The prediction machine
For a century we assumed perception flows one way: world in, picture out. Neuroscience now suggests we had the arrow backwards.
The framework is called predictive processing. Rather than passively receiving the world, the brain continuously generates a model of what is probably happening and compares incoming sensory data against it. What reaches your awareness is not raw reality — it's the brain's best guess, corrected by prediction errors whenever the world surprises it.3,4,5
Perception, in this view, is a Bayesian negotiation: prior expectations weighed against sensory evidence, each weighted by how reliable — how precise — the brain estimates it to be.7 Most of the time this is spectacularly useful. It's why you can read messy handwriting and hear speech in a noisy room.
But here's the turn that matters for therapy: the same machinery that predicts sights and sounds also predicts people. What a face means. Whether reaching out will be met with warmth or contempt. Whether you are the kind of person others stay for.
If it's hysterical, it's historical
Entrenched maladaptive patterns — the real target of every therapy — are old predictions that were once reasonable.
Strip away the brand names and virtually all psychotherapy is aimed at the same thing: entrenched maladaptive patterns — ways of responding that get us in trouble, that are sticky for a reason, and that, being patterns, recur.
Where do they come from? Usually from childhood — and not only from capital-T trauma. Sometimes the origin is simply a human event a child witnessed and couldn't understand, carried forward as a schematic memory.2,8 Children solve problems differently than adults: they turn to the grown-ups. And children don't compute nothing — "there are no monsters" doesn't land. When something goes wrong, it has to be someone's fault, and it is often easier to conclude the fault is theirs.
"When I express emotional needs, people reject or ignore me" starts as a repeated observation. With enough repetitions, it stops being a thought and becomes a high-confidence prediction: vulnerability → rejection. Decades later, the adult walks into every room already forecasting the rejection — an ambiguous expression, a delayed text, a tired partner all get rendered by the model as confirmation.
The 2025 paper behind this framework opens with a striking clinical moment: a patient with a history of childhood emotional neglect, eyes lowered, grows angry and accuses his therapist of looking at him coldly. She points out that he hasn't yet looked at her face. He looks up — and is astonished to find warmth there.2 His brain hadn't observed her expression. It had predicted it.
Update a belief
The indigo curve is a prior — this brain's confidence about what happens when it shows vulnerability (here, expecting rejection ≈ 78/100). Each button press delivers a disconfirming experience (a warm, attuned response ≈ 18/100). Watch the belief update — then drag the rigidity slider up and see why entrenched beliefs barely move.
This is the textbook Bayesian toy model, not a simulation of a brain — but it captures the two clinical facts the framework explains: change requires repeated disconfirming experience, and highly precise (rigid) priors can discount almost any amount of contradictory evidence.2,7
Why painful beliefs refuse to update
If brains learn from prediction error, why doesn't ordinary life fix us? Because not all errors are weighted equally.
The brain estimates the precision of every signal. When a deeply encoded model says people abandon me, and someone behaves lovingly, there are two available explanations: my model is wrong — or this person is an exception, or pretending, or hasn't really gotten to know me yet. A sufficiently confident prior simply discounts the evidence.2,9
Seen this way, maladaptive schemas aren't irrational glitches. They were usually adaptive models learned under earlier conditions — and they still deliver something the brain values enormously: predictability. A world in which "people will hurt me" is miserable, but it is legible. Revising the model means tolerating uncertainty, and the short-term comfort of a predictable misery can outweigh the long-term cost.2
Most of this machinery runs outside awareness. The oft-quoted figure that ~95% of mental activity is nonconscious is a heuristic, not a measurement — but the underlying point is well supported: automatic, nonconscious processes drive far more of perception, evaluation, and behavior than introspection suggests.10 This is part of why purely cognitive approaches — explaining the belief to the patient — often aren't enough on their own. The model has to be updated where it lives.
Three ways minds actually change
Words spoken in a session and an interpersonal relationship somehow rewire a brain. Three mechanisms — each a different route to the same event: a salient prediction error.
Extinction
Pavlovian conditioning in reverse. A combat veteran startles at every loud noise; reactivate that alarm in a safe, supported setting and the cortex learns the war is over — inhibitory signals descend and block the fear response, even though the amygdala still fires.11
Modern exposure therapy explicitly engineers this as expectancy violation: the bigger the gap between predicted catastrophe and actual outcome, the stronger the new inhibitory learning.11
Memory reconsolidation
When the neural pattern encoding an old prediction is vividly reactivated and simultaneously met with contradicting information — a prediction error — the memory becomes temporarily labile. For a window of several hours, new information can rewrite the network: the old prediction is revised and a new one takes its place.12
Cognitive dissonance points at a related engine: the mind is intolerant of inconsistency between simultaneously active beliefs and works to resolve it.13
Corrective emotional experience
Described by Alexander and French in 1946.14 A client braces for a specific negative response — judgment, dismissal, being thrown out of therapy — and the therapist responds otherwise. That gap is a golden opportunity: a salient prediction error, the raw material for extinction and reconsolidation to work with.
It's a core tenet of emotion-focused therapy: new lived experience with a supportive other heals, disconfirms pathogenic beliefs, and provides interpersonal soothing.15
Notice the shared recipe. The old pattern must be alive in the room — vividly activated, not just discussed — at the very moment the disconfirming experience arrives. This is why triggers, clinically, are friends to follow: a triggered pattern is an activated pattern, and an activated pattern is an editable one.12
The patient is running experiments on you
Predictive processing explains how beliefs update. Control-mastery theory explains why anyone would volunteer for it.
Why would a person deliberately walk into situations that threaten their deepest models? Control-mastery theory — developed by Joseph Weiss and the San Francisco Psychotherapy Research Group — proposes that patients arrive with an unconscious plan to master their traumas, and that they actively test the therapist against their pathogenic beliefs.16,17
A patient whose model says if I express my needs, I'll be abandoned may gradually become more needy and vulnerable in session — effectively running the experiment: will you abandon me too? Another may provoke, withdraw, or challenge boundaries, recreating an old relational configuration to see whether history repeats. CMT distinguishes transference tests (reproducing the old situation with the therapist in the old role) from passive-into-active tests (recreating it with roles reversed).2
Wait long enough — this is the clinical folk wisdom — and transference heats up on its own; old patterns surface, and then you have something to work with. The 2025 integration paper argues that this is precisely the point: effective therapy requires not just a therapist who supplies salient prediction errors, but a patient motivated and ready to generate the tests that make those errors possible.2
Process > method: one mechanism, many brand names
CBT, ACT, DBT, IFS, EMDR, EFT, MI, PE, MBSR — hundreds of named schools, each with its own vocabulary for overlapping phenomena.
One therapist says schema modification; another, working through the transference; another, extinction learning, cognitive restructuring, attachment repair, experiential processing. Decades of outcome research show these supposedly rival treatments perform far more similarly than their theories predict — the "dodo bird" pattern that common-factors researchers have documented since the 1930s.18 Predictive processing offers a candidate explanation: they may be different routes into the same learning machinery.1,2
| Therapy | In its own words | In predictive-processing terms |
|---|---|---|
| CBT | Examine the evidence; run behavioral experiments against distorted thoughts. | Deliberate prediction-error generation against explicit beliefs. |
| Exposure / PE | Face the feared situation until fear subsides. | Expectancy violation → inhibitory learning; the predicted catastrophe fails to occur.11 |
| ACT | Defuse from thoughts; "I'm having the thought that I'm worthless." | Reduce the precision of the prior rather than its content — loosening its grip on perception and behavior. |
| Psychodynamic | Work through the transference in the therapeutic relationship. | Relational predictions re-enacted with the therapist, met with disconfirming responses → model revision. |
| Attachment-based / EFT | Provide a secure base; transform emotion with new lived experience. | Prior: dependency → danger. Repeated experience: dependency → safety. Cumulative error → updated relational model. |
| Behavioral activation | Schedule activity despite anhedonia. | Depressed prior: nothing will feel rewarding. Engagement produces reward → prediction error → revision. |
140 years of arguing about the same mechanism
Every era of psychotherapy discovered a real piece of the puzzle and named it in its own language. Walk the timeline and watch each school's central insight translate into the same computational vocabulary.
The psychedelic connection
If entrenched beliefs are over-precise priors, then a drug that temporarily relaxes the precision of priors would make minds unusually revisable. That is exactly what one leading model of psychedelics proposes — and it gives the therapy story a pharmacologic lever.
The REBUS model — relaxed beliefs under psychedelics — proposes that psychedelics acting at the serotonin 5-HT2A receptor reduce the precision weighting of high-level priors, so information that would normally be suppressed or explained away gains influence over the model.19
Take the patient from Part II, whose prior says when I reveal who I really am, people reject me. Ordinarily the prior wins by discounting: the therapist is warm → she's being professional; the partner stays → they'll leave eventually; a friend expresses love → they don't really know me. REBUS predicts that when confidence in the prior temporarily falls, the same evidence can land differently — and not merely as an intellectual concession, but as something experientially compelling enough to revise the model underneath.
The first direct human test of the belief claim
Most psychedelic neuroimaging tests REBUS indirectly, by showing the brain becomes less constrained. One study went at the psychological claim head-on. Zeifman and colleagues gave healthy volunteers 1 mg and 25 mg psilocybin four weeks apart and measured confidence in personally held beliefs before, during, and four weeks after. Confidence in negative self-beliefs fell after the 25 mg dose and not the 1 mg dose; acute EEG entropy and the intensity of subjective effects tracked the size of that drop; and the reduction related particularly strongly to increases in well-being at four weeks.20
The authors call it the first empirical evidence that relaxation and subsequent revision of negative self-belief confidence mediates psilocybin's positive psychological outcomes — while stating plainly that replication in larger, clinical samples is necessary.20 With n = 11 healthy participants, that caution is warranted. But it moves the field from "psychedelics make the brain more flexible" to the questions that actually matter clinically: which beliefs lose certainty, which subsequently change, and does that change mediate symptom improvement.
The molecular floor under all of this
The computational story would be idle speculation without a plasticity mechanism, and that layer has firmed up considerably:
- 5-HT2A is necessary. Vargas et al. showed psychedelic-induced cortical structural plasticity depends on 5-HT2A signaling — and, surprisingly, on intracellular receptors, which may explain why serotonin itself doesn't produce the same plasticity.21
- Specific circuits carry the lasting effect. Shao et al. found that psilocybin's long-term behavioral action in mice required pyramidal tract neurons in medial frontal cortex; knocking out 5-HT2A receptors abolished both the behavioral effect and the structural plasticity.22
- A neurotrophic route. Moliner et al. reported that LSD and psilocin bind the BDNF receptor TrkB with far higher affinity than conventional antidepressants, and that plasticity effects in mice depended on TrkB while head-twitch (hallucinogenic-like) effects depended on 5-HT2A — suggesting the two may be separable.23 This one is genuinely contested; treat "psychedelics work through TrkB" as an open hypothesis, not a finding.
Taken together: 5-HT2A → cortical signaling → BDNF/TrkB-associated plasticity and synaptic remodeling is substantially more concrete than it was five years ago. And the two levels aren't competitors. Molecular plasticity permits learning; relaxed priors determine what can be relearned.
What the human imaging shows
Siegel and colleagues tracked healthy adults with precision functional mapping before, during, and for weeks after 25 mg psilocybin versus methylphenidate. Psilocybin disrupted functional connectivity in cortex and subcortex more than threefold beyond the active control, with some changes persisting past the acute state.24 The recurring pattern across this literature — reduced within-network integrity, increased cross-network communication — is what you would expect from a temporarily less constrained hierarchical system.
But flexibility is not automatically therapeutic, and one finding says so directly. Doss et al. found psilocybin therapy increased cognitive flexibility in patients with major depression for at least four weeks — yet those improvements did not correlate with the antidepressant response, and greater increases in neural flexibility were associated with less improvement in cognitive flexibility.25 That is exactly the sort of nuance a good theory has to survive.
The window is the medicine
The most clinically useful idea in this literature isn't "the trip heals you." It's that the hours and days after a session may be when the brain is unusually capable of learning — and what fills that window determines what gets consolidated.
Nardou, Dölen and colleagues showed in mice that a range of psychoactive drugs — ibogaine, ketamine, LSD, MDMA, psilocybin — reopened a developmental critical period for social reward learning. Strikingly, the duration of the reopened window was proportional to each drug's duration of acute subjective effects in humans, and the effect was paralleled by metaplastic restoration of oxytocin-mediated long-term depression in the nucleus accumbens.26
That is an experience-dependent plasticity model, and it makes the word integration far less woolly. Integration need not mean decoding what the jaguar meant. It can simply mean deliberately supplying adaptive experiences and behaviors while the system is unusually capable of relearning: psychotherapy, exposure, interpersonal repair, behavioral activation, sleep, exercise, social connection, approaching what has been avoided.
What fills the window?
A dosing session opens a period of unusual revisability. Pick what happens next and watch what consolidates. The model here is multiplicative, not additive — drug-induced plasticity × environmental learning — which is why the same molecule can produce very different durable outcomes.
Illustrative, not quantitative. The critical-period reopening is a mouse finding;26 that a comparable window makes human psychotherapy substantially more effective is very plausible and, notably, still unproven. It's among the most important open questions in the field.
The traditional framing — is the drug doing the work, or the therapy? — is probably the wrong decomposition. If the drug makes certain learning possible and the environment determines what is learned, then a single session producing changes that last months is no longer puzzling. The molecule is long gone. The learned model isn't. That is what you'd expect of a catalyst for durable learning rather than a continuously present pharmacologic antidepressant.
A loosened brain is not a healthier brain
REBUS is often summarized as relaxed beliefs → therapeutic change. The honest version needs two more terms: relaxed beliefs + salient experience + adaptive learning → potentially therapeutic change.
Reducing the precision of priors is not inherently good. It's neutral. Loosen I am fundamentally worthless and you have done something wonderful. But the same loosening applies to priors that were doing real work:
My perceptions aren't always reliable.
Coincidences aren't necessarily messages aimed at me.
The person running this retreat probably doesn't possess supernatural knowledge.
When top-down constraints loosen, the brain explores a wider hypothesis space. Some novel models fit reality better. Others are spectacular nonsense. And psychedelic confidence can make both feel profound. This is the strongest argument against the simplistic "psychedelics reveal truth" narrative: the state that produces insight is the same state that produces false insight.
Suggestibility as mechanism, not nuisance
Here's where it becomes an ethical matter and not just a theoretical one. If priors soften and environmental information gains weight, then the therapist becomes unusually influential — and so do music, room, expectations, preparation, therapist language, cultural narrative, and the post-session social environment.
This is usually discussed vaguely as "set and setting." Predictive processing gives it a mechanism: setting supplies prediction errors and candidate replacement models during a state of altered precision weighting. There is direct experimental support that psychedelics shift how social and suggested information is taken up: LSD increases suggestibility on standardized measures,27 and in a placebo-controlled crossover study it increased adaptation to others' opinions — though notably only when those opinions were already close to the participant's own, an effect blocked by the 5-HT2A antagonist ketanserin.28
"Notice what comes up."
versus
"That memory means your father abused you."
Spoken during a state of heightened plasticity and social uptake, the second could install a model rather than discover one. This is why non-directive stance, informed consent, monitor training, and documented boundaries are not bureaucratic ornamentation in psychedelic trials. They are part of the mechanism's safety envelope.
Two disorders that fit the model unusually well
PTSD, and the cleanest analogy in the field
A trauma survivor carries a powerful prior: trauma cue → danger. Successful exposure doesn't erase the historical association; it builds competing learning — trauma cue ≠ danger now — and it works best when it maximizes expectancy violation.11 In this framework, a psychedelic would do two things at once: lower the precision of the threat prior, and raise the salience of the contradictory experience. Precision of threat prior ↓, salience of disconfirming experience ↑, revised generative model. That's a more mechanistically interesting proposal than saying patients "process trauma."
Generalized anxiety, and the world's largest behavioral experiment
Anxiety is almost tailor-made for a predictive account, because so much of the pathology is overweighting predictions of future threat and uncertainty. In simplified form, GAD runs as a loop: an uncertain future meets a catastrophic generative model; high precision is assigned to threat predictions; attention searches for confirmation; worry and avoidance transiently reduce uncertainty — and the model never gets adequately falsified.
This is one reason the recent GAD data are theoretically interesting. In a phase 2b randomized, double-blind, placebo-controlled trial across 22 US sites, a single dose of MM120 (lysergide) produced a statistically significant dose–response on the Hamilton Anxiety Rating Scale at week 4, supporting 100 µg as the dose for pivotal trials.29 Durable improvement after essentially one pharmacologic exposure is far easier to conceptualize under a learning and model-updating account than under conventional receptor-occupancy pharmacology.
An important caveat cuts against over-reading this in the therapy direction: these were largely drug-plus-monitoring designs, not manualized psychotherapy trials. The framework predicts strongly that context matters. It does not predict that 8–12 hours of branded manualized psychotherapy are necessary. A supportive environment, a powerful corrective experience, and strategically timed integration may prove sufficient for many indications — and distinguishing those possibilities is an empirical job, not a theoretical one.
Loosen → experience → surprise → update → consolidate
Duane's model says psychotherapy changes us by giving the brain experiences its old model cannot successfully explain. REBUS adds that psychedelics temporarily make the old model less certain. The plasticity research adds that they temporarily make the neural system more capable of encoding a replacement. Chained together, with each link carrying its own evidence grade:
Reduced within-network integrity, increased cross-network communication, massive acute connectivity disruption.24
The claim that this window makes psychotherapy or environmental learning substantially more effective in people is surprisingly untested.
If the model is even roughly right
A mechanism is only worth having if it changes what you do on a Tuesday afternoon. Here is what this framework actually implies — for the person in the chair, and for the person across from them.
- Insight isn't the mechanism; experience is. Understanding where a pattern came from is useful scaffolding, but the model updates on lived evidence. Expect therapy to feel like doing something uncomfortable, not just explaining it.
- Once is not enough, and that isn't failure. A prior built from thousands of repetitions doesn't revise on one contradicting experience. Repetition is the mechanism, not a sign it isn't working.
- The stuck feeling has a reason. A rigid belief is usually a model that was once adaptive and still delivers predictability. Treating it as stupidity — yours or anyone's — misreads what it's doing.
- Triggers are information. A pattern that's activated is a pattern that can be edited. The moment it flares in a safe setting is the useful moment, not the wasted one.
- Notice the discounting. "They're just being nice." "They'll leave eventually." That reflex is the prior defending itself, and catching it in real time is a skill worth building.
- Activate, then violate. Every mechanism in Part IV needs the old pattern alive in the room at the moment the disconfirming experience arrives. Discussing it cold does far less.
- Expect to be tested. Neediness, provocation, withdrawal, boundary-pushing — control-mastery reads these as experiments, not resistance. How you respond is the intervention.16,17
- Maximize expectancy violation. In exposure work, the size of the gap between predicted and actual outcome predicts the strength of new learning better than how long the fear lasts.11
- Method matters less than process. If the mechanism is shared, arguing about school allegiance is largely arguing about delivery route. Pick what generates the most salient prediction error for this person.
- In high-plasticity states, restraint is technique. When priors soften and social uptake rises, a confident interpretation can install rather than reveal. "Notice what comes up" is not vagueness — it's mechanism-aware practice.28,29
- Treat the post-session period as part of the intervention. If the window is real, integration isn't aftercare. It's when the learning gets written.27
Where the science actually stands
"Predictive processing explains psychotherapy" is not an established fact. It is an attractive, testable integration — and the authors themselves say so.
Li and colleagues are explicit that they are proposing a conceptual integration that generates testable hypotheses, not reporting an experimentally established mechanism.2 That qualifier can get lost in an opinion-page treatment. Here is the claim ladder, graded rung by rung:
And the psychedelic layer deserves its own ladder, because the confidence levels differ sharply from link to link:
Two more honest cautions. First, the common-factors literature reminds us that alliance, empathy, expectations, and a credible ritual account for much of therapy's effect regardless of mechanism theory18 — predictive processing may eventually explain those factors, but it hasn't yet. Second, a framework flexible enough to redescribe every therapy risks explaining everything and predicting nothing; the value of this proposal will be decided by the specific, falsifiable predictions it generates.
And one clinical north star survives every theoretical fashion, so we'll end on it: the goal of therapy is not only to correct old predictions inside the relationship — it is to help the client keep self-correcting after therapy is over. A well-updated model, and the learned skill of updating.