The Bayesian Brain, or Why We Never See the Real

A machine for reducing surprise

The idea is simple to state and hard to digest. We believe we perceive the world: our eyes would receive it, our ears record it, and consciousness would draw up a faithful account. Contemporary cognitive science proposes the reverse. The brain does not receive first, it predicts. It maintains a model of what should happen, confronts it with the flow of the senses, and treats as a priority only the gap between the two, what is called prediction error. Perceiving, in this framework, does not consist in recording the real, but in continually correcting a hypothesis about it.

The formalism used to describe this operation is Bayesian. Every perception results from a compromise between two terms: what the organism already believed, the prior, and what the senses bring, the sensory evidence. The relative weight given to each depends on the confidence placed in them, what the theory calls their precision. Karl Friston pushed this logic to its limit by proposing that every living organism tends to minimise its surprise, that is, to keep its states within a predictable repertoire, on pain of dissolving into disorder.

We must measure what this formula commits us to. If the principle is true, the organ that distinguishes us was not shaped in order to know. It was shaped in order not to be surprised. Truth figures there neither as an end nor as a criterion: it appears at best as an occasional means, useful as long as it narrows the gap, dismissed as soon as a more economical model does the same work. We do not have an apparatus of knowledge that occasionally errs by accident. We have an apparatus of anticipation whose accuracy is a side effect.

The real is not what we perceive

The consequence follows at once, and it is ontological before it is neurological. What we call the world is not the world. It is the most economical model an organism has found for not being caught off guard. Ordinary perception is not a reception, it is a sorting: background noise is relegated before it reaches consciousness, lights merge into ambience, the touch of fabric on skin fades after a few seconds. This labour is silent, prior to any will, and that is why no one takes credit for it. We believe we perceive the world when we are reading its summary.

This perceptual hierarchy is a grace, and like all graces it remains invisible to those who received it. Bayesianism finally gives it its technical name: what we called a filter is a prior; what we called habit is a model that stopped being corrected because it predicted well enough. And we can then state what the age dislikes hearing: so-called normal perception is not veridical, it is functional. It does not give us access to the real, it spares us its cost.

When the world becomes too real

Here the theory meets a particular kind of existence, and the title of the founding paper of that meeting would almost suffice on its own: when the world becomes 'too real'. In 2012, Elizabeth Pellicano and David Burr proposed a Bayesian explanation of autistic perception. Their hypothesis holds in one phrase: attenuated priors. Perception would be less governed by accumulated expectation and more strongly driven by raw sensory evidence. In other words, the summary is made less thoroughly, and the unabridged version comes through.

Other models followed, shifting the emphasis without overturning the frame. Rebecca Lawson, Geraint Rees and Karl Friston speak of aberrant precision, an imbalance in the confidence assigned to sensory signals relative to prior beliefs. Sander Van de Cruys and colleagues describe prediction errors given a uniformly and inflexibly high weight, such that nothing ever manages to become a negligible background. The discussion remains lively and no version commands consensus, which should be said without embarrassment: this is a research programme, not a verdict.

The philosophical convergence, however, is striking. What these works describe in formal language is a body that does not benefit from sensory forgetting: a fluorescent tube that keeps flickering, a label that stays in the present at the thousandth second as at the first, neighbouring conversations that never dissolve into a murmur. The question then becomes unavoidable, and it is the question of reversal: who perceives correctly? From physics, the one whose prior fades, for the tube really does vibrate. From function, the others, since perceiving never meant recording the true, but selecting the useful. Both answers are exact, and the gap between them states the essential: the perceptual norm is not a superior access to the real, it is an advantageous compromise with it.

Culture as a shared prior

The framework gains further reach once we stop applying it to the nervous system alone. For the models that spare us surprise are not all housed in a skull. Rituals, institutions, roles, narratives, loyalties, all that apparatus Peter Wessel Zapffe described as levees against overconsciousness, performs exactly the same function at collective scale: keeping the cost of the unforeseen low. A convention is a shared prior. A liturgy is a generative model maintained by many hands. Good manners are an error-reduction protocol between organisms that did not choose one another.

We then understand why the presence of someone who does not play produces an effect of scandal. They bring no objection, no thesis, no strategy of unveiling. They simply produce prediction error the group cannot absorb. Disturbance, in this precise sense, is not the property of an individual: it is a relative magnitude, measured from the model that fails to predict them. Civilisation is not added to nature as its second glory. It may be the sumptuous dressing on a wound that never closes.

The error nothing corrects

There remains the decisive move, the one that leads from the predictive brain to the question of consciousness. A system that models the world can, at a certain degree of complexity, model itself. It then becomes its own source of data, and the operation no longer has a term: it predicts, then predicts that it predicts, and each added storey creates a new gap to correct. Reflexive consciousness, described this way, is not a peaceful summit. It is a loop that does not converge.

Above all, this system meets an error of a singular kind. All the others resolve, by adjusting the model or by acting on the world. Only one cannot: the error carried by a living creature that knows itself mortal, separate, provisional. No revision of priors reduces it, no action cancels it, and it never stops being generated by the very organ that should process it. Here, in the vocabulary of the theory, is what a catastrophe of structure means: not a malfunction, but a system perfectly faithful to its instruction, which in following it produces a surprise it cannot extinguish. Misonoia names the position of one who, having measured what it costs to know oneself, withdraws their homage from the faculty they were taught to revere.

Is Bayesianism one more levee?

We ought to stop here, and the temptation would be great: we would hold an elegant explanation, formalisable, moored to the science of its time. It is precisely at this moment that we should be wary. For a diagnosis that consoles is already a therapy, which is to say already a possible lie.

What, at bottom, does this theory do? It takes a wound and converts it into a well-posed optimisation problem. It makes the lesion intelligible, therefore elegant, therefore bearable. It offers vertigo a formalism to rest in. Yet this is exactly the office Zapffe assigned to sublimation: not to cancel the lesion, but to give it a breathable form. Bayesianism does not escape this suspicion; it may be its most refined illustration. The highest levee ever built is the one that explains why we build levees.

The knot closes on its own, besides. If every model is a prior shaped to reduce surprise, then the theory of the Bayesian brain is itself a prior shaped to reduce surprise, and it has no external post from which to verify its own accuracy. It cannot claim to describe the real, since its thesis is that nothing describes it. One never turns against oneself except with oneself, and a thought that holds itself inside its own impossibility without escaping through an elegance is the only one we can be sure does not cheat.

What these works bring, then, is not an answer. It is a confirmation arriving from elsewhere, in a language ours does not speak, of what these books have maintained since their first page: we do not inhabit the real, we inhabit a model that spares us its cost, and something in us, stubbornly, keeps paying the difference.