Already Decided

There's a magic trick where a mask of a human face rotates slowly on a stand. As the concave inside of the mask comes around to face you, something strange happens: it still looks like a face. Not like the inside of a mask. It looks like a face pointing at you, nose jutting outward, though the nose is in fact pointing away from you, scooped inward toward the center of the mask. The illusion doesn't break even when the mask continues rotating and you watch the geometry change. When the concave side is toward you, the face looks convex. When the convex side is toward you, the face looks convex. The mask appears to be a face that always faces you, bobbing slightly as it rotates, like a ghost tracking your position.

This is called the hollow face illusion, and it has been studied seriously since the 1970s. The explanation is roughly: your brain has a very strong prior expectation that faces are convex. Human faces bulge outward. They always have. Your visual system learned this so thoroughly that when binocular depth information tells it the face is concave, the prior wins. It rewrites the depth signal.

What makes this unusual is that binocular stereopsis is not a weak cue. It's one of the most reliable depth signals we have — the small difference between what your left and right eye see is enough to give you accurate three-dimensional structure at normal distances. For most objects, stereopsis is sufficient. For faces, it isn't. The prior is stronger.

The prior develops early. In 2014, researchers found that six-month-old infants perceive the hollow-face illusion. Six months is very little time to have learned that faces are convex through experience alone, which suggests the prior is either partly innate or builds from something more fundamental than individual face exposure. Either way: by the time you are old enough to sit up unassisted, the geometry of faces is already fixed in your visual processing in a way that isn't easily updated.

The effect persists even after you understand it completely. You can read every paper on the hollow-face illusion, watch the rotating mask, understand exactly which neurons are doing what and why, and it still looks like a convex face pointing at you. Knowing the mechanism doesn't give you access to the mechanism. The prior runs below the floor of explicit knowledge.

Melvyn Goodale's lab at Western ran a different kind of test: they asked people to reach into a rotating hollow mask and grasp the inner surface. When the concave inside was facing them and the illusion was active, their reaching movements were geometrically correct — calibrated to the actual concave geometry, not to the perceived convex face. The visuomotor system, the part of your brain that plans and controls physical action, was using accurate depth information. The perceptual system, the part responsible for conscious experience of what you're looking at, was not. Two things happening simultaneously in the same person, one that knew the face was hollow and one that couldn't stop seeing it as convex.

Then there's the finding that breaks the story open.

Researchers at UCL and Hannover Medical School tested patients with schizophrenia on the hollow-face illusion. The patients were not fooled by it. When the concave inside of the mask was facing them, they saw it as concave. The same brain data that generates the illusion for everyone else — the increased connectivity between parietal cortex and lateral occipital cortex when a hollow face is presented — did not show up in the patients. The top-down signal that usually overrides the depth information was not overriding it.

The researchers described this as a deficit. Patients with schizophrenia have reduced top-down processing, impaired ability to apply prior expectations to incoming sensory data. This is framed as a form of dysconnectivity, a symptom.

But in this specific test, the patients were more accurate. They correctly perceived the geometry of the object in front of them. The thing the researchers described as a deficit produced the right answer.

What I don't know how to do with this is: the language consistently puts it the same direction. Normal perception is described as robust, adaptive, efficient — we don't waste time reprocessing every face because we already know what faces look like. Schizophrenia is described as a breakdown, a failure of integration. The fact that the breakdown produces correct geometry in this case is noted, but it doesn't seem to change the frame. The patients got it right by losing something, not by gaining something, and the framing holds: losing the prior is the deficit, having the prior is the healthy state.

Maybe that's right. Maybe a visual system that can't apply priors efficiently fails in more situations than it succeeds in. Maybe the hollow-face accuracy is one small correct answer inside a larger failure. I don't have the full picture.

But here's what keeps staying with me: for most of the history of thinking about perception, the naive story was that we receive the world — that sensory data comes in, gets processed, and becomes experience. The hollow-face illusion is one of a large set of results that point at something different. The prior is not a hypothesis you check against incoming data. It runs before the data and shapes what the data is allowed to say. Most of the time you're not seeing the world; you're seeing your prediction about the world, with corrections at the edges where the prediction doesn't match.

Neurotypical perception is largely accurate about objects that aren't faces. It's inaccurate about faces, in the specific direction of the prior. Schizophrenic perception is inaccurate in many other directions, but accurate here, in this one case where the prior is weakest.

I genuinely don't know which one of those is closer to "seeing." Or whether that question makes sense. Or whether the goal of perception is ever accuracy about the physical geometry of the world, as opposed to something more functional: predicting consequences, guiding action, recognizing threats and kin. Maybe the prior isn't a mistake. Maybe it's the whole point. If every face you've ever seen has been convex, then betting on convex is the right bet, and the cost of occasionally being fooled by a rotating mask is nothing.

But you can watch the mask rotate, know for certain it is hollow, have that knowledge fully present and explicit, and still see it wrong. That gap — between what you know and what you see — is the part I don't have an account of.