← Concept Vault
Concept · Neuroscience · Maps to the Law of Mentalism

Predictive Processing

Your brain is not seeing the world. It is guessing the world, then checking the guess against the data — and most of the time you are living inside the guess.
Read · 17 minPillar · Neuroscience

This is the most important idea in 21st-century neuroscience that almost nobody outside the field has heard of, and it quietly demolishes the folk model of perception you've carried your whole life. You think your eyes are cameras and your ears are microphones — that the world pours in and your brain reads it. Predictive processing says you have it backwards. The traffic runs mostly the other way: your brain emits a prediction of the world, and the senses are reduced to a fact-checker that only speaks up when the guess is wrong. Perception is not reception. It is controlled hallucination, constrained by reality. Let's pull it all the way down to the synapses and back up to your lived experience — and mark carefully where the science is solid, where it's a contested framework, and where it's been oversold.

01

What it is

The plain claim: the brain is a hierarchical prediction machine whose fundamental job is to minimize prediction error. It is constantly generating a top-down model of what should be causing its sensory input, sending that model down toward the senses as a prediction, and comparing it against the bottom-up signal that actually arrives. What flows up the hierarchy is not raw data — it's the difference between what was predicted and what came in. The error. The match gets cancelled out and never travels far. Only the surprise propagates.

Stack that idea across the whole cortex and you get a layered system: high levels predict abstract, slow-changing causes ("I am sitting in my office, it is afternoon"), lower levels predict faster, more concrete features (edges, motion, phonemes), each level trying to explain away the level below it by sending down a prediction good enough that no error has to climb. When the predictions are good, the system is quiet. When reality violates the model, error surges up, and the model updates — or, if the model is confident enough, it suppresses the error and you literally don't perceive the discrepancy.

This has three big lineages that converged:

  • The Bayesian brain — perception as statistical inference, computing the most probable cause of ambiguous sensory data given prior beliefs (likelihood × prior → posterior). Helmholtz called perception "unconscious inference" in the 1860s, well before anyone could test it.
  • Predictive coding — the specific cortical implementation, formalized by Rajesh Rao and Dana Ballard in their landmark 1999 Nature Neuroscience paper, originally to explain why visual neurons respond the way they do.
  • The free-energy principle / active inference — Karl Friston's attempt (from roughly 2005 onward) to unify perception, learning, and action under a single imperative: minimize "free energy," a mathematically tractable upper bound on prediction error / surprise.

"Predictive processing" is the umbrella term — popularized in philosophy by Andy Clark (his 2013 Behavioral and Brain Sciences target article "Whatever next?" and 2016 book Surfing Uncertainty) and Jakob Hohwy (The Predictive Mind, 2013).

02

What it does

If the framework is right, prediction-error minimization isn't one trick among many — it's the common currency of an enormous range of brain functions:

  • Perception is inference: your best guess at the world's hidden causes.
  • Attention is precision-weighting: turning up the volume on error signals you've decided are reliable.
  • Action is "active inference": instead of updating the model to match the world, you change the world (move your body) to match the model's prediction. Reaching for a cup is your brain predicting the proprioceptive state of "hand on cup" and letting reflex arcs cancel the error by moving.
  • Learning is updating the priors so future predictions are better.
  • Emotion / interoception is the same machinery pointed inward — predicting the state of your own body (heartbeat, gut, hormones) and feeling the error as affect.

One algorithm, applied recursively, from the retina to the sense of self. That's the audacious part, and the reason it's the leading candidate for a "grand unified theory" of the brain — and also exactly why skeptics get nervous (more on that below).

03

The mechanism — molecular to behavioral

Here's the machinery, built from the bottom up. I'll flag what's established versus what's still inference about implementation.

Cortical microcircuit (the architecture). The neocortex has a famously repeating six-layer columnar structure, and predictive coding maps onto it cleanly enough that it became a major selling point. The working model (anatomically motivated, e.g. Bastos et al., 2012, Neuron, "Canonical microcircuits for predictive coding"): superficial pyramidal cells in layers 2/3 carry prediction-error signals upward (feedforward, via faster gamma-band activity); deep pyramidal cells in layers 5/6 carry predictions downward (feedback, via slower beta-band activity). Two streams, two directions, two frequency bands, anatomically segregated. This laminar/frequency split is supported but not settled — it's the strongest physical evidence for the framework, and also where the hardest critiques land.

Synaptic and molecular level. Predictions and errors aren't free-floating; they're implemented in synaptic weights and neuromodulation.

  • NMDA receptors and synaptic plasticity (long-term potentiation/depression) are how priors get learned — the model is literally stored in the strength of connections.
  • Precision (confidence in a signal — the inverse of its expected variance) is thought to be set by neuromodulators and the gain of specific neurons: acetylcholine, dopamine, and noradrenaline are the leading candidates for the "volume knobs." Dopamine, in this framing, doesn't signal "reward" so much as the precision of beliefs about reward — a reframing of the classic reward-prediction-error story (Schultz, Dayan & Montague, 1997, Science) into the broader predictive scheme.
  • This matters because if precision-weighting is neuromodulatory, then drugs and disease that hit these systems will systematically distort perception — and they do, which is one of the framework's best predictions (psychosis, psychedelics, Parkinson's; below).

Network level — the hierarchy. Each cortical region is both a teacher and a student: it sends predictions down to the region below and error up to the region above. Higher regions (prefrontal, association cortex) encode slow, abstract, invariant causes; lower regions (V1, A1) encode fast, concrete features. Crucially, feedback (prediction) connections often outnumber feedforward (sense-data) connections (the classic example: far more cortical fibers run down into the visual thalamus than arrive from the retina), so anatomically there is heavy wiring for top-down signaling. That asymmetry is real and is one of the framework's favorite pieces of supporting evidence: the brain is built like something that mostly talks to itself and only occasionally listens to the world.

Behavioral level. Out the other end, you get perception, action, and feeling. The phenomenology — what it's like — is the content of the winning prediction, not the sensory input. You experience the model. Reality only edits it.

The compression of the mechanism: predictions flow down, errors flow up, precision decides who wins the argument, and what survives the argument is what you consciously experience.

04

The landmark science

Real names, real dates. This is where established-versus-contested matters most.

  • Hermann von Helmholtz (1860s) — "perception as unconscious inference." The seed of the whole thing, a century early.
  • *Rao & Ballard (1999), Nature Neuroscience* — the foundational predictive-coding model. They showed that if you assume higher visual areas predict lower ones and only error propagates, you naturally reproduce "extra-classical receptive field" effects* (e.g., end-stopping) — neurons suppressing their response to predictable stimuli. This was a genuine, retrodictive* success: a known-but-puzzling neural phenomenon fell out of the model. Established.
  • The mismatch negativity (MMN) — an EEG response, studied since Risto Näätänen (1978), where the brain throws an error signal ~150–250 ms after a sound violates an established auditory regularity (the classic "oddball" paradigm). This is about as close to directly watching prediction error as non-invasive neuroscience gets, and it's robust and replicated. Strongly established.
  • Repetition suppression / expectation suppression — neurons fire less to expected stimuli. *Summerfield et al. (2008), Nature Neuroscience*** showed expectation (not just repetition) reduces responses in face-processing regions. Multiple fMRI and single-unit studies converge here. Established.
  • *Karl Friston — the free-energy principle (2005, 2010 Nature Reviews Neuroscience "The free-energy principle: a unified brain theory?") — the most ambitious and most contested piece. Friston argues any self-organizing system that resists disorder (stays alive) must, in effect, minimize the long-term surprise of its sensory states, which is mathematically equivalent to minimizing variational free energy. It's elegant, deeply influential, and also accused of being so general it's nearly unfalsifiable. Brilliant and contested — hold both.*
  • Anil Seth — popularized "controlled hallucination" (a phrase he credits to earlier researchers) and "the brain as a prediction engine" (his 2017 TED talk, 2021 book Being You). His lab's work pushes the interoceptive angle: the self and emotion as predictions about the body. Influential synthesis; the interoceptive specifics are active research, not settled fact.
  • *Lisa Feldman Barrett — the theory of constructed emotion (2017, How Emotions Are Made) — applies predictive processing to affect: emotions aren't hardwired reactions read off the body, they're predictions* the brain constructs to make sense of interoceptive signals using learned concepts. Influential, with real evidence against "basic emotions," but the constructionist claim remains debated.
  • Computational psychiatry — *Paul Fletcher & Chris Frith (2009), Nature Reviews Neuroscience* ("Perceiving is believing"), and later Philip Corlett, Rick Adams, Karl Friston — reframed psychosis** as aberrant precision: hallucinations and delusions as overweighted priors and/or misweighted error. A productive research program with growing but not conclusive evidence.
05

How it shapes experience — sharp examples

This is where it stops being abstract and starts explaining things you've actually felt.

You hear the word you expected, not the sound that arrived. Play someone a degraded, noise-masked sentence and they hear gibberish. Tell them what the sentence says, replay the identical audio, and now they hear it crisply — the words leap out. The sound didn't change. The prior did. Sine-wave speech demos do this dramatically: a whistly alien noise becomes clear speech the instant you're told what it says, and you can never un-hear it. Your prediction is being installed into your perception in real time.

The dress. The Laurel/Yanny clip. Ambiguous input + different priors = genuinely different percepts, with neither person "wrong." The input underdetermines the experience; your model fills the gap.

The hollow-mask illusion. A concave (hollow) face mask looks convex — popping toward you — because your lifelong prior "faces are convex" is so strong it overrides the actual depth data. Your model wins the argument with reality. Notably, the illusion is weaker in people with schizophrenia — they're less captured by the prior, more driven by raw sensory data. That's predictive processing making a falsifiable, confirmed prediction about a clinical population (work by Danai Dima, Jonathan Roiser and others). One of the framework's cleaner wins.

Why you can't tickle yourself. Your motor system predicts the sensory consequences of your own movement (an "efference copy") and cancels them as expected. Self-generated touch is predicted, so the error is small, so the sensation is dampened. Blakemore, Wolpert & Frith (1998/2000) showed exactly this. Someone else's hand isn't predicted — the error survives — so it tickles. You are perpetually subtracting yourself from your own experience.

Placebo as top-down prediction. A sugar pill given with the strong expectation of relief can genuinely reduce pain — and the mechanism is predictive: the prior "I will feel better" propagates down and attenuates the ascending pain error. This isn't "it's all in your head" dismissal; expectation recruits real descending pain-modulation circuitry (periaqueductal gray, endogenous opioids — block them with naloxone and the placebo analgesia partly vanishes; Levine, Gordon & Fields, 1978). Pain itself is now understood as substantially a prediction about bodily threat, not a direct readout of tissue damage — which is why chronic pain can persist with no ongoing injury: the model got stuck.

Psychosis as the model breaking. If hallucinations are overweighted priors and delusions are the high-level model's desperate attempt to "explain away" a flood of mis-precise error, then voices are predictions the brain can no longer tell apart from sensation, and a delusion is the best story available to a system drowning in surprise it can't cancel. Psychedelics (via 5-HT2A receptor agonism) are theorized in the "REBUS" model (Carhart-Harris & Friston, 2019) to relax high-level priors — loosening the grip of the model so raw, unfiltered error floods up. That's a mechanistic story for why the world looks newly strange, meaningful, and vivid on psychedelics, and why they may help conditions defined by over-rigid priors (depression's hopeless self-model, addiction's grooved expectation). Promising, actively researched, not proven.

06

Where it's distorted or oversold — the honest cut

This framework is powerful enough to be dangerous, and it gets abused in three directions.

1. "Your brain creates reality, so you can think your reality into being." No. This is predictive processing kidnapped by manifestation culture. The framework says your brain constructs your perception of reality from priors + sensory error — it says nothing of the sort about your priors editing the external physical world. The sensory error is the leash. Predict that the floor is gone and step out a window; reality sends an un-ignorable error signal on the way down. The whole point is that prediction is constrained hallucination — constrained by a world that does not care what you believe. Anyone who quotes "controlled hallucination" to sell you that perception is unconstrained has dropped the load-bearing word.

2. The free-energy principle as a theory of everything. Friston's framework is mathematically gorgeous and has a real tendency to "explain" any observation after the fact. Critics (some philosophers and computational neuroscientists) argue it's so general it risks unfalsifiability — that "the brain minimizes free energy" can be made compatible with virtually any data, which makes it more of a framework or modeling language than a testable hypothesis. There's a sharp version of the critique: a thermostat or a rock minimizes free energy too, so what exactly is the theory ruling out? Defenders say the value is in the specific generative models it lets you build, not the slogan. Both can be true: a fertile framework, a weak law. Don't mistake elegance for proof.

3. "It's all top-down, so the senses barely matter." Overcorrection. Yes, feedback connections outnumber feedforward ones; yes, priors are doing enormous work. But a brain that ignored bottom-up error would hallucinate freely and die. Healthy cognition is a balance — precise enough priors to act fast, precise enough error to stay corrected. The pathologies are exactly the imbalances (too-strong priors → hallucination/rigidity; too-weak priors → the world won't cohere). The framework's own logic forbids the "perception is basically made up" reading.

The unresolved technical fights, stated plainly:

  • Does the cortex actually implement predictive coding the way Bastos/Rao-Ballard say (error in superficial layers, prediction in deep layers)? Some careful studies support it; others find the layer/direction story is messier than advertised. Genuinely open.
  • Is "expectation suppression" really error-cancellation, or could it be plain neural adaptation/sharpening? Alternative accounts exist and aren't fully ruled out. Contested.
  • Is the free-energy principle empirical neuroscience or a normative math framework? Contested, possibly category-confused.

What is solid: the brain is deeply predictive and top-down; expectation measurably alters perception and neural response (MMN, expectation suppression, hollow mask, sine-wave speech, placebo); and prediction-error signals are real and recordable. What is framework-level and contested: that one free-energy imperative unifies everything, and the exact cortical wiring of the scheme. Keep those two buckets separate and you can't be sold the myth.

07

The debates and frontier

  • Interoception and the self. The hottest edge: Seth, Barrett, and others argue your sense of being a self — your body ownership, your emotions, your gut feelings — is a predictive model of your own viscera. If true, anxiety is mispredicted interoception, depression is a stuck low-arousal body-model, and conditions like depersonalization are precision failures in the self-model. Evidence is accumulating; the strong claims outrun it.
  • Consciousness. Does prediction-error minimization explain consciousness, or just correlate with it? Predictive processing tells a good story about the contents of experience (why you see what you see) but, like every neuroscience, runs straight into the hard problem — why there's something it is like to run the algorithm at all. It competes/overlaps with Integrated Information Theory and Global Workspace Theory here. Open, deep, unsolved.
  • Active inference and AI. Friston's active inference is now a live program in robotics and machine learning — agents that act to confirm their predictions and reduce uncertainty, an alternative to standard reinforcement learning. Whether it scales is an open engineering question, but it's a real frontier, not a metaphor.
  • Computational psychiatry. The most clinically promising frontier: re-describing autism (high precision on sensory error → overwhelmed by detail, weak priors), psychosis, anxiety, and chronic pain as specific precision-weighting failures, and aiming treatments at the weights. Early, but it's giving psychiatry a mechanistic vocabulary it never had.
08

How to work with it

You can't opt out of running a predictive brain. You can work with the knobs.

  1. Treat your first perception as a hypothesis, not a fact. "I am seeing the world" is false; "I am seeing my brain's best guess" is true. The single most useful habit this framework grants you: a half-second of doubt about your own immediate read of a situation — especially an emotional read, which is your interoceptive model guessing, not a fact about reality.
  2. Update your priors deliberately, because they will run you otherwise. What you expect, you tend to perceive and confirm. A prior of "people are hostile" surfaces hostility (you'll perceive neutral faces as threatening — measurably true in anxiety). Priors are learned in synaptic weights through repeated experience, which means they're changeable through repeated new experience. This is the mechanistic floor under exposure therapy: you don't argue a fear away, you feed the model disconfirming error until the prior updates.
  3. Manage precision, i.e., manage attention and arousal. Precision is the volume knob on error, set partly by arousal/neuromodulation. High anxiety = misallocated precision (everything feels urgent and threatening). Practices that regulate arousal (breath, sleep, exercise) are, in this language, recalibrating which signals your brain trusts. Meditation, on the predictive read, is partly attending to raw sensory/interoceptive error without immediately explaining it away — loosening the grip of habitual priors. That's why long-term meditators report the world looking fresher: less prediction, more data.
  4. Use action, not just thought, to change the model (active inference). You can change your state by changing the world your senses report — posture, environment, who's in the room. Acting "as if" isn't fake-it-till-you-make-it mysticism; it's feeding your predictive system new sensory evidence that updates the prior.
  5. Respect the leash. Constrained hallucination means constrained. The discipline is to keep your model open to error — to actively seek the bottom-up signal that proves you wrong — because the failure mode of a confident predictive brain is to suppress the very errors that would have corrected it. Stay correctable.
09

Connects to

The Law of Mentalism — "The All is Mind; the Universe is Mental." This is the neuroscience case for it, and it's a startling fit. The Hermeticists claimed reality, as experienced, is mental in nature — that the world you live in is constituted by mind. Folk intuition rejects this as obvious nonsense: the rock is out there, not in my head. Predictive processing splits the difference with brutal precision: there is a mind-independent world (the error signal proves it — something out there pushes back), but the world you consciously inhabit is built by your mind, top-down, from priors. Your experienced reality is a model — literally a mental construction — and the external world only edits it at the margins. The Hermetic line "the Universe is Mental" lands almost exactly where Anil Seth lands with "controlled hallucination": you are not living in reality, you are living in your brain's rendering of reality. The law was metaphysical intuition; predictive processing is the mechanism — your experienced universe genuinely is mental, generated by a prediction machine, constrained but not dictated by what's outside. Where Mentalism overreaches (mind creates the external world) is exactly where predictive processing draws the line (the error signal is reality refusing to be wished away). Hold both: your experience is mind; your world is not entirely yours to write.

Related laws and structures:

  • Vibration — the brain runs its predictions and errors on neural oscillations; predictions ride slow beta-band feedback, errors ride fast gamma-band feedforward (Bastos 2012). The "frequency" intuition meets the "prediction" intuition in the same cortical microcircuit.
  • Cause & Effect — the framework is built on it: your priors are the brain's compressed model of which causes produce which sensory effects. Inference is causal reasoning run backward (from effect to most-probable cause). Active inference is causal reasoning run forward (predict the effect of my action, then act).
  • Correspondence ("as above, so below") — the hierarchy is the corresponding-levels structure: the same predict/error/precision algorithm repeats from the retina to the self-concept, micro to macro, the part mirroring the whole.
  • Polarity / Rhythm — precision-weighting is a dynamic balance between two poles (trust-the-prior ↔ trust-the-data); pathology is the imbalance, health is the oscillating middle.

Related concepts: Bayesian inference · the free-energy principle · active inference · predictive coding · the constructed theory of emotion · interoception · the hard problem of consciousness · computational psychiatry · neuroplasticity (how priors physically update) · System 1/System 2 (System 1 ≈ fast, confident priors; System 2 ≈ effortful error-checking).

The Compression

You don't see the world — you see your brain's best prediction of it, and you only feel the world when the prediction is wrong; mastery is keeping your model confident enough to act and humble enough to be corrected.

Your Practice
My Depth
My Notes
← Previous
Hemispheric Lateralization
Next →
The Triune Brain (and Why It's Wrong)