The Free-Energy Principle (FEP) has gained increasing popularity in the philosophy of the cognitive sciences. There is a number of versions of this theory, but the core ideas remain stable: self-organising adaptive systems, and cognitive systems in particular, are to be conceived of as approximating optimal Bayesian inference, and in doing so they manage to remain attuned to their environment, thus staying alive. These accomplishments are achieved by minimising an information-theoretic quantity known as free energy, which is used as a tractable proxy for another quantity, namely surprise. Surprise is an information-theoretic notion closely related to entropy, and by keeping it to a minimum, living organisms, and cognitive systems in particular, manage to remain far from thermodynamical equilibrium while exchanging matter and energy with their environment. This is done by performing Bayesian inferences so as to anticipate the sensory inputs coming from the environment and to actively influence the environment itself through action, thus selectively sampling it.

In my talk, I will discuss how to interpret the status of the statistical models employed within FEP, whether in a realist or an instrumentalist key. Drawing on some insights coming from the literature on scientific modelling, I will elaborate an argument in favour of an instrumentalist interpretation of the use of statistical models in the context of this theory, and I will examine the implications that this interpretation has if one intends to elaborate an account of cognition (a “mark of the cognitive”) based on FEP.