JUANA SUMMERS, HOST:
Tech companies have been going to great lengths to find ways to train their AI models, including purchasing, scanning and then destroying rare books. An investigation by 404 Media traced one purchase of rare books with a tracking device and found that the shipment ended up at an Amazon AI training warehouse. Let's bring in Emanuel Maiberg. He's the reporter behind this story. Welcome to the program.
EMANUEL MAIBERG: Hey. Thanks for having me.
SUMMERS: Thanks for being here. I mean, Emanuel, this is an incredible story. I wonder if you could just start by walking us through how it came together? Where did the idea to put a tracking device into a rare book come from?
MAIBERG: Yeah, so we've known that AI companies have been doing this since a lawsuit from book authors against Anthropic, which is one of the major AI labs. And that lawsuit revealed that they were purchasing books in large quantities, that they were scanning them, and that in order to scan them really fast, they cut the spines off the books and fed the pages to the scanner. I was interested in that. At some point, a bookseller who's been in the business for a long time reached out to me and said that they have been getting very large orders. I obviously wanted to know where these books ended up. So we managed to get a tracking device into one of these orders, which is how we discovered the Amazon warehouse.
SUMMERS: So interesting. And how were they able to tell that these books were being purchased for AI training?
MAIBERG: There are a few signs. One of them is, historically, the kind of people who buy these books don't make such large orders. Another sign is that there was no coherent theme to the kind of books that the clients were purchasing. So, for example, if there's a library at a law school, they might purchase a few books about a specific legal issue. That makes sense. You can tell that there's someone who is interested in a specific subject. When you get an order for hundreds of books that have no coherent theme, that is another sign.
SUMMERS: And if I understand correctly, these companies are going after old books. Why is that?
MAIBERG: Well, there is this problem with training AI models, which is up until now, the AI models are mostly trained on all the information that they scraped off the entire internet. At the point at which they did that, that content was human-generated. Articles from newspapers, from blogs, scientific articles, just people chatting online. And what AI companies discovered is that - once they scraped all that data and they were looking for more data to make their AI models even better, they tried generating data with AI and then having the AI train on that text.
It turns out that that causes a problem called model collapse, where if an AI trains an AI training data, it gets worse. So now they're looking for more human-generated text because they already scraped the internet. The next logical step is to scan books that are not easily available online. This is why, at least the books that I've seen that end up in these orders, are not popular, common, easy to find books. They're more rare books.
SUMMERS: Interesting. What do you hear from critics of this practice? Like, what do they say the issue is here?
MAIBERG: So I think people obviously have a very negative reaction to anyone destroying books in large quantities. I think people hear the term rare book, and they're imagining a first edition of "Oliver Twist."
SUMMERS: Right. Yeah.
MAIBERG: Or something like this. That is not the case. A rare book is most likely a book that was never printed in large quantities because it is about a niche subject. It is what is called vanity publishing, where someone self-publishes a book. It's an instruction manual for a product that's no longer available. It's really things like this that people at the moment don't really care about. They don't have high emotional value. But we don't necessarily know what we might find valuable in the future.
SUMMERS: Emanuel, how did Amazon respond to your reporting?
MAIBERG: Amazon confirmed that the warehouse that I reported on, VGT3, was actually used for AI training data. Their specific quote was, "Amazon purchases books through commercial channels to help develop and improve the products and services our customers use." That is a vague statement about why they're doing this exactly. I believe that the reason is that it's not just for AI training data and that originally the facility wasn't established for that reason. But it is one reason that they're scanning the books now.
SUMMERS: That was Emanuel Maiberg, reporter with 404 Media. Thanks so much.
MAIBERG: Thank you.
(SOUNDBITE OF MENAHAN STREET BAND'S "TIRED OF FIGHTING") Transcript provided by NPR, Copyright NPR.
NPR transcripts are created on a rush deadline by an NPR contractor. This text may not be in its final form and may be updated or revised in the future. Accuracy and availability may vary. The authoritative record of NPR’s programming is the audio record.
