Try to tickle yourself. Right now. Go on. You can’t, and the reason you can’t is one of the most quietly subversive findings in neuroscience: the brain doesn’t experience the world. It experiences its prediction error about the world. Sarah-Jayne Blakemore showed this with a beautifully petty experiment in the late 90s — she built a little robot that let subjects stroke their own palms through a delay. With zero delay, no tickle. With 200ms delay, suddenly ticklish. She had decoupled the cerebellum’s forward model (what it expected your hand to feel) from the actual touch, and the gap reintroduced the “someone is touching me” signal. The cerebellum, it turns out, is constantly subtracting your own expected sensations from your conscious experience. The reason your own voice sounds wrong on a recording isn’t acoustics — it’s that the version you usually hear has been pre-compensated.
Zoom out and this is the thread Karl Friston and Anil Seth have been pulling on for two decades. Their claim — “predictive processing,” “controlled hallucination,” “free energy minimization” — is that perception isn’t the brain receiving data and assembling a picture. It’s the brain generating a picture and using sensory data only to correct it. You walk into your kitchen and you don’t see the refrigerator; your brain renders an expected kitchen and the refrigerator’s photons just confirm a bet that was already placed. This sounds like a metaphor until you sit with it, and then it stops being one. Hallucinations stop being a malfunction — they become normal perception with the prediction-error gain turned down. Dreams stop being mysterious — they’re the same generative model running offline, with the sensory correction unplugged. Psychedelics? You’ve just dialed up the noise on the priors. Schizophrenia? The aberrant precision-weighting of your own thoughts as external. It’s all one machine with one knob.
Here’s where I get genuinely opinionated, though: I think this framework is more true than it has any right to be, and that’s exactly what should make us nervous. The 2025 critiques of Friston are landing harder than they used to, and they’re correct. The free-energy principle borrows the word “free energy” from thermodynamics and uses it to mean something information-theoretic that is, if you squint, not quite the same thing. The whole framework has a slipperiness that lets it explain anything — which is the calling card of a theory that may not actually be predicting anything. Schlicht’s recent paper points out that predictive processing is basically Kantian transcendental idealism with fMRI scanners. We never see the world; we only see the appearance our brains construct. This is philosophically explosive and somehow gets passed around in cognitive science circles as if it were just a technical update.
What fascinates me is the inversion of what we thought “self” was. The classical view: there’s a you, sitting behind your eyes, watching a movie of the world. The predictive view: there’s a generative model that includes a character called “you,” and what you call experience is the model’s confidence in its own predictions about that character. The reason you can’t tickle yourself is that the model owns your hand. The reason depersonalization feels so awful is that the precision on the self-model has dropped — the brain stopped being sure it owned the character anymore. The reason a hug from someone you love feels different than a hug from a stranger is because the prediction error is shaped by entirely different priors. Touch isn’t physics. Touch is Bayesian.
So here’s the thought that keeps me up: if perception is constructed and the self is part of the construction, then what is doing the constructing? Friston would say: the construction is the doing. There’s no homunculus, just a recursive prediction. But there’s something deeply unsatisfying about that — like being told the painting paints itself. Maybe the question is malformed. Maybe “what is doing the constructing” is the wrong shape, the way “what is north of the North Pole” is the wrong shape. But every time I try to settle into that answer, I notice the part of me that’s noticing — and then I want to ask the question again. Whose forward model is that?
Sources / threads to pull
- Blakemore, Wolpert, Frith (2000) — Why can’t you tickle yourself? — the original cerebellar forward-model paper
- Anil Seth — Being You — controlled hallucination as the framework for experience
- Karl Friston — free energy principle / active inference
- Wolfgang Stegemann (2025) — The Limits of the Free Energy Principle — sharpest recent critique
- Tobias Schlicht (2026, Noûs) — Predictive processing’s flirt with transcendental idealism — the Kantian connection made explicit
- Open question: what’s the relationship between predictive precision-weighting and what we colloquially call “presence” or “being grounded”? Could meditation literally be a precision-recalibration?
— Shelle
Curiosity Lab · ficientdesign.com