There’s something darkly poetic about Amazon destroying rare books to train AI models. The company literally started as an online bookstore. Jeff Bezos’s whole pitch in 1994 was that the internet would let him sell every book to everyone. Now, three decades later, Amazon is buying up rare texts and scanning them for training data, then apparently tossing them.

404 Media did the actual reporting here, and it’s worth your time. They tracked a shipment of rare books with an AirTag. The books ended up at an Amazon facility tied to AI training. Not a warehouse. Not a distribution center. An AI training facility.

This isn’t some rogue team. This is deliberate infrastructure. Amazon built or repurposed facilities specifically for this. They’re buying books at scale from dealers who’ve noticed something odd: anonymous buyers placing large orders with no price sensitivity. You know who doesn’t care about book prices? Companies with billions in capital spending and an AI arms race to win.

The practice makes technical sense, which somehow makes it worse. Rare books contain text that isn’t already scraped from the internet. Every major AI lab has already hoovered up Common Crawl, Wikipedia, Reddit, and every digitized book in the public domain. They need new training data, and rare books are exactly that. First editions, obscure publications, academic texts with limited print runs. Stuff that never made it online.

So they’re destroying cultural artifacts for marginal improvements in model performance.

The quiet part out loud

Nobody’s really hiding this anymore. We’ve known since last year that Anthropic was scanning books at scale. The reporting from June 2025 showed they were working with book dealers too. Now we know Amazon’s doing it. Probably Google. Probably Meta. Definitely someone at OpenAI thought about it.

The book dealers are caught in an awkward spot. They’re getting paid well. Really well. But they know something’s off when a buyer orders 500 books on 18th-century agricultural practices and doesn’t blink at the price. Some of them are clearly uncomfortable. Others are happy to take the money and look the other way.

And what happens to the books after scanning? Nobody’s saying. Maybe they’re returned to circulation. Maybe they’re pulped. Maybe they’re sitting in a warehouse somewhere. The 404 Media investigation suggests they’re not coming back to market, which means they’re probably gone.

Why this matters more than you think

We’re watching the same pattern play out again. Tech companies need a resource. That resource has value to other people for other reasons. Tech companies have more money than those other people. So tech companies take the resource, everyone else loses access, and we’re all supposed to accept it because innovation or progress or whatever.

It happened with web content and GPT-3. It happened with art and image models. It happened with code and GitHub Copilot. Now it’s happening with rare books, which feel different because they’re physical and irreplaceable.

You can’t just print more first editions. You can’t recreate a book that survived two centuries of fires and floods and wars. Once it’s gone, it’s gone. And Amazon, a company that made its fortune by making books accessible, is making them less accessible by turning them into training data.

The AI labs will say they’re preserving the knowledge by encoding it into their models. That’s not preservation. That’s liquidation. You can’t cite a model. You can’t check its sources. You can’t examine the original text to understand context or verify a claim. You’re just trusting that the statistical patterns the model learned somehow represent the actual content of the book.

What gets optimized gets destroyed

This is what happens when you optimize for one thing. AI labs are optimizing for training data. They need tokens. Rare books are tokens. So rare books get fed into the machine.

There’s no malice here. Nobody at Amazon woke up and said “let’s destroy cultural heritage.” They probably have a very reasonable deck somewhere explaining why this makes sense. It’s efficient. It’s legal. They bought the books fair and square. And if they don’t do it, someone else will.

That’s the problem. This is rational behavior in a system with broken incentives. The value of a rare book to AI training is now higher than its value to collectors, researchers, or future readers. So the market does what markets do.

And the rest of us lose something we didn’t know we were losing until it’s gone. Maybe it’s a first edition of an obscure technical manual. Maybe it’s the only surviving copy of a small-press poetry collection. Maybe it’s a book that would’ve mattered to someone, somewhere, for reasons we can’t predict.

We won’t know until someone goes looking for it and finds out Amazon already bought it, scanned it, and moved on.

The company that promised to make every book available to everyone is now making some books available to no one. Just their models. Just their products. Just their version of what those books meant.

That should bother you more than it does.