Active Inference — Acting to Change Predictions

Have you ever wondered, halfway through some deliberate action, whether you were doing this thing because you wanted it — or because you were quietly trying to make the world confirm what you already believed? The predictive processing literature has a productive frame for that question, called active inference, and it is worth understanding.

In the predictive frame, the brain is running an ongoing forecast of what the next moment will contain. When forecast and data disagree, there are two ways to resolve the disagreement. The system can update the forecast to match the data — this is what we usually call learning. Or the system can act on the world so that the data matches the forecast. This second move is active inference (Friston et al., 2010) [Hypothesis]. The evidence level here is Hypothesis because active inference remains an evolving theoretical framework rather than a settled empirical fact, though it has been productive in generating testable predictions.

A concrete example makes the frame clearer. Imagine a nervous system that predicts, from its prior, that it is not the kind of person who is welcomed in a particular social group. Two paths are available. The system can gather sensory data — actual behaviour from the group — and update the prior if the data disagrees with the prediction. Alternatively, the system can act in ways that quietly make the prediction come true — showing up half-hearted, leaving early, not accepting invitations — so that the world produces data confirming the model. Both paths reduce prediction error. Only one of them updates the model.

What active inference makes visible is that motivated behaviour is not always about achieving goals in the ordinary sense. A significant portion of what a nervous system does moment to moment is subtly organising the world to keep its predictions confirmed. This is not conscious sabotage. It is what a prediction-minimising system will do by default, in the absence of some deliberate counter-move.

Why this matters practically is that it explains a phenomenon everyone has some version of: the situations we most complain about are also, oddly, the situations we keep recreating. On the active inference view, this is not mysterious. The nervous system has a strong prior about how these situations tend to go. Rather than sitting with the discomfort of prediction error long enough to update the prior, the system acts to reinstate the familiar situation, which reduces immediate prediction error at the cost of never actually updating the model.

Breaking this loop requires something specific. It requires being willing to enter situations where prediction error is likely, and then resisting the pull to act in ways that would restore the familiar outcome. That resistance is uncomfortable — the nervous system reads it as unresolved prediction error, which feels like tension, hesitation, wrongness. Staying inside that tension long enough for genuinely new data to arrive is what allows the prior to update rather than being confirmed once again.

This is one of the more useful frames for understanding why change is hard. It is not that the person lacks motivation. It is that the nervous system's default move, when facing prediction error, is to act to restore prediction rather than update the model. Recognising the default is the first step toward choosing differently.

For the underlying mechanism, see Prediction Error — The Currency of Learning. For where the priors driving this come from, Priors — The Brain's Working Assumptions is the companion piece.