We Tracked a Shipment of Rare Books. It Ended at an Amazon AI Training Facility

Amazon is buying monolithic quantities of books, scanning them for AI training data, and destroying them successful the process. 

A 404 Media investigation was capable to uncover Amazon’s book buying operation, which hasn’t been antecedently reported, by placing a search instrumentality successful a uncommon book we suspected would beryllium acquired by an AI institution for training data, and pursuing it astir the state to its last destination. 

That last destination was an Amazon storage successful Las Vegas, Nevada. Amazon labor who activity astatine this location opportunity each they do is person monolithic shipments of printed books which they past trim the bindings disconnected successful bid to scan the books much quickly. The printed book is destroyed successful the process. The logo of the Amazon squad that useful astatine this warehouse, called VGT3, is simply a dinosaur, brandishing its teeth and pinch a book successful its hands.

“Amazon purchases books done commercialized channels to thief create and amended the products and services our customers use,” an Amazon spokesperson told maine successful a statement. 

The world’s AI companies are perpetually looking for, and spending utmost resources to locate, much worldly to train their AI models. With books, that sometimes intends destroying them successful the process, thing that ample parts of the nationalist person spoken up against, and which we tin now corroborate Amazon is doing.

In July, I published a story astir booksellers who reported a humanities spike successful income starting successful the past year. They suspected this spike successful income was owed to AI companies acquiring immoderate books they tin successful hunt of caller training data. Printed books are valuable arsenic training information because a batch of the matter they incorporate is not readily disposable connected the internet, which AI companies person already scraped. The information is besides conveniently organized and, if the book was printed earlier 2022, is guaranteed to beryllium free of AI-generated text, which tin make immoderate AI exemplary that is trained connected it worse via a recursive process called “model collapse.” 

These booksellers suspected AI companies were down these ample bulk purchases because of the precocious number of books they were buying, the seemingly random prime of books, and the truth that these buyers, dissimilar libraries and universities, did not look value delicate astatine all. But booksellers couldn’t opportunity for definite who was down the ample purchases because the marketplaces wherever they waste their books support the buyers anonymous. When an bid comes in, a bookseller ships the sold books to a storage operated by the marketplaces, wherever books are sorted and past sent to the buyer.

In July, 1 bookseller told maine they received a very ample bid of astir 1,000 books connected Biblio, 1 of these marketplaces. The seller agreed to put an Apple AirTag provided by 404 Media successful 1 of the books included successful this bid truthful we could spot wherever the book was going. And by extension, which company, AI aliases otherwise, was down this monolithic order.

404 Media granted the bookseller anonymity because they worried sharing this accusation would harm their business. Biblio did not respond to a petition for comment.

The bookseller sent the shipment to Biblio, and it arrived astatine a California airport. It past traveled by level to Milwaukee International Airport successful Wisconsin. Later that day, the book traveled to a storage belonging to a specialized shipping and distribution institution called Trifinity, correct extracurricular of Kenosha Regional Airport, astir 30 miles south. The book remained location for astir 2 weeks, astatine which constituent it began walking westbound by what appeared to beryllium a truck. I could spot the book recreation via the road and walk a nighttime complete astatine a trucking recreation halfway astir Grand Junction, Colorado. The adjacent day, the book arrived astatine its last destination, an Amazon storage called LAS8 successful Las Vegas.

LAS8 is 1 of respective ample Amazon warehouses successful the area, each pinch a different specialty. LAS7, correct crossed the street, for example, is simply a fulfilment center, while LAS8 appears to mostly run arsenic 1 of Amazon’s “print connected demand” operations, which will people and vessel books to Amazon shoppers arsenic they are buying them. Initially, I was confused astir why the book would get there, but Amazon labor who activity astatine this location and who talk moving conditions location pinch different Amazon labor online explicate that the the northbound extremity of the LAS8 warehouse, wherever I saw the book arrived, housed a different Amazon cognition pinch a different code: VGT3. 

Many Amazon warehouses person unsocial symbols to correspond that circumstantial site. VGT3’s symbol, painted connected the entranceway to the tract and inside, is of a Tyrannosaurus rex, the monolithic carnivorous dinosaur, pinch its rima open, holding an unfastened book. 

The entranceway to Amazon's VGT3 facility.

“I activity astatine VGT3 present successful Vegas, and each we do is scan books,” 1 Amazon worker wrote connected a forum for Amazon workers. “Some are assigned to trim books, and others spell to person wherever they get books and scan the barroom codes. We didn't person rates, but now we do, but it's not stressful.”

“Working astatine VGT3 is bully each we do is scan books,” different Amazon worker wrote. “It's truthful cool.”

I saw VGT3 labor online talk astir really moving successful this cognition is simply a bully but boring occupation because workers do the same, easy, repetitive tasks each day. I besides saw immoderate chat indicating that VGT3 jobs are desirable for this reason, and because the storage sometimes offered nighttime shifts. Earlier this year, immoderate labor expressed concerns that Amazon would unopen down the tract because they had worked done the books they had and weren’t getting capable shipments of caller books, though the tract is still operational today.

That labor are scanning the barcodes aliases ISBNs connected books — a unsocial serial number fixed to each published book — earlier scanning their contented gives further credence to different mentation put distant by booksellers: AI companies are trying to methodically scan each printed book successful the world by moving done the database of ISBNs. One bookseller told maine they suspected this was the lawsuit because the very ample orders they were getting ne'er included very uncommon books that do not person ISBNs.

Elsewhere online successful 2024, booksellers who waste their books connected Amazon said they received a spike successful orders to beryllium shipped straight to the VGT3 location for a customer named “Amazon FC.” This was truthful different fixed its size they suspected the orders were immoderate benignant of scam. An Amazon typical chimed successful to opportunity they checked and that the orders were legitimate.

Like different awesome tech companies, Amazon is processing its ain ample connection models (LLMs). Amazon considers its family of models branded Nova to beryllium “frontier” models, meaning it considers them to beryllium competitory pinch different cutting separator LLMs from Google, OpenAI, and Anthropic. These models require monolithic amounts of training information successful the shape of human-written text. Internally, Amazon besides utilized an AI coding supplier called Kiro to create its ain software.  

We first learned that AI companies wanted to scan millions of books for training information because of a suit from book authors against Anthropic, which revealed Anthropic’s “Project Panama.” The extremity of the task was to get books from commercialized bookselling marketplaces, trim the spines disconnected the books, and scan them. It’s imaginable to scan books without destroying them, but cutting the spine makes it cheaper and faster. Additionally, the judge successful the suit ruled it was adjacent use and not a copyright usurpation for Anthropic to scan a book for training information successful portion because it destroyed the original, printed copy. Essentially, it’s the customer’s correct to return beingness media and shop it digitally, and destroying the original transcript intends that transcript isn’t duplicated and resold, and isn’t cutting into the publisher’s business. 

We’re not revealing the titles of the books included successful the shipment we tracked, but they are rare, meaning location are not galore copies of them successful circulation. Sometimes that’s because not galore copies of them were ever printed, and sometimes because they are successful a overseas connection not galore group speak. As the bookseller who sold them told me, location are not galore group successful the world who would attraction astir them successful the aforesaid measurement group mightiness attraction astir the first version of Oliver Twist, but that doesn’t mean they’re not valuable. 

“There are different types of value,” the bookseller said. “There's monetary value, obviously, but location are a batch of different types of value. There's humanities value, intelligence value, sentimental value. All sorts of things, and each of those the AI companies don't attraction about. They conscionable want the contented arsenic a bunch of words strung together.”

Amazon provided its connection supra but did not respond to questions astir why it trim the books it scanned, really galore accommodation for illustration this it has crossed the world, and their consequence to the backlash from group upset that AI companies are destroying books.