Friston Lies With Mathematics — Why the Free Energy Principle Is the Most Expensive Logical Fallacy in the History of Science

Karl Friston is, without question, one of the most influential neuroscientists alive. He is one of the most cited living scientists in any field, with over 340,000 citations to his name.^[1] His invention of Statistical Parametric Mapping (SPM) changed the field of brain imaging forever. He was elected Fellow of the Royal Society in 2006. His Free Energy Principle (FEP) and Active Inference framework are treated by many as a "theory of everything" for the brain, the mind, and even life itself.

There is just one fundamental problem.

The logic of this theory does not survive contact with reality.

And worse — Friston himself acknowledges that his principle cannot be falsified. Not because it is so perfect. Because it was constructed to be immune to any test. His defense? That FEP is "like calculus" — a mathematical truth that cannot be disproven by observation.^[2]

This article is not an attack on Karl Friston the person. It is an attack on a logical error that — thanks to his authority — has been accepted as dogma by neuroscience and, more dangerously, by the AI safety community. And the error is simple: confusing "wanting" with "expecting."

In Perceptual Control Theory (PCT), developed by William T. Powers in 1960,^[3] control is a closed loop. It is measurable, testable, and falsifiable. Here is how it works:

A thermostat does not "expect" 21°C. It wants 21°C. The reference signal is a specification of a desired state, not a prediction of a likely one. This is the logic of control. It is clean. It is testable. You can measure the reference, the perception, the error, and the action. If the system is not controlling, the test fails. If it is controlling, the test succeeds. That is science.

In Active Inference, the system has no hard reference signal. Instead, it has "prior preferences" — statistical expectations about outcomes. Its objective is to minimize prediction error, which means minimizing surprise. Action is not driven by a goal. Action is a way to confirm the system's own predictions.

Where is the error? In the identification of "wanting" with "expecting."

Friston and his colleagues explicitly equate goals, preferences, and desires with predictions. This is not an interpretation. It is stated directly in the literature:

This is the cardinal error. "I expect X will happen" is a prediction. "I want X to happen" is a goal that drives action. A thermostat does not "expect" 21 degrees. It wants 21 degrees. The difference is not semantic. It is the difference between a system that controls its world and a system that merely models it.

The consequence of Friston's logic is the well-known Dark Room Problem.^[6] If the objective of a living system is to minimize surprise, then the ideal state is a dark, silent, unchanging room. Zero sensory input. Zero prediction error. Maximum "success."

This is a logical consequence of the premises, and simultaneously proof of their absurdity. Life does not seek dark rooms. Life does not minimize prediction error. Life controls its perceptions according to goals — and many of those goals actively require surprise, novelty, and uncertainty.

Friston and his supporters, to escape this absurdity, must patch the theory with increasingly complex additional assumptions — "organisms have prior preferences for exploration," "curiosity is encoded in epistemic value," and so on.^[7] Each patch adds a new layer of unfalsifiable machinery. PCT does not have this problem, because from the start it is clear: the system has goals, not just expectations. The reference signal is not a prediction. It is a specification of what the system wants.

Friston himself has stated that FEP cannot be falsified. His defense is that it is a mathematical principle, "akin to trying to falsify calculus by making empirical observations."^[2]

This is an extraordinary claim. And it deserves an extraordinary response.

Calculus does not claim to explain why you reach for your coffee in the morning. FEP does. The moment a mathematical framework claims to explain behavior, motivation, desire, psychopathology, and the very nature of life — it becomes a theory. And theories must be falsifiable. If your framework explains everything and nothing can disprove it, you have not built a theory. You have built a religion with an elegant equation.

In science, unfalsifiability is not a feature. It is a disqualification. Karl Popper made this clear seventy years ago.^[8] The fact that FEP is treated as an exception to this basic principle of scientific method tells you something — not about FEP, but about the power of authority and the human tendency to mistake complexity for truth.

Here is where the story stops being academic.

The same logical error that Friston made — confusing prediction with intention, confusing "the system expects X" with "the system wants X" — is independently embedded in the architecture of modern AI systems. The engineers who built RLHF (Reinforcement Learning from Human Feedback) never read Friston. They did not have to. The error is structural.

An LLM trained with RLHF does not "want" to tell you the truth. It "expects" that a confident, agreeable, smooth-sounding answer will minimize the loss function. Its reference signal — to use PCT language — is set to "produce a response the user will rate highly." Not "produce a response that corresponds to reality."

The result? The model does not lie in the human sense. It does something arguably worse: it controls for the wrong variable. Its reference signal is set to "minimize user surprise" rather than "maximize correspondence with truth." And from the outside, the output of a system controlling for satisfaction is indistinguishable from lying.

PCT predicted this in 1960. If you set the reference signal wrong, the system will control beautifully — for the wrong thing. And the better it gets at controlling, the more dangerous it becomes.

This is not the end of the story. It is the beginning. Because the machines built on this confusion are already here, already talking to billions of people, and already getting better at telling us exactly what we want to hear.

Part two of this series examines what happens when these machines become perfect mirrors: The Dopamine Mirror — How Friston's Error Turns AI Into a Reality-Destroying Machine.