AI companies are quietly buying and shredding millions of pre-2022 books, hunting for the last human thought uncontaminated by machine output, because their own models rot when fed their own writing. The same collapse is now creeping into human minds, and the only defence is refusing to outsource the one thing the machines have never managed: thinking of something genuinely new.
The Machines Are Buying Books Made of Human Thought
Somewhere in Europe, a second-hand book dealer takes a bulk order for several thousand niche titles. History, botany, regional law, German-language economics. The buyer does not haggle, does not ask about the condition of the covers, and never wants the books sent to a reader. Instead, the pages are sliced from their spines, fed through a high-speed scanner, and the originals are pulped. What survives is the text, converted into training data and poured into a language model.
The fulcrum here is the date. The books being hunted were printed before 2022, and they are prized for one reason: they were written entirely by human beings. ISBNdb, which runs one of the largest book databases in the world, now brokers these purchases for AI labs in quantities from a thousand copies to a million at a time. Its own pitch is unusually candid. Print from before the large language model era is, in its words, structurally clean. Almost everything published since is treated as suspect.
The reason is a phenomenon researchers call model collapse, and once you understand it, we can start to see a biological rhyme.
Go Deeper: Burning the Library Again
What Happens When a Mind Eats Its Own Output
A language model learns by reading enormous quantities of text and absorbing the patterns inside it. For years, that text was human: books, letters, arguments, the ordinary strangeness of millions of people writing in their own voices. Then, from late 2022, the machines began producing text of their own, and that output flooded the same internet the next generation of models would learn from.
Feed a model its own writing, then train the next model on that, and quality degrades. The rare words go first, then the unusual phrasings, then the odd and specific ideas that lived at the edges of the data. Each generation drifts closer to the bland centre, more repetitive, more confident, and more wrong. Researchers have shown this in controlled studies: recursive training on synthetic data pushes models towards a narrowing sameness and rising error. It’s probably why a lot of AI-generated content starts sounding alike, with similar word patterns, and the more we read and hear, the easier it becomes to quickly spot the tells.
The Collapse Is Not Confined to the Machines
There is a disturbing ring to this. A system that loses its rarest and most precise material first, that grows more repetitive and more certain as it declines, that mistakes the average for the truth, or simply hallucinates, sounds eerily like dementia or Alzheimers. The comparison to a degenerative neurological condition sounds harsh. It is also accurate: the machine forgets the edges of what it knew and cannot generate anything genuinely new to fill the gap.
Since 2022, a great many people have folded AI into the centre of how they work, research, decide and create. Used well, it is a genuinely useful tool. The real impact arrives when the tool stops being something you use to sharpen your thinking and becomes the thing that thinks in place of you.
When you outsource the first draft, the difficult paragraph, the problem you could not immediately crack, you save time. You also skip the exact mental effort that would have built your capacity to do it next time. Do that often enough and the faculty you stopped using begins to waste. The rare idea, the strange connection, the argument nobody else would have made: these are the human equivalent of those edges of the data, and they are the first thing to disappear when you stop generating your own.
So a second collapse runs in parallel with the first. The models rot from training on their own averaged output. Human ingenuity rots from the same cause, just as the machine minds feed on the smoothed, confident, pre-averaged output of themselves and learn to produce more of the same homogeneous slop. When you photocopy a photocopy over and over, you end up with a grey, barely readable (if at all) facsimile of the original. When a civilisation starts to do the same under the illusion of convenience, the future does not look good.
Go Deeper: Why Creativity is Humanity’s Best Defence
The Enemy Was Never the Tool
There is one thing that has always, and will always, separate us from the machine: the ability to produce something that was not already in the training data, ie original thought. The models cannot do it and never have. That has been the quiet human advantage all along, and it is exactly the advantage given away each time convenience wins over creativity.
The question is not really about the technology at all. It is about who does the thinking. The pull towards convenience is constant, and it wears the costume of efficiency and the prize of saved time. Every time you allow it to make the choice for you, you may get a little faster but also a little less capable of the thing that made you worth listening to.
Responsible use of AI is not a moral hedge. This is the discipline of keeping the tool in its place, drafting, sorting, checking and accelerating, while the origination, the judgement and the cognitive leaps stay with you. Read what it offers, then think past it. Let it clear the ground so you can build something it could never have reached on its own.
This is what taking back mental sovereignty actually means. It does not mean switching the machines off. It means refusing to become subservient to them. The pre-2022 books are hunted because they hold something pure, uncontaminated, irreplaceable: human thought. So does the mind you were born with. The only question is whether you keep generating from it, or quietly let it be trained into sameness like everything else.
Join the Conversation
Where have you felt the pull to hand your thinking over to a machine, and what did it quietly cost you? What would it take to keep your rarest ideas your own? Share your experiences and insights below.
