Try to remember the last time you picked up a container you expected to be full and it was empty. Your arm flew up. Nothing about the object had deceived your eyes — you saw it perfectly well — and yet your body had already committed to a force that turned out to be wrong. That small, slightly comic moment is the most direct everyday evidence that perception and action run on predictions, and that you only notice the prediction when it fails.

The Ordinary Version of the Idea

The intuitive model of perception is a camera. Light arrives, the eye transduces it, the brain assembles an image, and you see what is there.

The empty-box moment is difficult to fit into that model. If perception were purely reception, there would have been nothing to be wrong about — the arm would simply have lifted the weight it encountered. Instead a prediction had already been made, the motor system had acted on it, and the error was what reached awareness.

Predictive processing generalises that observation. It proposes the brain is continuously generating expectations about incoming signals and using the mismatch between prediction and input to update itself, rather than building experience from raw data each time.

Two Kinds of Prediction

A 2026 study set out to examine something the framework implies but that had been understudied: that priors come in different kinds and must be weighed against each other.

It distinguishes long-term priors, formed through lifelong experience — how bodies move, what physics does, what faces do — from short-term priors, emerging from immediate contextual cues in the current moment (Journal of Vision, 2026).

To pull them apart, participants watched animated human and robot agents performing biological or mechanical motion inside a probabilistic-cueing paradigm, with cues supplying either form-based or motion-based short-term expectations.

What Happened When Expectations Held — and When They Broke

When the stimuli were congruent with long-term priors, those long-term priors guided perceptual decisions across conditions. A lifetime of watching how living things move is a powerful default, and it won.

When expectations were violated, the pattern changed: performance reflected greater weighting of motion-based short-term cues, while form-based cues contributed minimally.

That asymmetry is the interesting result. It is not simply that short-term cues took over when the long-term prior failed — it is that one type of short-term cue took over. How something moved carried the weight; how it looked did not.

The authors describe this as highlighting the hierarchical nature of predictive processing in action perception: not one prediction, but several, competing, with the balance shifting according to which is proving reliable.

Why the Framework Travels Beyond Vision

If perception generally works by prediction and correction, then the same account should apply to perceiving the body’s own state, not only the outside world.

A 2026 paper in a general practice journal takes exactly that step, applying predictive processing to persistent physical symptoms. It notes that these are among the most common and challenging presentations in general practice — whether investigations come back normal or reveal pathology insufficient to explain the severity — and that patients and clinicians alike often experience them as a dead end.

Its central claim is that traditional biomedical models assume symptoms passively reflect structural pathology, and that normal test results therefore imply the absence of illness. Predictive processing proposes instead that symptoms are active perceptual constructions, generated as the brain infers the state of the body from incoming signals and prior beliefs.

One label matters here and is easy to lose: that paper is explicitly an opinion paper. It develops a clinically applicable account from existing literature and illustrates it with scenarios. It does not report new data, and it should be read as a proposed framework rather than a finding.

What the Framework Does Not License

It does not mean perception is unreliable. Prediction is what makes perception fast and usually correct. A system that waited for complete evidence before committing would be too slow to catch anything, and the empty-box error is notable precisely because it is rare.

It does not mean symptoms are imagined. ‘Actively constructed’ describes how a perception is produced, not whether it is real. Everything anyone has ever perceived is actively constructed on this account, including the ones with obvious physical causes.

The vision study is a laboratory paradigm. Animated agents in a cueing task, chosen because they let two kinds of prior be separated cleanly. That precision is bought by distance from ordinary seeing.

And a framework is not evidence for a treatment. Applying predictive processing to persistent symptoms is a proposal about how to understand and explain them. No intervention follows automatically, and the paper making the case is an opinion piece.

The related question of how the body’s signals get interpreted in the first place is covered in why the physical often arrives before the thought, and the pain version of the same mismatch in what central sensitisation describes.

The Useful Reframe

It is that seeing, hearing and feeling are best understood as inference under uncertainty rather than as measurement. The brain is not reporting the world; it is making its best account of the world and revising it.

That is why illusions work, why a familiar route feels shorter, why the empty box catches you out, and why expectation changes what a sensation feels like without anyone deciding it should.

It is also a genuinely young framework with a lot of enthusiasm attached to it. The vision result above is one careful experiment about one kind of stimulus. The clinical application is an argued proposal. Both are worth knowing about, and neither is settled.