Sensation & Perception: What You See Is the Brain's Best Guess

Any given retinal image could have been produced by infinitely many external scenes. The brain's solution to this inverse problem is not passive reception but guess first, then let input correct the guess. Anil Seth calls it a "controlled hallucination" — what you experience is the guess; the senses only constrain it.

Helmholtz proposed "unconscious inference" back in 1867. The modern version is predictive processing (Clark, Friston): cortex is a hierarchical generative model in which predictions flow downward and only prediction error flows upward. Whatever matches is never forwarded — which is why the contents of consciousness are set mainly by the model, not the input.

The critical dial is precision weighting: the brain's running estimate of how trustworthy the sensory evidence is right now. Good light and focused attention raise the weight on the senses; dimness, fatigue, and noise hand more authority to the prior. Sine-wave speech sounds like noise the first time and turns into a crisp sentence the moment you are told the words — install a prior and perception itself changes.

This is structurally the same claim as the Yogācāra account of the perceived aspect being generated by consciousness: the object is not sitting out there waiting to be picked up, it is constructed. Barrett's theory of constructed emotion (Day 9) is the same mechanism applied to interoception.

Exercise: take one strong judgment ("this plan won't work," "she's brushing me off") and write two columns — the left holds only what a recording could verify, the right holds everything else. How much longer the right column is, is the current weight of your prior.

An illusion is not a sensory malfunction — it is the working trace of a normal inference engine exposed by atypical input. It is a contrast agent, not a bug report. Treating illusion figures as proof that "the brain is dumb" is like mistaking the dye for the disease.

In Adelson's checker-shadow illusion, squares A and B have identical pixel values and you firmly see different shades. The visual system does not output luminance but reflectance — it is inferring the illumination away, discounting the shadow. In the real world this is almost always right; the cost is failing on artificially constructed images.

More importantly: knowing the answer does not dissolve the illusion. Perceptual modules are sealed off from your beliefs (Pylyshyn's "cognitive impenetrability"). Any bias that lives at the perceptual or intuitive level is essentially immune to "I'll be more careful next time" — only process and environment changes reach it (Day 58).

The hollow-mask illusion supplies the mirror-image evidence: a concave mask is forced into a convex face because "faces are convex" is an extremely strong prior. Dima et al. (2009) found people with schizophrenia are markedly harder to fool, with lower top-down connection weighting — turn the prior down and you see more accurately here, while paying for it in hallucination elsewhere.

Exercise: stare for ten seconds at an illusion whose mechanism you already know but still see. Then ask: which of my judgments has the same structure — fully understood, yet the feeling doesn't budge? That one needs a mechanism, not an argument.

When the senses disagree, the brain neither votes nor averages — it sums them weighted by each one's reliability, and in the lab this comes close to statistically optimal. What feels like a holistic impression is a fairly precise estimate.

Ernst & Banks (2002, Nature) had participants judge an object's height by sight and touch at once, then added noise to the visual channel. The visual weight declined smoothly as noise rose, closely matching the maximum-likelihood prediction. Ventriloquism follows the same rule: vision localizes far more precisely than hearing, so the voice gets pulled to the lips.

The McGurk effect (1976) shows that integration is mandatory: hear /ba/ while watching lips say /ga/ and most people hear /da/ — a syllable nobody produced. The crucial addendum is cross-linguistic variation. Sekiyama (1997) found the effect is markedly weaker in native Chinese and Japanese speakers: integration weights are shaped over years by language experience, not fixed biology.

Integration extends to the body itself. In the rubber-hand illusion (Botvinick & Cohen, 1998), synchronized visual and tactile stimulation makes a fake hand feel like your own: body ownership is also an inference. A caveat is required — Lush et al. (2020) showed the effect correlates strongly with suggestibility, and the classic interpretation is under re-examination.

Exercise: take one important message you were about to send as text this week and hold it on video or in person instead. Afterwards write one line: what I assumed their stance was, and what it actually was. The gap is what the text-only channel let your prior fill in.

Looking and seeing are two different things. In Simons & Chabris (1999), participants counting basketball passes missed the chest-thumping gorilla about half the time. That isn't inattention — it is the price of attention.

Expertise grants no exemption. Drew, Võ & Wolfe (2013) embedded an obvious gorilla in lung-nodule screening CT images and 83% of radiologists failed to report it, with eye-tracking showing most had looked straight at it. Fixation is not consciousness.

Lavie's perceptual load theory gives an actionable rule: when a task's perceptual load is high, resources are consumed and irrelevant stimuli are filtered early, so distraction drops; when load is low, spare capacity spills over and attention goes hunting for something else. What makes you most distractible is usually a task that is too easy.

Filtering is also not all-or-none: in the cocktail-party paradigm your own name in the unattended channel still gets through (Cherry, 1953). Change blindness is more extreme — Simons & Levin (1998) swapped the person asking for directions during a brief occlusion and many passersby never noticed. The richness of the world is largely an artifact of it being checkable on demand.

Exercise: after reading AI-generated content that matters, don't ask "does this read as correct?" — check three pre-listed items one by one (numbers, citations, logical jumps). The first invokes your prior; the second bypasses it, and the miss rates are not in the same league.