Anthropic bought millions of used books.

Then it cut the bindings off, scanned every page and threw the paper away. Mulched it.

Not metaphorically. Not "ingested" in some vague AI press-release sense. Industrial cutters. High-speed scanners. Pallets of books.

The stinger here?

This destruction helped make the process legal.

Fahrenheit 451. With lawyers.

Now … this is a tricky one. I LOVE books. And find it appalling that books were destroyed.

However (and lord it pains me to say this): there were reasons for this. And it’s not quite as bad as the viral posts are making out.

Still…bad.

Here’s my Youtube video on the story:

Project Panama

Anthropic called the operation Project Panama.

It wanted "all the books in the world" for a central research library. So it bought millions of physical books, hired outside companies to remove the bindings, scanned the pages and kept the digital copies.

The mangled originals were discarded.

The scans went into a private library. Anthropic then selected material from that library for training Claude models.

That last distinction is important btw. "Anthropic scanned a book" does not automatically mean every Claude model was trained on that specific book.

Someone asked this on Youtube. And it’s a good question. They asked if they now had access to all these books inside Claude:

The answer is nope. The library and the model-training sets were separate things. You can’t just “call up” any book from a vast library that Anthropic now owns. Instead the models are trained on parts of the library. Very different.

Why destroy the originals?

OK so even if we are ok with training on books (which is a whole other complicated story…) why destroy them?

Is it just because it’s cheaper? Well yes. It is cheaper to use destructive scanning because it’s much faster. But that’s not the reason…

The reason is legal.

If Anthropic bought one physical copy and created one digital copy while keeping both, it had made a surplus copy.

That’s a problem. They’d fall foul of fair use laws and get sued into oblivion.

Instead it used what the court called a one-in, one-out format change:

One purchased paper copy became one private digital copy. The paper version was destroyed. The scan was not redistributed.

In the June 2025 fair-use ruling, Judge William Alsup compared this to replacing purchased print copies with more convenient digital ones. His wonderfully blunt summary was: "There was no surplus copying."

That is what made this a better path for Anthropic.

So yes…

The destruction was not some accidental bit of waste. It was part of the legal logic. A legal loophole.

The same ruling also found that using books to train specific language models was "quintessentially transformative". Anthropic was not selling replacement copies of the books. It was using the text to build a model capable of generating new material.

Whether you agree with that or not is a different question. But that is what the ruling says.

The $1.5 billion was for something else

This is where most posts about the story collapse two separate things into one. I’ve seen a lot of confusion about this online.

Anthropic have just settled a $1.5b lawsuit over using books in training. They are paying out to publishers and authors who are aggrieved that their work was used.

That’s cheap for Anthropic btw…they’ve done very well out of the settlement. A story for another time!

This settlement is a different matter entirely. Anthropic downloaded millions of books from pirate libraries including LibGen, PiLiMi (a mirror of Anna’s Archive). It kept those pirated files in its central library.

That was NOT excused as fair use.

And that is what the $1.5 billion settlement was about. Not the paper cutter stuff but good ol’ fashioned Piracy.

The court gave final approval on 20 July. More than 482,000 works are covered, with an estimated payment of roughly $3,000 per work.

The distinction is bizarre but super important:

Buy a used book, destroy it and keep one private scan? Fair use.

Download the same book from a pirate site and keep it? Copyright infringement claim.

You can see why people are confused.

AI ruined the internet.

Why go to all this trouble in the first place?

Because books are good training data.

They are long. Edited. Dense. Specialist. Structured.

But most importantly anything written before 2022 is golddust now.

A lot of them were written before the internet filled up with AI-generated slop. If you train off AI work you poison your model.

We have talked before about why models hallucinate when the underlying signal is thin. Better source material does not magically make a model truthful. But a library of carefully edited human writing is vastly more useful than another billion scraped AI generated listicles repeating each other.

Or, as one bookseller told 404 Media: the world's best AI training data is sitting on a shelf.

And now AI companies appear to be buying the shelves…

Book dealers in Europe are reporting strange bulk orders for thousands of unrelated titles. One request apparently asked for 3,000 books with little concern for subject or condition. Rare-book dealers are worried that obscure editions are disappearing into scanners.

BUT…

We do not know who every buyer is. We do not know what every book is being used for. And there is no evidence Anthropic ran around hunting down the last surviving copy of priceless first editions for a giant book bonfire.

Most of the books going in are books we would otherwise think of as junk.

The books that have been sitting high up on a shelf in your local second hand bookshop for a decade.

Dummies Guide to Microsoft Access ‘95, Lonely Planet Czechoslavakia, Boise Idaho Regional Almanac 1987.

Books that literally sell by the pound in wholesale markets.

There is a LOT of junk out there that would have (let’s be honest here) gone to the pulper eventually. In a weird way Anthropic (and other AI labs) have saved some books that were essentially valueless. These items are still useful for AI models, even if they haven’t been valued by us humans for a few decades.

A lot of the viral posts show images of antiquarian books being sliced and diced. The main video you’ll see (of the covers been cut of hardbacks) is actually from a book restoration art project… quite the opposite of what is being pushed.

This is why we have to be a bit careful here.

Destroying books is gross. It’s icky. I get that.

But we as humans recycle a LOT of books every year anyway. This does not excuse Anthropic but as always it’s important to put things in perspective, even if it’s uncomfortable.

The private library

For me personally the shitty part is the private library.

Anthropic (and I imagine other labs) are building these private labs. And removing physical books from circulation. The digital replacement stays inside a private company. The knowledge improves a paid product, but the scan is not returned to a public library or archive.

I’ve seen some people argue that Anthropic should make their scans available.

Come on. Think!

To be clear: Anthropic cannot simply publish copyrighted scans for everyone.

They would get sued into oblivion for piracy. They legally cannot share without yet again falling prey to fair use laws.

Rights still exist. Authors and publishers still need paying.

But "we bought it, chopped it up and now the only useful copy we made lives behind our lock" is not exactly public preservation. This also makes this week’s fight over open weights and Anthropic's closed-model stance feel a bit more pointed. The valuable inputs are gathered into a private library. The resulting model stays private. The public gets access on Anthropic's terms…

Good business. Icky precedent.

At minimum there should be proper provenance records, protection for genuinely scarce copies, compensation for rights-holders and some route for public-domain scans to end up in public archives rather than disappearing into corporate training pipelines.

Remember though this is not Farenheit 451. Anyone saying so hasn’t read the book.

Anthropic did not destroy the books because the knowledge was dangerous.

It destroyed them because the knowledge was valuable.

Kyle