Please join us for the final installment of our interview series on AI and Agency, featuring Karl Friston in discussion with Brains editor Majid D. Beni!
Please join us for the final installment of our interview series on AI and Agency, featuring Karl Friston in discussion with Brains editor Majid D. Beni!
In this interview, Karl Friston presents the Free Energy Principle as a universal account of perception, action, and agency, grounded in a mathematically rigorous framework that purportedly unifies physics, biology, and cognition. While this vision is compelling in its scope, it founders on three foundational conceptual flaws that render it scientifically inadequate as a theory of mind or intelligent behavior.
First, the Free Energy Principle is explicitly framed as unfalsifiable. Friston himself compares it to Hamilton’s principle of stationary action, asserting that it is a mathematical truth rather than an empirical hypothesis. This immunizes the theory against empirical disconfirmation. Any observation can be retrofitted as an instance of free energy minimization, rendering the principle explanatorily empty. Scientific theories must generate risky, testable predictions; a framework that accommodates all possible outcomes explains none. The distinction Friston draws between an unfalsifiable core principle and falsifiable process models functions not as a clarification but as a deflection, shielding the theory from genuine scientific scrutiny.
Second, the theory rests on a category error between thermodynamic and information-theoretic notions of free energy. Although Friston acknowledges that the variational free energy used in his models is not physical energy, the persistent use of terms like entropy, surprise, and minimization invites a misleading analogy with physical systems. This conflates negative log-probability with joules, Shannon entropy with thermodynamic entropy, and statistical uncertainty with energetic dissipation. Such conceptual slippage generates an illusion of depth without explanatory substance. Biological systems are not closed physical systems minimizing thermodynamic potentials; they are open, historically contingent, and adaptive. Transferring principles from statistical mechanics to cognition by way of superficial formal parallels is not insight but metaphor masquerading as mechanism.
Third, the framework smugglingly reintroduces teleology under the guise of mechanistic explanation. By invoking prior preferences, goals, pragmatic value, and goal-directed behavior, Friston attributes intentionality to subpersonal processes without explaining how such goals arise. The system is described as if it wants to reduce surprise, as if it plans to fulfill preferences. This reproduces the homunculus problem: who or what within the system holds these preferences and executes these plans? The origin of value, affect, and purpose is presupposed rather than derived. In doing so, the theory fails to bridge the explanatory gap between mechanistic dynamics and intentional agency. It does not dissolve teleology; it relocates it, leaving the hard problems of motivation and meaning untouched.
These issues culminate in a broader confusion between description and explanation. The Free Energy Principle offers a formal redescription of adaptive behavior in Bayesian terms, but it does not identify causal mechanisms or intervene in the causal chain between neural activity and behavior. Its core claim that organisms survive because they minimize free energy is circular; it explains survival by appeal to a principle that is itself defined by survival. The mathematics obscures this tautology but does not resolve it. Consequently, the theory cannot generate counterfactual predictions or specify conditions under which it would fail—hallmarks of genuine scientific theories.
In contrast, empirically successful models in neuroscience—such as predictive coding in early visual processing—remain local, mechanistically specific, and testable. They do not claim universality, nor do they invoke teleological primitives. Their explanatory power stems from their bounded scope and empirical anchoring, not from mathematical grandeur.
Friston’s vision of artificial agents endowed with curiosity and genuine agency may be philosophically intriguing, but it remains speculative. Current artificial systems, including large language models, lack the necessary embodiment, generative world models, and intrinsic value structures. More critically, the Free Energy Principle provides no clear pathway for how such structures could emerge from physical processes alone.
In sum, the Free Energy Principle is not a scientific theory but a metaphysical narrative dressed in mathematical formalism. It excels in rhetorical unification but falters in mechanistic explanation. A more fruitful cognitive science demands theoretical modesty: precise models, empirical accountability, and a clear separation between descriptive formalism and causal explanation. Only then can we move beyond elegant tautologies toward genuine understanding.
These are relevant readings:
Friston, Karl. “Am I self-conscious?(Or does self-organization entail self-consciousness?).” Frontiers in psychology 9 (2018): 579.
Friston, Karl, Philipp Schwartenbeck, Thomas FitzGerald, Michael Moutoussis, Timothy Behrens, and Raymond J. Dolan. “The anatomy of choice: active inference and agency.” Frontiers in human neuroscience 7 (2013): 598.
Friston, Karl J., Tommaso Salvatori, Takuya Isomura, Alexander Tschantz, Alex Kiefer, Tim Verbelen, Magnus Koudahl et al. “Active inference and intentional behavior.” Neural computation 37, no. 4 (2025): 666-700.