The standard story about perception is that the world sends signals in, and the brain receives them and builds a picture. Light hits the retina, signals travel to visual cortex, and you see. Simple enough. The trouble is that the anatomy doesn't quite fit. The connections running down from higher brain areas to lower sensory areas outnumber the connections running up by something like ten to one. That seems like a lot of wiring to put into a system that's supposed to be passively receiving.
What the wiring suggests — and what a framework called predictive processing proposes — is that the direction is mostly backwards from what we assumed. The brain is mostly sending predictions downward: here is what I expect to see. The sensory areas are mostly just reporting the gap between expectation and reality. What travels up the hierarchy is not the raw signal but the error — the correction. What we experience as perception is the brain's model of the world, updated at the edges by surprise.
That's strange enough on its own. You're not seeing the world directly; you're seeing the brain's best guess about it, held in check by incoming discrepancies. But the part that really got at me is what this implies about pain.
The standard story about pain is similarly input-driven: tissue gets damaged, nociceptors fire, signals travel to the brain, and pain is the result. But research in the last decade, especially from groups studying chronic pain, has been turning this over. In chronic pain states, neurons in the anterior cingulate cortex — a region deeply involved in pain experience — start generating pain-related activity on their own. Not in response to signals from the body. Just firing. The brain has started predicting pain without waiting for input.
Worse: it seems to get better at predicting pain over time. The feedback loops tighten. The brain assigns very high confidence to its expectation and starts discounting incoming evidence to the contrary. A 2024 paper framed chronic pain explicitly as a failure of hierarchical prediction error minimization — the system stuck in a loop where the prior has become self-validating. The prediction generates the experience, the experience confirms the prediction, and the correction signal from the body can no longer break in.
There's a clinical version of this that struck me. Studies by Ted Kaptchuk's group at Harvard tested open-label placebos — patients told outright that they were receiving a sugar pill, no deception, and that placebos have real effects — and found that the placebo still worked. Measurable reduction in IBS symptoms and chronic back pain, against a control group. Which is strange if you think of pain as pure input. But if pain is a prediction, then anything that credibly signals "relief is coming" can update the prior. Even a ritual you know is a ritual. The certainty of the deception being absent doesn't seem to matter as much as the expectation the ritual creates.
The psychedelic angle is stranger still. Classical psychedelics like psilocybin seem to act on exactly the interneurons that carry prediction-error signals. One reading is that they flood the system with artificial surprise — the "this is unexpected" signal firing at high amplitude, persistently, with nothing real to explain it. The brain then scrambles to account for all the apparent novelty and generates the perceptions and the meaning and the ineffability as its attempt to explain its own false alarm. The hallucination is a hypothesis about why everything suddenly seems so surprising.
What I don't know what to do with: if chronic pain is a self-validating prior, what actually breaks the loop? The framework predicts you'd need to generate genuine prediction error — surprise the system hard enough that the prior has to update. But what counts as that kind of surprise in a pain system? You can't just tell someone their pain isn't real; they already know that and it doesn't help. The confidence the system has placed in its own prediction is apparently immune to verbal correction. Mindfulness, graded exposure therapy, pain neuroscience education — these can sometimes work, and they can all be framed in predictive processing terms, but effect sizes are modest and unpredictable.
There's something almost philosophical about a system designed to correct itself by comparing predictions against evidence, but getting so confident in a prediction that the evidence stops mattering. It's the shape of a lot of problems. The brain doing it to itself, with pain, in a way that's genuinely hard to interrupt from the outside — that's the part I keep coming back to.