A Most Comfortable Dead End — Monokai

A Most Comfortable Dead End

No AI is flattening us. We're happily planting our own monoculture.

July 30, 2026 7 min. Artificial IntelligenceModel CollapseCulture flattening

Perfection and mistakes, AI generated

Right now, without asking you, your body is making mistakes on purpose. In a lymph node somewhere, a white blood cell is copying the gene for an antibody and getting it wrong, introducing errors at roughly a million times the normal rate, producing thousands of garbled variants, nearly all of them useless. Nothing in there is deciding anything. The cell cannot know which virus you will meet next month, so it does not aim. It scatters strangeness in bulk, and whatever happens to fit gets kept. This is not a flaw in the immune system. It is how the immune system works. A body that produced only the one correct, average antibody would be defenseless against anything it had not already seen, which is to say against the future.

Error is not the opposite of intelligence. In any system that has to survive a world it cannot predict, error is where the intelligence comes from.

We are now building, at staggering expense, a machine that does the reverse.

A large language model is an averaging engine. It can recombine its material in ways that genuinely surprise you, but its center of gravity is the middle: the likeliest next word, the sentence most people would have written, the median thought in clean prose. That is the design, not a defect, and it is seductive because the middle works. It answers the email, clears the filter, passes for competent. Sounding strange is something you pay for in being misunderstood.

The usual story says this is being done to us. But nobody is taking anything. We hand it over with a small exhale of relief, the way you tap the suggested reply instead of finding your own words. The machine did not steal our strangeness. It offered to spare us the cost of having any, and the cost is real, and the offer is free.

A million mistaken hands

For most of history people made things by copying: songs set to borrowed tunes, icons painted to a pattern. And the copying was never clean. That is the part that matters. A scribe’s hand slipped. A singer misremembered a verse and the mistake was better than the line. Consider the pea on your plate. For centuries the vegetable was “pease,” a mass noun like rice, with no singular at all. People heard the final sound as a plural, sliced it off, and invented a lone “pea” that had never existed. Your language is full of these fossilized errors. You used several this morning.

Variation came from everybody, thrown off by people reaching for something and missing by a little — ten thousand wrong antibodies to find the one that works. The people we call original did not invent novelty. They foraged it from drift that nobody meant to make.

So the loss does not fall on the few first. It falls on the ground they stood on. What the machine removes is the scatter: it copies without error, and asked for the average sentence it gives you the average sentence, immaculate, none of the grit a tired hand leaves behind. You cannot mutate your way out of a monoculture when there is nothing left to mutate from.

The collapse that counts

We have heard all this before, and it was wrong every time. The printing press was going to flatten the scribal hand; photography would kill painting; recorded sound, the photocopier, the sampler, the web. Each was going to drown us in sameness, and each instead split the culture open and handed tools to people who had never had any.

Why is this different? Because those were machines for reproduction, and this is something stranger: a machine that generates, always toward the middle, and then feeds on what it makes. The press spread what humans wrote, errors and oddities and all. Many presses, many hands, every one of them wrong in its own way. It never read its own output, regressed toward its own average, and trained on the result. This one does. Each pass writes the likeliest version back into the pool it learned from, and the pool gets flatter. Researchers call the far end of this model collapseShumailov, Ilia et al. “AI models collapse when trained on recursively generated data.” Nature vol. 631,8022 (2024): 755-759..

Collapse in the lab is conditional and probably fixable. It sets in when a system trains mostly on synthetic data with no fresh human input, and the people building these things know it: they curate, mix in real text, strain out the slop. A training set can be groomed. So grant them the win, and suppose the output stays as varied as the writing it learned from.

It still does not help, for two reasons that have nothing to do with the weights.

The first is that the slip requires the attempt. “Pease” became “pea” because somebody had to produce the word with their own mouth and got it wrong. Drift is a byproduct of reaching. A sentence that arrives finished was never...