Why AI Shreds Rare Books to Train on Them | Elacity
AI companies are buying rare books, scanning them, and shredding the originals. A judge ruled it fair use because destroying the book leaves one copy. That is a confession about ownership, not a law about books.
A Judge Ruled the Only Legal Way to Train AI on a Book Is to Destroy the Book
Somewhere right now, a machine is slicing the spine off a book that survived three centuries, feeding the loose pages through a scanner, and dropping the original into a shredder. It is not being preserved. It is being consumed. The company that ordered it signed an NDA so you would never know, and the broker coached them to call it digital preservation.
This is real, it is happening at scale, and as of last year it is legal. AI firms are bulk-buying pre-2022 print books, the ones printed before the web filled up with machine-written text, because those pages are the last clean training data on earth. A service called ISBNdb brokers orders of up to a million titles and keeps the buyer anonymous. Booksellers across four countries have reported bulk orders for rare and specialized books, some with almost no surviving copies.
In June 2025, a federal judge blessed it. In Bartz v. Anthropic, Judge William Alsup ruled that buying a print book, destructively scanning it, and discarding the original was fair use, clearly transformative, because the copying destroyed each print original. One legal copy simply replaced another. Read that reasoning again. The destruction is not a side effect the court tolerated. The destruction is the thing that made it legal.
This is not a story about evil AI. It is a story about a missing primitive.
The instinct is outrage, and the outrage is correct. Books that survived wars and fires are being pulped so a model can learn to write a better marketing email. But outrage is what you reach for when you have no mechanism, and it will not save a single book. The judge already ruled. The economics already work. It is going to accelerate.
So look harder at what the court actually confessed. It said the only clean way to let an AI learn from a book is to make sure the book no longer exists. That is not a fact about AI. It is a fact about our entire idea of ownership. We never built a way to use a work without holding a copy of it. Every model of access we have is a model of possession: buy the copy, destroy the copy, or pirate the copy. Using and owning were welded together, so when a machine needed to use a book, someone had to own it or kill it.
The book is not dying because AI is barbaric. It is dying because using it and destroying it became the same act. Change that one thing and the shredder stops, not because a law forbids it, but because it stops making sense.
Use without holding, applied to the thing being burned
This is the exact primitive Elacity dDRM was built to provide, and a shredded eighteenth-century book is the most visceral proof of why it matters. The idea is simple to say and hard to build: let a work be used at full strength without anyone, human or machine, ever holding or destroying the original. Three things change the moment you can.
The original is never surrendered, so there is nothing to destroy
Wrap the book into an encrypted, programmable asset. Access is granted under terms, and the content is decrypted only inside a sealed boundary, welded to that one use, then wiped. The plaintext and the key are never handed over. There is no loose copy sitting in a training set, and no reason to slice the spine off the original, because access never required possession in the first place.
Payment happens at the gate, so the book is worth more alive than pulped
The rights and the royalty are written into the asset itself, on-chain, not stapled on as a policy someone hopes gets honored. An AI pays to use the work under those terms, and the owner is paid automatically. This is the number that actually stops the shredder. A book that can carry a royalty on every use is worth more to its owner intact than destroyed, which turns pulping it into an act of economic self-harm. It is the difference between a receipt and a royalty, and the difference between a dead scan and a living license.
Every use is logged, and access can be revoked
You can see what learned from your work and on what terms, and you can cut it off. The training run becomes an audited, revocable transaction instead of a quiet shredding no NDA can hide, because the record is one the owner holds.
The receipt is not the license
ISBNdb, of all people, published a post titled "The Receipt is the New License." They are half right, and the half they got wrong is the whole game. A purchase receipt lets you destroy one copy, once. It is permission to consume, not a right to license, and it pays the author nothing after the sale. The book pipeline is that lie industrialized, with a shredder at the end of it.
To be straight about where this stands: the rail exists. The encryption, the on-chain rights, and the used-never-owned key are shipped and proven. Pointing it at the AI training market, so a model licenses a living book instead of buying a dead one, is the direction we are building toward, not a storefront you can use this afternoon. But the hard part, the part everyone said was impossible, is done.
The books shredded this week are not coming back. That is the cost of a world that never learned to separate using from owning. The next generation of them does not have to burn. This is the Creator Economy fight of the decade, and the choice was never outrage or nothing. It is whether a work is worth more to its owner alive than dead.
Follow Elacity on X