegas.
“Amazon purchases books through commercial channels to help develop and improve the products and services our customers use,” an Amazon spokesperson said.
The bookseller – which 404 Media kept anonymous – agreed to slip an Apple AirTag provided by the publication into a huge order the vendor received in July for around 1,000 works through a book marketplace.
Warehouse workers at the Las Vegas location said their days are spent taking in heaping shipments of printed books and preparing them for AI scanning. The process – also detailed in a January Washington Post article on Anthropic and other AI labs’ own book-scanning efforts – involves automated machines that chop the bindings off the orphaned volumes.
The mascot of the Amazon team that works at the warehouse, named VGT3, is a Tyrannosaurus rex showing its chompers while holding a book, according to 404 Media.
“I work at VGT3 here in Vegas, and all we do is scan books,” one Amazon employee wrote on an online forum for the company’s workers. “Some are assigned to cut books, and others go to receive where they get books and scan the bar codes.”
Another laborer wrote: “Working at VGT3 is nice all we do is scan books,” adding, “It’s so cool.”
Some booksellers logged huge spikes in sales over the past year – which they suspected were driven by AI companies seeking to get ahold of new training data, according to 404 Media.
While they couldn’t be sure, booksellers believed AI giants were behind the bulk orders because of the exceptional number of books they were buying and the seemingly random selection of volumes, 404 Media reported. Another giveaway was that the buyers didn’t appear to care about prices, unlike libraries and universities.
Books have become particularly valuable to AI labs, especially ones authored prior to 2022 – the year OpenAI first released ChatGPT – as they are certain to not contain AI-generated text. Companies have said training models on AI text can actually wreck havoc on models, a concept known as “model collapse.”
One Anthropic co-founder believed that training AI on books could teach them “how to write well” rather than imitating “low-quality internet speak,” according to a January Washington Post report. A 2024 Meta internal email called securing a digital cache of books “essential” to competing in the AI race.
Anthropic executives in early 2024 ramped up their book-scanning project – code named “Project Panama” – and sought to keep it under wraps, according to the Washington Post.
“Project Panama is our effort to destructively scan all the books in the world,” Anthropic executives said in an internal planning document that was unsealed in legal filings in January. “We don’t want it to be known that we are working on this.”
The report noted that Anthropic initially considered purchasing books from used bookstores and libraries including the New York Public Library — or “a new library that is chronically underfunded,” according to documents cited by WaPo.