← Writing

ADHD · AI

Your brain is a prediction machine — and that reframes ADHD

The intuitive model of the mind is a camera: the world comes in through the senses and the brain records it. The current best model is almost the reverse. The brain is a prediction engine that guesses the world first and uses the senses only to check the guess. Take that seriously and two things change — what attention actually is, and why "just try harder" is the wrong prescription for an ADHD-shaped mind.

Perception is a controlled guess

The framework is called predictive coding, or more broadly predictive processing. Its claim is that the brain does not sit and wait for sensory data. It runs a continuously updated model of the world and generates top-down predictions of what the senses should report next. What actually travels up from the eyes and ears is mostly not "the image" — it is the error: the difference between what the model predicted and what arrived. Perception is the brain's best guess, corrected by surprise.

Andy Clark summarised the shift in a 2013 paper whose title is the whole idea: "Whatever next?" The neuroscientist Anil Seth puts it more provocatively — perception is a "controlled hallucination," a prediction the brain holds until the world pushes back hard enough to revise it. This is not mysticism. It is the most economical way to build a fast agent out of slow, noisy neurons: transmitting only the surprising part of a signal is far cheaper than re-transmitting the whole thing, moment after moment.

Karl Friston generalised this into the free energy principle: a system that persists must act to minimise the long-run gap between its predictions and its inputs — it must, in effect, keep surprise low. You can quarrel with how universal that claim is, and critics do. But you do not need the grand version to use the local one.

Attention is precision, not effort

Here is the part that matters for anyone who has been told they have an attention problem. In predictive processing, attention is not a spotlight you point by willpower. It is the weight the brain assigns to prediction errors — what the literature calls precision. When the system decides a given error signal is trustworthy and important, it turns up its gain, and that channel dominates. When it decides an error is noise, it turns the gain down, and the signal is ignored no matter how "important" it is on paper.

Dopamine is one of the messengers that sets this gain. It carries, among other things, a reward-prediction error: the difference between the reward you expected and the reward you got. So the machinery that decides what is worth attending to and the machinery that decides what is worth wanting are the same machinery, tuned by the same chemical. That single fact reframes the whole conversation.

ADHD as a dysregulated prediction loop

The folk theory of ADHD is a moral one: a deficit of effort, a weak will. The predictive-processing reading is mechanical and, I think, both truer and kinder. If attention is precision-weighting tuned by dopamine, then an atypical dopamine system does not produce a person who won't focus. It produces a person whose gain control is miscalibrated — who assigns high precision to the wrong error signals and low precision to the ones a tax deadline says are important. The boring, predictable task generates little prediction error and little dopaminergic gain, so it is not that the ADHD mind refuses to do it; the signal never gets loud enough to win the competition for control.

This also explains the paradox that confuses observers: the same person who cannot start a routine form can disappear for nine hours into a hard problem. Hyperfocus is not the opposite of the deficit. It is the same precision system, locked onto a source that is generating a rich, fast stream of prediction error and reward. The knob is not stuck low. It is unstable, and it swings.

The problem with ADHD is not knowing what to do. It is doing what you know.Russell Barkley (paraphrased)

Russell Barkley's long argument is that ADHD is better described as a disorder of executive function and self-regulation than of attention as such — a problem of using knowledge to govern behaviour across time. Predictive processing gives that a substrate: regulating behaviour over time is exactly the job of keeping a stable model running against a noisy, tempting present.

The design consequence: lower the cost of prediction

If the loop is the problem, then the intervention is not more willpower applied to a badly tuned system. It is redesigning the environment so the system is easier to predict. This is the practical core, and it is where I spend real effort as a founder who runs on this kind of brain.

Concretely, lowering the cost of prediction means a few different moves:

  • Defaults over decisions. Every open choice is a live prediction problem the brain has to price. A default — same first task each morning, same tools, same commit ritual — removes the error signal entirely. The decision that never happens costs no precision.
  • Externalise the model. A second brain — a wiki, a written system — offloads the part of the world you would otherwise have to hold in an unstable head. You are not remembering; you are reading. The prediction is made once, in writing, and reused.
  • Make the right action the low-error path. Structure the environment so the correct next step is the one with the least friction and the most immediate, legible feedback. Do not fight the gain control — feed it. A visible progress bar generates the reward-prediction error the task itself withholds.
  • Ration novelty deliberately. Novelty is dopaminergically expensive and seductive. Treat it as a budget: a little, placed where it powers work, not scattered where it fragments attention.

None of this is a productivity hack in the LinkedIn sense. It is the opposite of self-optimisation theatre. It is accepting that the machine is what it is and building the room around it so the machine's cheap path and the right path are the same path.

A worked redesign: one impossible morning

The theory earns its keep only if it changes what you actually do, so here is the loop applied to a concrete failure most people with this wiring will recognise. The task is a tax form due in a week — low stimulation, no urgency yet, no reward until it is finished. The neurotypical advice is to schedule an hour and sit down. For a miscalibrated precision system, that hour arrives and the form generates almost no prediction error, so the gain never rises to meet it, and the hour is spent on something the loop found louder. The failure then gets stored as a moral fact about your character. It is not one. It is a predictable output of the machine described above.

Redesign the morning to lower the cost of prediction instead of raising the demand for willpower. First, remove the decision: the form is the first task, every relevant morning, before email exists, so the loop never has to price a choice between it and something more stimulating. Second, manufacture the missing error signal: break the form into a two-minute opening move — one field, one number — because a task you can start in two minutes generates a small completion reward the whole form withholds, and the loop will chase that. Third, borrow urgency the environment is not supplying: a fixed external deadline, a person expecting it, a timer that makes the abstract "next week" concrete enough to register as a real prediction error now rather than a vague one later.

None of that is discipline. Every move works on the environment — the defaults, the size of the first action, the visibility of the reward and the deadline — so that the cheap path for a miscalibrated loop and the correct path become the same path. When it works, it does not feel like heroic focus. It feels like the form was simply easier to do than to avoid, which is the entire goal: you did not out-muscle the system, you rebuilt the room until the system's laziness pointed at the right target.

The honest footnote is that this fails on the days the loop is swinging hard, and no environment design fully fixes that. The point is not perfection; it is moving the base rate. A morning engineered this way wins more often than a morning built on the assumption that you can summon precision on command — and "more often" compounds.

Where the model breaks

Predictive processing is a framework of unusual reach, and reach is exactly what should make you suspicious. Three limits are worth stating plainly.

One: the grand version may be unfalsifiable. Friston's free energy principle is elegant and arguably explains too much — a theory that can absorb any result is not obviously a theory. The local claims about prediction error and precision have empirical support; the cosmic claim that everything that exists is minimising free energy is closer to a framework for building theories than a theory you can test. Use the parts that pay rent.

Two: a mechanism is not a permission slip. "My dopamine system miscalibrates precision" explains a difficulty; it does not dissolve agency, and treating it as an excuse is its own trap. The value of the model is that it points at the environment as the lever — which is a place you can actually act — not that it relocates responsibility to your neurochemistry and stops there.

Three: ADHD is heterogeneous, and one elegant story flattens it. The precision-loop reading is a lens, not a diagnosis. Real presentations vary enormously; some of what gets called ADHD is anxiety, or sleep debt, or a mismatch between a person and a badly designed job. A model this satisfying is tempting to apply everywhere, and the honest move is to hold it as one good description among several, not the key to a person.

So stop asking a miscalibrated system to try harder. Ask instead what in the room is generating the wrong prediction errors — and change the room before you blame the brain.

Sources

  1. primaryAndy Clark, "Whatever next? Predictive brains, situated agents, and the future of cognitive science," Behavioral and Brain Sciences (2013); Surfing Uncertainty (2016).
  2. primaryKarl Friston, "The free-energy principle: a unified brain theory?" Nature Reviews Neuroscience (2010).
  3. primaryAnil Seth, Being You: A New Science of Consciousness (2021) — perception as "controlled hallucination."
  4. primaryRussell A. Barkley, ADHD and the Nature of Self-Control (1997) — executive function and self-regulation.
  5. secondaryLisa Feldman Barrett, How Emotions Are Made (2017) — prediction and interoception.
  6. secondaryWolfram Schultz et al., work on dopamine and reward-prediction error (1997 onward).