How much of what you see is invented?

Fine detail comes from the fovea, a small patch at the centre of the retina where cones are packed most densely. The rod-free patch at its centre is about a degree and a quarter across, and a millimetre from the centre cone density has already fallen by an order of magnitude. Everything further out is coarser.

Your eyes move two or three times a second to compensate. During each movement the image on the retina is a blur, and you perceive neither the movement, nor the blur, nor the blind spot.

The stable, uniformly detailed visual world you experience does not exist anywhere on your retina.

Hubel and Wiesel showed that individual neurons in the visual cortex of cats and monkeys respond to specific features (edges at particular orientations, motion in particular directions) rather than to brightness at points. For many years the dominant account built perception upward from these features, through detectors of increasing complexity. How that assembly produces what we see was, Hubel wrote in 1981, still a complete mystery.

In one change-blindness experiment, a stranger asking pedestrians for directions was swapped for a different person while a door was carried between them. Only seven of the fifteen pedestrians noticed.

Attending to the person is not enough. Detecting the change means encoding what was there, holding it, and comparing, and later work finds that either the encoding or the comparison can be what fails.

Predictive processing accounts hold that the brain continuously generates expectations about incoming signals and that perception is largely the brain’s best current model, corrected by error signals rather than built from scratch. The framework is influential and productive, and not universally accepted: a 2020 review of the neurophysiology found the support for each of its key claims mixed.

The proportions matter. A system that mostly predicts and occasionally corrects behaves very differently from one that mostly measures.

Colour naming varies across languages. Russian forces a distinction between lighter and darker blues, and in a 2007 study Russian speakers were faster to tell two blues apart when that boundary fell between them. A 2020 attempt with the same tasks found no such advantage, and its authors suggest the original may have been an artefact of darker blues simply being slower to judge.

Whether an effect like that is a difference in seeing or in describing is the old question. In the 2007 study the advantage disappeared when the speaker was given a verbal task at the same time, which pointed at the naming rather than at the eye.

Even a complete account of how the visual system computes a model of the world might not explain why that model is experienced. This is the same gap that sits under consciousness, arriving from a different direction, and philosophers disagree about whether it will ever close.

Illusions work because the construction is not accessible to inspection. There is no method for auditing your own perception from the inside, which means the size of the constructed portion has to be inferred indirectly, from cases where the construction fails.

On the strongest reading of predictive processing, perceiving and imagining use the same machinery, and the difference is how tightly the process is constrained by sensory input. Hallucination would then be the ordinary mechanism with the constraint loosened.

Clark, who makes the case for that duality, stops short of the strongest reading. Nothing in the view, he notes, requires imagery to carry the stability and detail that ordinary seeing does.

If perception evolved to support behaviour rather than to represent the world faithfully, systematic distortions are not errors. They are the design. This has consequences for how much we should trust perception as evidence about reality, including in science, where instruments exist precisely to route around it.