The Bayesian brain hypothesis claims that the brain performs probabilistic inference — updating prior beliefs with sensory evidence to compute posterior estimates of the world. The framework explains visual illusions (rational decisions under uncertainty), multisensory integration (optimal weighting by reliability), and motor learning (prediction error minimization). The mathematics is elegant. The fits to data are good. The conceptual unification is seductive.

Mangalam (2025) argues the seduction is the problem. The hypothesis occupies an ambiguous territory between metaphor and mechanism. When challenged on biological plausibility — where are the probability distributions stored? how are priors updated in neural hardware? — proponents retreat to metaphor: the brain operates “as if” it were Bayesian. When challenged on metaphorical looseness — if it’s just “as if,” then it’s not a mechanistic claim — proponents reassert mechanism and point to neural implementations of prediction error.

The retreat is not accidental. It is the structural signature of an unfalsifiable theory. A framework flexible enough to accommodate any finding post hoc — reinterpret the data, adjust the prior, add a complexity term — has the explanatory power of a tautology. The Bayesian brain can explain why you perceive the illusion and why you don’t. It can explain why prediction errors arise and why they don’t. The theory that explains everything constrains nothing.

The through-claim is about the relationship between mathematical elegance and empirical content. Bayesian statistics is a powerful formalism. The brain processes information. The question is whether the formalism describes the mechanism or merely recapitulates the outcome. A thermostat behaves “as if” it computes temperature gradients. The analogy is useful; the mechanism is a bimetallic strip. The distance between “as if” and “is” is the entire content of the theory, and the Bayesian brain hypothesis systematically avoids committing to which side it occupies. The claim retreats exactly when it would become testable.