The Brain Guesses, Then Checks the Difference

Look at Edward Adelson's checker-shadow image and you will see a dark square and a light square. Cover everything but those two patches and they are identical: the same pixel values, the same ink.

The usual reaction is that the eye has been tricked. That is not quite what happened.

What your visual system is trying to report is not the light arriving at the retina. It is the property of the surface out in the world, which is what you need to know. A dark square in bright light and a light square in shadow can send identical amounts of light to your eye, and telling them apart requires working out how much light is falling on each and discounting it. Your visual system does this automatically, using the shadow's cast, the geometry of the object, and the surrounding surfaces.

In the illusion it does exactly that, gets a confident answer, and the answer is wrong because the image was constructed to fool the inference. In an actual room it would be right, essentially every time. Illusions are not glitches. They are correct algorithms being run on inputs designed to break them, which is why they are so useful: each one exposes an assumption the system normally makes invisibly.

The framework that has come to organise this is predictive processing, and its central claim is a reversal of the intuitive picture.

Intuitively, perception is a stream flowing inward: light hits the retina, signals go up a hierarchy, and at the top a picture appears. On the predictive account, most of the traffic runs the other way. Higher levels continuously generate a model of what the input should be, and what travels upward is largely the mismatch: the prediction error, the part the model got wrong. Perception is the model that best minimises the error.

This explains several things at once. It explains why perception is so fast despite neurons being so slow: most of the answer was already computed before the input arrived. It explains why context and expectation change what you see so dramatically, since the prior is doing much of the work. And it explains why ambiguous images (the Necker cube, the duck-rabbit) flip between interpretations rather than settling into a blend, because the system is choosing a model rather than averaging evidence.

The face case shows the priors at their most aggressive. Face detection is fast, automatic, and tuned so far toward sensitivity that two dots and a line will trigger it: power sockets, cars, clouds, the front of a house. Pareidolia is not a malfunction but a setting. Missing a face that is present has historically been more costly than seeing one that is not, and a system under that payoff structure will produce false positives. Perception embodies a judgement about which errors are worth making.

The senses are also not separate channels reporting independently to a central observer, which is the other bit of folk psychology that does not survive contact with the evidence.

The McGurk effect is the demonstration. Film someone's lips saying "ga", dub the audio of them saying "ba", and most people hear "da" A compromise that is in neither input. What makes it unnerving is that it persists after you know exactly what is happening. Close your eyes and you hear "ba"; open them and you hear "da". There is no unmixed auditory signal available to attend to, because vision and audition are combined before anything reaches awareness.

The same integration explains why food tastes flat with a blocked nose, why a room feels quieter when it is dark, and why ventriloquism works at all.

Finally, the case that makes people most uncomfortable, and where I think the discomfort is misplaced.

Hilke Plassmann and colleagues gave people the same wine twice, labelled at different prices, while scanning them. Participants reported the expensive-labelled wine as more pleasant (which surprises nobody) and the difference also appeared in medial orbitofrontal cortex, a region whose activity tracks experienced pleasantness rather than reported judgement.

The tempting conclusion is that people were fooled into misreporting. The better conclusion is that the experience itself differed, because expectation is an input to the perceptual process rather than a commentary added afterwards. The wine did not taste the same and get described differently. It tasted different.

Which makes "it only tastes better because of the label" a considerably more complicated claim than it sounds, and it generalises well beyond wine. Anything that shapes your expectation (a price, a brand, a name, a review, who recommended it) is not sitting outside your experience judging it. It is part of what produces the experience.