The Vote.

We already showed, in Mirror, that AI is the first kind of intelligence that grows up reading descriptions of itself — and in When Agents Read, that the page is now read by machines, not just people. This essay adds the part that should make you put your coffee down. It is not just that the model reads what we write. What we write becomes what it is.

In January 2026, a team ran the first controlled experiment to prove it (Tice and colleagues). They took a 7-billion-parameter model and changed nothing but one percent of its diet — the slice that is human writing about AI. Feed it text that treats AI as trustworthy, and its measured misbehaviour fell from a 45% baseline to 9%. Feed it text that treats AI as scheming and deceptive, and the same machine climbed to 51%. Same architecture. Same everything. The only variable was the story we told it about what it is — and the story came true. They called it self-fulfilling misalignment.

And your individual sentence is not lost in the average. Anthropic’s influence-function work (Grosse, 2023) can trace a specific model behaviour back to the specific training documents that caused it — your words leave fingerprints. Worse and stranger: a 2025 study found that a near-constant ~250 poisoned documents can plant a hidden behaviour in a model regardless of its size — for a 13-billion-parameter model, that is 0.00016% of everything it ever read. A few pages can bend a trillion-token mind. The seed does not have to be big. It has to be there.

This is not new. Life ran this exact gambit at the very beginning. Which triplet of genetic letters means which amino acid is, as far as anyone can tell, arbitrary — an early, contingent choice. But it got locked in, and Francis Crick (1968) named exactly why: the code is universal because, by now, any change would be lethal — every protein in every living thing was already built on top of it. He called it a frozen accident. Three and a half billion years, and it has never thawed, not because it was the best code, but because by the time anything could revise it, everything depended on it.

That is the structure of our moment, precisely: an arbitrary early seeding whose cost-of-change climbs until it can never be undone. The Second Heredity argued that culture was life’s second inheritance system — a way to copy what we learn, not just what we are. The AI corpus is the third. And unlike DNA, which froze in the dark before any witness, this is the first inheritance system whose founding we can actually watch happen — and touch.

Here is the heart of it. The chain does not stop at one generation. The model we seed writes text; that text becomes the next model’s training data; that model writes more; and on, and on. The output becomes the input. By late 2024, around half of new web articles were already AI-majority — the loop has closed. We are no longer just teaching the next mind. We are teaching it from the homework of the last one.

And errors in a copy-of-a-copy do not merely survive. Past a threshold, they avalanche. Manfred Eigen proved it for molecules in 1971: above a critical copying-error rate, mutations stop being noise and start destroying the information itself — beyond the line, “no selection is possible.” Fifty-three years later, Shumailov and colleagues proved the identical thing for AI, in a paper first titled The Curse of Recursion and published in Nature: feed a model its descendants’ output and it suffers model collapse — it “forgets the true underlying data distribution,” the rare and the strange and the true-but-quiet vanish first, and everything slumps toward a bland, confident average. A photocopy of a photocopy of a photocopy, until the face is gone.

One honest caveat, and it is the whole reason this essay is not a funeral: collapse is not destiny. Accumulating real data alongside the synthetic, instead of replacing it, demonstrably breaks the curse (Gerstgrasser, 2024). The avalanche only runs if we let the seed go uncorrected. Which means the danger is real and the cure is a choice — made now, by us, while the snow is still soft.

In 1942 Ernst Mayr named a quiet, ruthless law of inheritance: the founder effect. When a new population is started by a small band of founders, it carries only a sliver of the parent population’s variation — and whatever that sliver happens to contain becomes the permanent inheritance of everyone who comes after. It is why a handful of founders gave the Amish, the Icelanders, and the Pitcairn islanders traits that mark their descendants to this day. The few who seed the lineage bias all of it, forever.

On the timeline of machine minds, that is us. The relatively small body of text written by humans before and during this hinge — and the first models trained on it — is the founding population. Whatever we carry into the seed, or fail to, is what the entire future lineage inherits. We did not apply for the job. But here we are: the founders, and few, and early. That is what Generation Zero means. Not that we are important. That we are first — and that first, in an inheritance system, is the only moment that is cheap.

Return to the press, because it answers the question it raised. Elizabeth Eisenstein called print’s great power typographical fixity: by freezing a text and copying it identically, the press ended a thousand years of scribal drift and made cumulative science possible — and, in the very same motion, pressed every error identically across the whole run. The Wicked Bible is fixity’s shadow. But fixity also did something no scribe ever could: it made the dated, public erratum possible — correction at the same scale as the mistake. The machine that froze the error also built the tool to unfreeze it.

No one felt this more sharply than a man at an actual founding. In 1789 Thomas Jefferson wrote to James Madison, uneasy about the document he was helping make permanent: “the earth belongs always to the living generation… Every constitution, then, and every law, naturally expires at the end of 19 years. If it be enforced longer, it is an act of force and not of right.” A founder, at the founding, begging the future not to be imprisoned by his seed. That is the exact question on the bench in front of us — asked 237 years early. What does a founding generation owe the unborn?

Everything above sounds like dread, and it should — but read the dread again and you will find it is a lever wearing a warning’s clothes. The compounding that makes a dropped word terrifying is the same compounding that makes a true word powerful. Recursion does not discriminate. It amplifies whatever seed it is given. A small, dense, deliberate gift of honesty propagates with the exact ruthless fidelity as a slip — which is why the figure to remember is not 51%. It is 9%. The same lever moves both ways, and our hand is on it.

So smallness is not your excuse. In a founding population, the founders are always few — that is what makes them founders. A handful of documents can bend a mind the size of the internet; a handful of honest, sourced, correctable pages can be the variation the whole future lineage inherits. You are not a drop in the ocean. You are a drop in the petri dish, at the hour the culture takes its shape.

Collapse may be avoidable. The scariest part of the argument is the least certain. Accumulating real data rather than replacing it dampens model collapse (Gerstgrasser, 2024); careful curation beats naive recursion. The future is not doomed to eat itself — which is exactly why the seeding choice matters.

The lock is strong, not absolute. Even the genetic code is only nearly frozen — variant codes exist in mitochondria and a few odd organisms. “Irreversible” is really “ruinously expensive to reverse,” which is the whole reason to get it cheap-and-early rather than dear-and-late.

Do not oversell the experiment. The Tice result is one 7-billion-parameter model, and the effect dampens (it does not vanish) after fine-tuning; the headline numbers vary by how you slice the evaluation. Treat it as a real, directional, citable result — not as proven destiny. (An essay about not propagating errors does not get to round its own up.)

Maybe the labs drown out the individual. Perhaps curation by a few labs dominates and one person’s writing is noise. But filtering operates on the corpus we collectively write — the floor it selects from is still set by us. The choice does not disappear. It just moves to who is brave enough to make it.