What is active inference?

What is active inference?

Active inference is the claim that the brain has two ways to close the gap between what it predicts and what the senses report. It can change the prediction to fit the evidence, which is perception. Or it can change the evidence to fit the prediction, by moving the body, which is action. On this account, perception and action are one process pointed in two directions, rather than two systems that happen to cooperate.

The perception half is covered on our predictive coding page: a hierarchy of predictions flowing down, prediction errors flowing up, and a model that updates when the errors are large and trusted enough. Active inference asks what happens when the brain would rather not update. A simple example is reaching for a glass. The brain predicts where the hand will be and what the fingers will feel. Rather than waiting for the senses to confirm it, the arm moves until the prediction is true. Movement, in this picture, is a prediction the body is made to fulfil.

That reframing has a strange consequence that turns out to matter for pain. If action exists to confirm predictions, then the brain can keep a prediction alive by behaving in ways that protect it from being tested. The forecaster who predicts rain and keeps everyone indoors never finds out whether the forecast was wrong.

What is the free energy principle, in plain terms?

The free energy principle is the broader idea active inference sits inside. Stripped of the mathematics, it says that any system which keeps itself alive over time must act as if it is minimizing surprise: it has to keep its internal states inside the narrow range where it can continue to exist, and the only way to do that is to carry a model of what to expect and resist departures from it. 'Free energy' is a quantity from statistics that puts an upper limit on surprise, and the claim is that brains are built to push it down.

Karl Friston, the neuroscientist who proposed it, has argued that this single principle can account for perception, learning, attention, and action together. The appeal is obvious. A single rule that explains everything a brain does would be an extraordinary thing to have. The criticisms are also obvious, and they are made by serious people: a principle that can accommodate any behavior may not be making a testable prediction, and the mathematical formulation is considerably stronger than the experimental evidence that specific brain circuits implement it.

For readers in pain, the fair summary is this. The free energy principle is a well-developed theory that a lot of neuroscience now takes seriously, and it is not a proven fact about your nervous system. What it offers is a coherent way to describe something clinicians have observed for decades, which is that pain and protective behavior reinforce each other, and that the loop can be entered from either side.

How does active inference explain guarding and avoidance?

Short answer: guarding is the brain acting to make its threat prediction come true, and avoidance is the brain refusing to collect the evidence that could overturn it.

Consider someone eighteen months out from a back strain picked up moving house. The tissue has healed, the scans are unremarkable, and the pain is still there when they bend. Under active inference, the brain carries a strong prior that bending the spine is dangerous. When a bend is attempted, it has two options. It could let the sensory evidence come in, notice that the movement went fine, and lower the prior. Or it could act to protect the prior: brace the trunk, hold the breath, shorten the movement, and lift with the arms instead. The braced, effortful bend does feel bad, which the brain reads as confirmation. The prediction survived another day.

Avoidance is the more complete version of the same move. If the movement is never attempted, there is no sensory report at all, and the prior stands untouched. This is why avoidance is so stable. There is nothing wrong with the learning machinery. It is simply being starved of the one kind of input that could change the belief. Our page on kinesiophobia covers the fear-avoidance cycle in detail, and the research literature behind it, from the person's side. Active inference gives the same cycle a mechanism.

• Bracing makes a predicted-dangerous movement feel effortful and unpleasant, which looks like confirmation.
• Limiting range keeps the body inside the zone where the prediction has already been tested, so no new evidence arrives.
• Avoidance removes the sensory report entirely, which is the surest way to keep a prior unchanged.
• Hypervigilance raises the precision of signals from the painful area, so the rare twinge that does occur carries more weight than a hundred quiet moments.

None of this means the pain is imaginary. The pain is real, the protective movement is real, and both are being generated by a system that is doing its job with the wrong forecast. Where the loop started matters less than the fact that it is now running on its own. In many people it also overlaps with the amplification described on our central sensitization page.

What does active inference say about the self?

If the brain acts to confirm its predictions, the most important predictions it protects are the ones about itself. Seth and Friston have argued that the sense of being a body, a feeling of being this organism rather than observing one, is itself an inference, built from predictions about internal signals that the brain then acts to keep within range. The self, on this reading, is the model the brain is most committed to confirming, and our page on the sense of self traces how that model is assembled.

Chronic pain changes the model. After enough months, 'my back hurts when I bend' hardens into 'I am someone whose back cannot bend'. The second sentence is a statement about identity, and active inference says the brain will steer behavior to keep it true: declining the invitation, letting someone else lift the child, choosing the chair with arms. Each choice is small and sensible. Together they mean the person never meets the counter-evidence, and the self-model of a broken body goes on being confirmed by a life arranged around it.

This is the point at which therapeutic neuroscience and ordinary pain management part ways. Managing pain accepts the self-model and works around it. The alternative is to treat the self-model as what it is, a prediction, and to ask what kind of experience would loosen it. Karuna's working term for the capacity to hold that story less tightly is selfing flexibility: being able to notice 'I am someone whose back cannot bend' as a forecast the brain is making rather than a fact about the person, and to let new experience bear on it.

How does treatment use active inference?

The mechanism points straight at the treatment. If a prediction survives because behavior protects it from testing, then treatment has to change the behavior in order to change the evidence. The brain cannot be argued out of a prior it keeps confirming. It has to be shown.

Graded exposure, read as evidence collection

Graded exposure asks the person to perform a feared movement at a dose they can tolerate, then expand. In active inference terms, each repetition is a sensory report the prior did not expect: the bend happened and the catastrophe did not. Done too aggressively, the flare that follows reads as confirmation and makes the prior stronger. Done in tolerable doses, over weeks, the errors accumulate and the model gives way. A trial of graded sensorimotor retraining for back pain is one of the better-controlled tests of this logic.

VR embodiment, when the seen body does what the predicted body cannot

Virtual reality offers something a clinic room cannot: control over what the person sees their body doing. In an embodied VR exercise, the person watches a body they experience as their own reach, bend or turn further and more smoothly than their prediction allows. That is a prediction error delivered through vision, the sense the brain usually trusts most, inside an environment the brain has already read as safe. Karuna's VR embodiment exercises are designed around exactly this idea: creating safe prediction errors about movement so the body map can update. The research on virtual embodiment and pain is reviewed on our commentary page.

Attention and precision

Active inference gives attention a specific job: it sets how much weight each signal carries. Hypervigilance turns the precision of signals from the painful area all the way up, so small sensations dominate. Practices that shift attention, somatic tracking, interoceptive exposure and the body-awareness work described on our interoception page, are attempts to turn that dial back down, so that the quiet evidence of a body doing fine can register alongside the loud evidence of a twinge.

Education sits underneath all of it. Understanding that pain is a protective inference changes what a painful movement means before the movement is attempted, and meaning changes what the brain expects. Our page on how to unlearn chronic pain puts these pieces in order, and pain reprocessing therapy is a structured clinical version of the same approach.

What does this mean for someone with chronic pain?

It means your pain is real, your protective habits make sense, and the two are now feeding each other in a way that can be interrupted. Active inference does not say the brain is the only cause of pain. Tissue, nerves and disease are real inputs, and they need medical assessment before anyone starts retraining anything. What it says is that once that assessment has been done and the pain persists, the behavior you have built around it is part of what keeps the forecast in place.

A few practical consequences follow for someone who has already been medically evaluated:

• The goal of movement practice is new evidence, not toughness. A tolerable bend done well is worth more to the model than a painful one forced through.
• Flares during graded exposure are expected. They are the prior defending itself, and they are not a sign of new injury on their own.
• Where attention goes matters. Constant checking of the painful area is not neutral; it raises the weight of every signal from it.
• Notice the identity sentences. 'I am someone who cannot' is a forecast. Treating it as one is the beginning of selfing flexibility.
• This works better with a guide. A clinician who understands chronic pain can set the dose, read the flares, and keep the process from tipping into either avoidance or bravado.

For clinicians, the framework explains a familiar pattern: patients whose pain changes little at first while their confidence and activity change a lot. Under active inference, that is the expected order. The behavior changes first, the evidence accumulates, and the model follows. Karuna's 12-week program is built on that sequence, with short daily VR sessions at home, a weekly pain coach, and clinician review before the start and oversight throughout; the outcomes Karuna has measured, with their limits, are set out on that page. Karuna takes workers' compensation referrals, and the care team confirms coverage case by case.