Karl Friston's Free Energy Principle is one of the most ambitious frameworks in modern neuroscience. It is also, on close inspection, a description dressed as a law — one that survives every counter-example by absorbing it. This is a structural audit: the strongest case for the theory, then four faults, each tested against the evidence, one at a time.
An honest audit begins by putting the target at its strongest. The Free Energy Principle is not a fringe idea, and pretending otherwise would only build a straw man worth nothing.
Its core claim is genuinely elegant. Any system that persists over time must keep itself within a narrow band of viable states — and such a system can be described as if it minimises the discrepancy between what it predicts and what it senses. From that single premise, Friston derives perception, action, learning, and attention as facets of one underlying process. It reaches across thermodynamics, Bayesian inference, and biology and ties them to one piece of mathematics. It has produced real, testable models in computational psychiatry. And it takes seriously the deepest question in biology: how anything alive resists dissolving into disorder.
We grant all of it. The framework is mathematically serious, genuinely generative, and it asks the right question. The critique that follows is not that FEP is unserious — it is that a framework this powerful invites one specific failure. It can be tuned to explain any outcome after the fact, and where its central mathematics has been checked line by line, crucial steps did not hold. Those are the claims we now test.
The collapse of desire into belief
The most load-bearing move in Active Inference is quiet, and almost everyone waves it through: it encodes what a system wants as a prediction the system holds with very high confidence. A goal is modelled as a prior belief that a preferred state will occur. The organism does not act to reach the goal — it acts to confirm its own optimistic forecast.
This forces two categorically different things into one variable. In decision theory, philosophy of mind, and cybernetics, the split between what a system wants and what it expects is foundational. You can want something you believe is nearly impossible — a miracle recovery. You can confidently expect something you desperately do not want — the wall arriving during a car crash. FEP dissolves that distinction by fiat.
The consequence is not cosmetic. If a system's only imperative is to reduce prediction error, then a strong desire that reality refuses to satisfy generates a large, sustained error — and when the body cannot act to close it, the only mathematically available move is to update the belief. To stop wanting what cannot be had. Under strict FEP, chronic frustration, defiance in the face of hopelessness, and resistance to unavoidable pain become difficult to represent at all: the theory says the intention should simply decay to restore zero surprise.
The evidence runs the other way. Recording from the parietal cortex of a person with tetraplegia as she attempted individual finger movements, researchers found the neural representational structure was not only intact but matched the temporal profile of an able-bodied forward model — stable across ten sessions even when that stability actively hurt task performance [eLife 2022]. The intention did not decay to erase surprise. It held. That is the signature of an autonomous reference signal — a set-point the system defends — not a probability distribution obediently updating toward the incoming data.
Category error. FEP forces the will to obey the statistics of probability. The neurobiology shows goal representations that persist independently of whether they are ever confirmed — the exact opposite of a belief minimising its own error.
What the theory says should be paradise
Take the Free Energy Principle at its word. If a living system exists to minimise surprise, then the ideal environment is one with no surprise at all: silent, unchanging, perfectly predictable. A sealed, sterile room where nothing ever deviates from expectation should be, in FEP's own terms, a homeostatic paradise.
Biology says the exact opposite.
Strip a nervous system of varying input — sensory-deprivation studies, prolonged isolation, solitary confinement — and it does not settle into blissful zero-error. It destabilises, fast. Deprived of external signal, the brain amplifies its own internal activity and manufactures experience to fill the void [a]: hallucinations, dissociation, the collapse of time-sense, symptoms documented across decades of isolation research [b]. A brain denied something to control does not accept the silence. It generates its own signal rather than go quiet.
I put this parallel to Kent McClelland, Professor Emeritus of Sociology at Grinnell College, framing the maximum-control cell as a real-world instance of Friston's Dark Room: an environment engineered for near-total predictability, which the theory should treat as optimal and which every prisoner experiences as punishment. His reply:
Brains are control engines, and a brain deprived of any opportunity to keep things in control stops learning and starts to atrophy.
McClelland, who spent several years volunteering in a medium-security prison in Iowa, noted the same thing from the sociological side: that the crushing predictability of the environment was itself a core part of the punishment — the reason variety and novel ideas mattered so much to the men inside. The parallel is exact. An FEP-optimal environment — zero surprise, perfect prediction — is, in practice, one of the most destructive conditions a human nervous system can be placed in.
FEP's defenders answer this with a patch: Expected Free Energy, an added "epistemic value" term that pushes the agent to explore. But notice what that concedes. Exploration does not fall out of the original principle — it has to be bolted on as an extra term precisely because the base theory, projected forward honestly, discourages the very behaviour biology requires. When a theory needs an epicycle to stop predicting the wrong outcome, the epicycle is the tell.
The base axiom predicts the zero-surprise environment is optimal; reality demonstrates it is pathological. "Expected Free Energy" is a correction added after the fact — an admission, in the theory's own machinery, that minimising surprise gives the wrong answer.
The Popperian audit
Here is the move that makes FEP nearly impossible to pin down. Its proponents state that the principle cannot be proven or disproven — that trying to falsify it is itself a category error — because it is a mathematical truth on the order of the calculus or Hamilton's principle of least action. Any system with an attractor and a statistical boundary must, by definition, appear to minimise free energy.
And yet the same framework is aggressively applied to explain highly specific empirical phenomena: dopamine release, decision optimism, autism, schizophrenic hallucination, the evolution of culture, aesthetic experience. This is the shield. When a physiological counter-example lands, FEP retreats into mathematics — "it's only a tautology, the system is merely in its attractor." When the mathematics is challenged, it advances back into biology — "it's a powerful engine for explaining neural pathology."
A theory cannot be both at once. An unfalsifiable tautology carries no empirical content by definition — and a detailed causal mechanism with real predictive power is, by definition, falsifiable. "Why did organisms survive? Because they keep free energy low. How do we know they keep free energy low? Because they survived." That circle is airtight and says nothing. The abstract principle may be untouchable; the specific claims of Active Inference, sold as if they flowed from that principle with the same necessity, do not inherit its immunity.
The framework claims the invulnerability of a mathematical identity and the explanatory power of an empirical mechanism, and switches between them under pressure. That is not a demarcation problem at the edges — it is the load-bearing structure of the defence.
Biehl, Pollock & Kanai, Entropy (2021)
The dense formalism — Markov blankets, non-equilibrium steady states, variational densities — functions in practice as armour. It makes the theory expensive to attack on its own terms. So the decisive test is not rhetorical. It is what happens when specialists sit down and check the proofs.
They did. In a peer-reviewed technical critique, Martin Biehl, Felix Pollock, and Ryota Kanai worked through the core argument and reported three findings that matter here. Stated precisely, because precision is the whole point:
Be careful with what this does and does not say — overstating it would repeat Friston's own sin. It does not mean "all of FEP is false." It means the specific, foundational derivation that lets Friston claim any system with a Markov blanket appears to perform Bayesian inference does not go through as originally written, and holds only inside a narrow class of systems under assumptions that were never made explicit [Entropy 2021, 23(3):293]. The universal generalisation — the part that makes FEP a "theory of everything alive" — is exactly the part that fails to generalise.
The claim of mathematical inevitability does not survive line-by-line checking. The lemma is either wrong or trivial, the blanket is defined inconsistently, and the leap to "every living thing does Bayesian inference" rests on assumptions that were smuggled in.
Free Energy Principle vs Perceptual Control Theory
Every disturbance FEP struggles with — sudden, novel, un-modelled — a control loop handles without breaking a sweat, and without a single probability distribution. Perceptual Control Theory, developed by William T. Powers, describes the same organism-environment stabilisation FEP is reaching for, using a closed negative-feedback loop instead of a predictive open loop wearing a closed-loop costume.
Balance an inverted pole on your palm. You correct for gusts, twitches, and bumps you never anticipated and never modelled — you are simply protecting the perception of "upright." The Test for the Controlled Variable makes this measurable: apply a disturbance, and if the system compensates to hold a specific perception steady, that perception is what it controls. No variational estimation. No marginal likelihood. Just a loop defending a set-point. FEP models this same reflex as a cascade of precision-weighted predictions propagating down a hierarchy — an ornate and redundant computational load for something a thermostat does with a bimetal strip.
The charge against FEP is that it evades falsification. It would be hollow to level that charge from behind the same shield. So here is the exposure the theory never offers — the specific findings that would break the audit above.
Each of these is a concrete, checkable condition. That is the difference between a critique and a creed — and, not incidentally, the difference this whole audit is about.
Status: open research essay · not peer-reviewed · a formal preprint version is planned for Zenodo.