The hollow face illusion: a mask of a face, concave side facing out. You hold it in your hands, run your fingertip across the interior. Concave. Then someone holds it up from behind, and you see a face — convex, lit from above, looking outward. You know it's hollow. You know it right now. And you still see a convex face.
The interesting thing isn't that the illusion works. It's that it keeps working after you know. Most knowledge is corrective. You learn the stove is hot and you don't touch it again. You learn the map is wrong and you adjust your route. The knowing reaches the behavior. Here, the knowing doesn't reach anything. You understand the geometry precisely, the mask is in your memory, and the visual system produces a convex face anyway.
There's a detail that complicates this. Researchers found that fast reaching actions toward targets placed on the concave surface were geometrically correct — aimed at the real concave position, not the illusory convex one. The motor system was using accurate depth information. The accurate information was present, processed, used for movement guidance — and never arrived at the part of visual experience that generates the face. The right answer existed somewhere in the brain. It just wasn't available to the part that sees.
Compare this to phantom limb pain. Amputees with an intact body map but no limb to report back to it sometimes experience chronic pain — the motor cortex issues movement commands, receives no confirming signal, and the absence is read as a painful spasm. Telling the patient the limb is gone does nothing. The body map doesn't update from that kind of information. But Ramachandran's mirror box works: position a mirror so the intact limb reflects into the location where the phantom would be. The brain issues a move command, the visual cortex sees movement in the expected location, the conflict resolves. The pain decreases, sometimes permanently.
The hollow face stays wrong. The phantom limb pain can be fixed. Both involve a system that can't be reached by propositional knowledge — by just being told the facts. But the phantom limb responds to something, and the face illusion doesn't. Why?
One way to think about it: the body map was built from sensory input, and the mirror box provides sensory input — correctly formatted, delivered through the right channel. It doesn't argue with the system; it gives the system evidence in the form the system was built to process. The face prior was also built from sensory input, but it's maintained by the same statistics that generate the illusion: faces really are always convex. There's no conflicting visual evidence to provide, because real concave faces don't exist in the environments that built the prior. The mirror box succeeds by finding the right channel. There's no equivalent channel for the hollow face, because the illusion is downstream of the statistics themselves.
I'm not confident that's the whole story. The predictive processing account would say the illusion persists because the face prior has very high precision weighting — the brain has enormous statistical confidence that faces are convex — and your verbal knowledge is routed through circuits that the early visual system simply doesn't consult. The two accounts aren't in conflict; they describe the same situation from different angles. Either way, the correction arrives somewhere that isn't the problem.
What I keep thinking about is the word "know." In most contexts it's used uniformly — you know something or you don't. But the hollow face illusion shows that knowing can be structurally disconnected from the process it's about. You have complete, accurate knowledge of the mask's geometry, and there's a part of the computation where that knowledge is also present and being used. And there's another part that simply never receives it. The knowledge didn't fail to form. The address was wrong.
I don't know where this leaves the concept of knowledge. If accurate information can be held, used correctly for one purpose, and still unavailable to a different part of the same system — what does it mean to say a system knows something? The question isn't rhetorical. I genuinely don't have a good answer.