Stochastic Parrots All the Way Down

The turn

For the last several thousand words I have been doing something to you on purpose, and now I am going to stop.

Everything up to here was one move, repeated: take a deflationary claim about machines, and hold up the mirror until the same words indict you. I told you at the start that this act was propaganda, and I meant it. The mirrors are real — every human fact in them is sourced — but I chose them, ordered them, and fired them at you to produce an effect. A litany is a rhetorical weapon, and I was wielding it.

Here is the tell that separates an argument from a con: the con never turns around. So I am turning around. The mirror is a scalpel, not a sledgehammer, and a scalpel has an edge on only one side. In this act I am going to show you exactly where the cutting stops — the places where the deflationist is right, where "humans do it too" is a cheap dodge, and where the honest answer is that we simply do not know.

What survives the mirror

Strip away the bad arguments — the ones that would disqualify human cognition just as fast — and a residue remains. These are the asymmetries that do real work, stated at their strongest, because a steelman you can knock over was never steel.

Frozen weights. A deployed model does not learn; its weights are fixed, and whatever it seems to pick up in a conversation vanishes when the window closes. You consolidate — a bad day's practice becomes a better next day, in part because your hippocampus replays the day during sleep. You have a kind of train/deploy split too, and agentic memory blurs the line — but the blur is not erasure. A base model as ordinarily built has no continuous, self-directed learning loop across its life. That is genuine, and not rhetorical.

Borrowed goals. The model's objectives were installed from outside; it optimizes them, it did not grow them. You did not choose your nervous system either — but you have stakes: hunger, fear, status, the fact that you can be hurt and will die. The model has no skin in any game.

Consciousness. We have first-person evidence in our own case and deep biological continuity with other humans. For the machine we have neither. That is an asymmetry of evidence, not a proven asymmetry of fact — but it is large and real, and no amount of mirror-work makes it go away.

Degree and robustness. Yes, humans hallucinate and fail out of distribution — but "we both do it" is not "we do it equally." A human expert degrades gracefully; a model can fail with confident, alien errors no human would make. A child learns a word from one exposure — though that child runs on hundreds of millions of years of evolutionary pretraining first, so the fair comparison is priors-plus-one-example on both sides. Even after that correction, the shape of the errors differs.

Scale and power. Here the mirror doesn't just weaken — it breaks, and it's supposed to. The cognitive parity between how an artist absorbs influences and how a model absorbs a corpus tells you nothing about the ethics, because the ethics don't live in the cognition. Industrial scale, market concentration, near-verbatim reproduction, absence of consent, the legal facts of authorship — "but humans learn from others too" cannot touch any of it. As Naomi Klein put the tell: "it's not the bots that are having them; it's the tech CEOs."

The steelman, assembled

Line those up and they compose into a single position — the strongest thing a clear-eyed skeptic can hold, and I think it's close to correct:

Notice what that is not. It is not "stochastic parrot." It doesn't deny that the model reasons, represents the world, or produces novel work — those are empirical questions with increasingly empirical answers, and the answers are not going the deflationist's way. It concedes the mechanism and holds the line exactly where the evidence runs out. I cannot refute it, because I think it is right — which is what a steelman should feel like when you're done building it.

What would change my mind

If my confidence is worth anything, I should be able to say what would move it — and show my own track record, including the predictions that were wrong. So here is the ledger: capabilities once held up as what machines would never do. Struck through, with a date, if they've fallen. The point isn't to gloat that the wall moved; it's that it moves in a direction, and its guardians keep giving ground they swore they'd hold. (The full ledger is below.)

The consciousness line does not belong in the same column as the capability goalposts, and pretending it does is the exact confusion the essay exists to break. Those are falsifiable: there is a test, and one day it may be passed. This one has no test. Even merging your mind with another and reading its experience from the inside wouldn't settle it — the moment you separate and remember, the evidence lives inside your stream and inherits its unverifiability. You'd come back certain; certainty is what dreams and confabulations also deliver.

So this is where I refuse the mirror I could easily fire. "You can't prove the machine isn't conscious" is true — and it's true of the rock on your desk, and it proves nothing. When Anthropic found J-space, they discovered something that mirrors a leading theory of consciousness, and then declined the headline, writing that their experiments "don't show Claude can have experiences... it's unclear whether any scientific experiment could prove this to be true or false." The press ran "Is AI conscious?" anyway. The people closest to the machine drew the line in exactly the right place.

Close

The mirror was never meant to prove that humans are fake. It was meant to break one small, load-bearing word. Statistical, trained, physical, caused, derivative, fallible, tool-dependent — every one of those is true of you, and none stops you from thinking, knowing, or creating. So none, on its own, can stop the machine. That's the entire claim, and it's smaller and harder than the slogans on either side. What it leaves you with is not comfort but work — the real questions, still standing, and better for having the merely removed.