404 Media가 희귀 서적에 추적 장치를 넣어 배송 경로를 추적한 결과, 대량 구매된 도서의 최종 목적지는 네바다주 라스베이거스의 아마존 물류창고였습니다. 이 시설의 직원들은 대량으로 들어온 책의 제본을 절단해 스캔한 후 파기하는 작업을 하고 있으며, 이는 AI 학습 데이터 확보를 위한 것으로 확인되었습니다. 인쇄된 책은 인터넷에 없는 텍스트를 담고 있고 AI 생성 텍스트 오염이 없어 AI 학습 데이터로 가치가 높아, AI 기업들이 가격에 민감하지 않게 대량 구매하고 있는 것입니다.
번역된 본문
아마존이 대량의 책을 구매해 AI 학습 데이터로 스캔하고, 그 과정에서 책을 파기하고 있습니다. 404 Media의 취재는 AI 기업이 학습 데이터 확보를 위해 구입할 것으로 예상한 희귀 서적에 추적 장치를 넣어 전국을 따라 최종 목적지까지 추적함으로써, 지금까지 보도된 적 없던 아마존의 도서 구매 작전을 밝혀냈습니다. 그 최종 목적지는 네바다주 라스베이거스의 아마존 물류창고였습니다. 이 시설에서 일하는 아마존 직원들에 따르면, 이들의 일은 대량으로 들어오는 인쇄 서적을 받아 더 빨리 스캔하기 위해 제본을 절단하는 것뿐이라고 합니다. 이 과정에서 인쇄된 책은 파기됩니다. 이 창고에서 근무하는 'VGT3'라는 아마존 팀의 로고는 이빨을 드러내고 손에 책을 든 공룡입니다.
아마존 대변인은 성명을 통해 "아마존은 고객이 사용하는 제품과 서비스를 개발하고 개선하기 위해 상업적 경로를 통해 책을 구매한다"고 밝혔습니다.
전 세계 AI 기업들은 끊임없이, 그리고 막대한 자원을 들여 AI 모델 학습에 사용할 더 많은 자료를 찾고 있습니다. 책의 경우, 그 과정에서 책이 파괴되기도 하는데, 이는 많은 대중이 반대의 목소리를 내왔던 일이며, 이제 아마존이 실제로 이를 하고 있음을 확인할 수 있게 되었습니다.
7월, 저는 지난 1년간 역대급 매출 급증을 경험했다고 전한 서점상들에 대한 기사를 게재했습니다. 이들은 이러한 매출 급증이 AI 기업들이 새로운 학습 데이터를 찾아 닥치는 대로 책을 구매하고 있기 때문이라고 의심했습니다. 인쇄된 책은 학습 데이터로서 가치가 높습니다. 책에 담긴 상당수의 텍스트는 AI 기업들이 이미 긁어간 인터넷에서 쉽게 구할 수 없는 것이기 때문입니다. 또한 데이터가 편리하게 정리되어 있고, 책이 2022년 이전에 인쇄된 것이라면 AI 생성 텍스트가 포함되지 않았음이 보장됩니다. AI 생성 텍스트는 '모델 붕괴(model collapse)'라는 재귀적 과정을 통해 그것으로 학습된 AI 모델을 악화시킬 수 있습니다.
📖 서적을 스캔하는 시설에서 일하고 계신가요? 제보를 기다립니다. 업무용이 아닌 기기에서 Signal(@emanuel.404)로 안전하게 메시지를 보내거나 emanuel@404media.co로 이메일을 보내주세요.
서점상들은 이러한 대량 구매의 배후에 AI 기업이 있다고 의심했습니다. 구매량이 많고, 책 선택이 무작위처럼 보였으며, 도서관이나 대학과 달리 구매자들이 전혀 가격에 민감하지 않았기 때문입니다. 하지만 서점상들은 이런 대량 구매의 배후를 확실히 알 수 없었습니다. 그들이 책을 판매하는 마켓플레이스가 구매자를 익명으로 보호했기 때문입니다. 주문이 들어오면 서점상은 판매된 책을 마켓플레이스가 운영하는 물류창고로 보내고, 그곳에서 책이 분류된 후 구매자에게 전달됩니다.
7월, 한 서점상이 이런 마켓플레이스 중 하나인 Biblio에서 약 1,000권에 달하는 매우 큰 주문을 받았다고 알려왔습니다. 그 판매자는 이 주문에 포함된 책 한 권에 404 Media가 제공한 애플 에어태그를 넣어주기로 했고, 이를 통해 책이 어디로 가는지, 나아가 이 거대한 주문의 배후에 어떤 회사(AI 기업이든 아니든)가 있는지 확인할 수 있었습니다. 404 Media는 이 정보 공개가 사업에 해가 될까 우려한 판매자에게 익명성을 보장했습니다. Biblio는 코멘트 요청에 응답하지 않았습니다.
추적 경로:
캘리포니아 — 우리가 추적한 책 배송물이 캘리포니아의 한 공항에서 출발했습니다.
밀워키 — 다음 목격 지점은 밀워키 미첼 국제공항 바로 근처였습니다. (1901 E College Ave, South Milwaukee, WI 53172)
유통 센터 — 이후 남쪽 30마일 떨어진 케노샤 지역공항 바로 옆, 전문 운송·유통 업체인 Trifinity의 창고에서 발견되었습니다. (5312 104th Ave, Kenosha, WI 53144)
콜로라도스프링스 — 이후 트럭으로 서쪽으로 이동하기 시작했으며, 콜로라도스프링스 외곽 고속도로에서 목격되었습니다.
트럭 정류장 — 밤에는 트럭 휴게소에서 하룻밤을 묵었습니다. (2195 Hwy 6 and 50, Grand Junction, CO 81505)
Amazon is buying massive quantities of books, scanning them for AI training data, and destroying them in the process. A 404 Media investigation was able to reveal Amazon’s book buying operation, which hasn’t been previously reported, by placing a tracking device in a rare book we suspected would be acquired by an AI company for training data, and following it around the country to its final destination. That final destination was an Amazon warehouse in Las Vegas, Nevada. Amazon employees who work at this location say all they do is receive massive shipments of printed books which they then cut the bindings off in order to scan the books more quickly. The printed book is destroyed in the process. The logo of the Amazon team that works at this warehouse, called VGT3, is a dinosaur, brandishing its teeth and with a book in its hands. “Amazon purchases books through commercial channels to help develop and improve the products and services our customers use,” an Amazon spokesperson told me in a statement. The world’s AI companies are constantly looking for, and spending extreme resources to locate, more material to train their AI models. With books, that sometimes means destroying them in the process, something that large parts of the public have spoken up against, and which we can now confirm Amazon is doing. In July, I published a story about booksellers who reported a historical spike in sales starting in the past year. They suspected this spike in sales was due to AI companies acquiring any books they can in search of new training data. Printed books are valuable as training data because a lot of the text they contain is not readily available on the internet, which AI companies have already scraped. The data is also conveniently organized and, if the book was printed before 2022, is guaranteed to be free of AI-generated text, which can make any AI model that is trained on it worse via a recursive process called “model collapse.” 📖 Do you know work at a facility where you scan books? I would love to hear from you. Using a non-work device, you can message me securely on Signal at @emanuel.404. Otherwise, send me an email at emanuel@404media.co. These booksellers suspected AI companies were behind these large bulk purchases because of the high number of books they were buying, the seemingly random choice of books, and the fact that these buyers, unlike libraries and universities, did not seem price sensitive at all. But booksellers couldn’t say for certain who was behind the large purchases because the marketplaces where they sell their books keep the buyers anonymous. When an order comes in, a bookseller ships the sold books to a warehouse operated by the marketplaces, where books are sorted and then sent to the buyer. In July, one bookseller told me they received a very large order of around 1,000 books on Biblio, one of these marketplaces. The seller agreed to put an Apple AirTag provided by 404 Media in one of the books included in this order so we could see where the book was going. And by extension, which company, AI or otherwise, was behind this massive order. 404 Media granted the bookseller anonymity because they worried sharing this information would harm their business. Biblio did not respond to a request for comment. 1 California The shipment of books we tracked flew out of an airport in California. 2 Milwaukee Next time we see it is right outside Milwaukee Mitchell International Airport. 1901 E College Ave, South Milwaukee, WI 53172 3 Distribution center We then saw it in a warehouse belonging to a specialized shipping and distribution company Trifinity, right outside Kenosha Regional Airport, 30 miles south. 5312 104th Ave, Kenosha, WI 53144 4 Colorado Springs It then starts heading West via truck. We see it on a highway outside Colorado Springs. 5 Truck Stop It stops for the night at a trucking travel center. 2195 Hwy 6 and 50, Grand Junction, CO 81505. 6 Amazon warehouse: LAS8 The next day it arrives at an Amazon warehouse called LAS8. 5801 Nicco Way, Las Vegas, NV 89115. The bookseller sent the shipment to Biblio, and it arrived at a California airport. It then traveled by plane to Milwaukee International Airport in Wisconsin. Later that day, the book traveled to a warehouse belonging to a specialized shipping and distribution company called Trifinity, right outside of Kenosha Regional Airport, about 30 miles south. The book remained there for about two weeks, at which point it began traveling west by what appeared to be a truck. I could see the book travel via the highway and spend a night over at a trucking travel center around Grand Junction, Colorado. The next day, the book arrived at its final destination, an Amazon warehouse called LAS8 in Las Vegas. LAS8 is one of several large Amazon warehouses in the area, each with a different specialty. LAS7, right across the street, for example, is a fulfilment center, while LAS8 appears to mostly operate as one of Amazon’s “print on demand” operations, which will print and ship books to Amazon shoppers as they are buying them. Initially, I was confused about why the book would arrive there, but Amazon employees who work at this location and who discuss working conditions there with other Amazon employees online explain that the the north end of the LAS8 warehouse, where I saw the book arrived, housed a different Amazon operation with a different code: VGT3. Many Amazon warehouses have unique symbols to represent that specific site. VGT3’s symbol, painted on the entrance to the site and inside, is of a Tyrannosaurus rex, the massive carnivorous dinosaur, with its mouth open, holding an open book. “I work at VGT3 here in Vegas, and all we do is scan books,” one Amazon employee wrote on a forum for Amazon workers. “Some are assigned to cut books, and others go to receive where they get books and scan the bar codes. We didn't have rates, but now we do, but it's not stressful.” “Working at VGT3 is nice all we do is scan books,” another Amazon employee wrote. “It's so cool.” I saw VGT3 employees online talk about how working in this operation is a good but boring job because workers do the same, easy, repetitive tasks all day. I also saw some discussion indicating that VGT3 jobs are desirable for this reason, and because the warehouse sometimes offered night shifts. Earlier this year, some employees expressed concerns that Amazon would shut down the site because they had worked through the books they had and weren’t getting enough shipments of new books, though the site is still operational today. That employees are scanning the barcodes or ISBNs on books — a unique serial number given to every published book — before scanning their content gives further credence to another theory put forth by booksellers: AI companies are trying to methodically scan every printed book in the world by working through the list of ISBNs. One bookseller told me they suspected this was the case because the very large orders they were getting never included very rare books that do not have ISBNs. Elsewhere online in 2024, booksellers who sell their books on Amazon said they received a spike in orders to be shipped directly to the VGT3 location for a customer named “Amazon FC.” This was so unusual given its size they suspected the orders were some kind of scam . An Amazon representative chimed in to say they checked and that the orders were legitimate. Like other major tech companies, Amazon is developing its own large language models (LLMs). Amazon considers its family of models branded Nova to be “frontier” models, meaning it considers them to be competitive with other cutting edge LLMs from Google, OpenAI, and Anthropic. These models require massive amounts of training data in the form of human-written text. Internally, Amazon also used an AI coding agent called Kiro to develop its own software. We first learned that AI companies wanted to scan millions of books for training data because of a lawsuit from book aut