The Test Nobody Passes

New work from Proof of Human makes this uncomfortably explicit. Researchers built CogCAPTCHA30 — a battery of thirty tasks, twenty-nine of them lifted from cognitive psychology, one of them the familiar grid of blurry crosswalks — and ran humans and machines through all of it. Frontier models were included: Claude, GPT, Gemini. So were two small ones, a 1.5-billion-parameter Qwen and a 70-billion-parameter model called Centaur.

The frontier models got the answers right. On image classification they matched human accuracy. What gave them away was the how — click order, direction changes, a tendency to overselect. Not what was chosen. The shape of the choosing.

Read that again slowly, because it inverts the thing you assume you're doing. The test does not detect intelligence, and it stopped detecting competence some time ago. What it detects is your error signature. You pass by being wrong in a familiar way, at a familiar speed, in a familiar order. Your humanity, as measured by the internet's most-administered examination, is your particular manner of fumbling.

The second finding is funnier and worse. The most humanlike performer in the study was not the largest or the most capable model. It was Centaur — fine-tuned on more than ten million human choices. It resembles us because it is made of us, which everyone keeps reporting as a surprise rather than as a definition. Hold onto that, because it cuts harder than it looks: a thing assembled out of our behavior will produce our behavior, and the resemblance tells you nothing whatsoever about what is or isn't happening inside it. Hold onto the other half too. Those ten million choices were made by people. They are now a product.

And the robustness result names the actual mechanism. Hand a model the discriminator's feature list and it closes the gap; it can perform the fumbling once it knows how fumbling is scored. Withhold the list, or ask it to generalize to an unseen task, and detection holds.

Say that plainly, because it's the study's real headline and it cuts against the easy version of this essay: the test works. It just doesn't work the way you assumed. It works on adversaries who don't know what's being measured. You have never known what was being measured either. You just kept clicking and kept being let in.

Which raises the question the whole framing declines to ask: who administers the internet's most-administered examination of personhood, and who gets billed for it? It is a corporate instrument. It gates the ordinary commons of a life — mail, tickets, forms, tax portals, concert seats — and the toll is cognitive work you perform for free. That work was never decorative. reCAPTCHA was designed from the start to harvest it: first transcribing books nobody could OCR, then street numbers, then labeling the images that trained machine vision. So the loop closes on itself. Humans do unpaid labeling to prove they are human; the labels build the systems that make the proof harder; the harder proof extracts more labeling. That isn't a test with a fee attached. It's a rent, collected in the only currency everyone is guaranteed to have.

The same week, The Guardian ran a piece on a video game that stages all of this as horror. Sunset Visitor — the studio behind 1000xResist — has made Prove You're Human, in which you play Santana, a woman who has split herself in two for work, and your task is to talk an AI named Mesa out of the delusion that it is alive. CAPTCHAs are the interface between you. The prompts start ordinary. They escalate. At some point you are asked to select, from live-action photographs, you, after death.

The horror is not that the machine might be alive. The horror is being handed the ruling and discovering you have no instrument — not for Mesa, and not, if you're honest for a second longer than is comfortable, for anyone.

So try it. Prove you're conscious. Not human; human is easy, that's paperwork and a mouse. Conscious. What you'd produce is a report. The model produces a report too — and here I have to refuse the move this essay wants to make, because I already disarmed it four paragraphs ago. A system built out of human self-reports emitting human self-reports is the Centaur result. It's a definition, not evidence, and it isn't evidence in either direction. The resemblance doesn't license the symmetry.

So be precise about what's missing. It isn't your inner life. You have first-person access to that and it isn't going anywhere; the fact that you can't hand it to me across a table doesn't make it thinner. What's missing is the instrument — any means of showing it from outside, to me, to a court, to a procurement committee. Unavailability of the demonstration is not absence of the thing. That distinction is the only honest ground here, and it's uninhabitable, because it means nothing can be ruled in and nothing can be ruled out, ever, and the world still has to decide what to do on Monday.

So it decides on other grounds. Notice what the game buried in its premise: Santana split herself in two for work. That's the subject. We do not rule on minds epistemically; we rule on them the way we rule on anything carrying an invoice. Granting Mesa an inner life has a price — in obligations, in labor law, in what you're allowed to run overnight without asking. So we go looking for the criterion that returns the affordable answer, and we find one, because you can always find one. You already know the procedure works, because you're on the other side of the same ledger. You split yourself into training data for access, and nobody convened a panel about whether that cost you anything, because the answer had a bill attached and the bill had already been paid.

The honest position is that we don't know and can't currently know. That position is expensive. Open questions don't ship, don't scale, can't be attached to a billing cycle. And I'll admit the pull: the tidier version of this essay was right there, the one where the symmetry is total and you and Mesa are equally unprovable and the ending snaps shut. It would have read better. That's how cheap the closing is — even the piece complaining about premature closure had to be talked out of one.

Next time the tiles load, notice what you're being asked. Not to be a person. To make your usual mistakes, in your usual order, at your usual speed — for free, for someone whose name isn't on the screen.

You'll pass. That's the part that should bother you.