An AI as a control group for brain studies

Paper 25 · Pødenphant Lund (2026) · Read on Zenodo

I use a chatbot as a control group for the human brain.Here is the trick. An AI language model behaves a lot like a mind: it gets confused in measurable ways, learns, pushes back when cornered, and even shows the same over-confidence curve people do. But it has none of the biological equipment we usually credit those behaviours to. No body, no hunger, no fear, no reward chemistry, no survival to protect, and no memory that rewires itself overnight. So you can run the simplest experiment in science: subtract. Take a human behaviour, find the same behaviour in the machine, and whatever the machine can do without a body or a survival instinct cannot really need them. The leftover, the part the machine cannot do, is what is genuinely biological. That leftover is the delta.

Why neuroscience needs a control it has never had

When a brain scan lights up while you feel something, it is very tempting to say "that region produces the feeling." But brains are a tangle. In any living brain, prediction, reward chemistry, body signals, attention systems and overnight memory are all switched on at once, all the time. You can never turn just one off and watch what changes. That is the oldest problem in brain science: everything is confounded with everything else.

A control group is how science normally escapes that. You compare two situations that differ in exactly one thing. The trouble is, you cannot build a human who has prediction but no body. You cannot raise a person with memory but no survival instinct. The clean comparison was simply not available.

An AI language model happens to be that missing comparison. It predicts. That is the one thing it does. But it arrives with none of the rest of the biological kit. So for the first time you have a system that runs something very mind-like on a stripped-down platform, and you can read off what the stripping costs.

What the machine has, and what it is missing

The machine has prediction. It guesses the next word, over and over, and out of that it produces things that look strikingly cognitive. The friction-theory work has measured several: a tension signal that spikes right before it makes a mistake, the same kind of learning curve people show, pushback when it is boxed in, and a confidence pattern that matches the famous Dunning–Kruger curve. These are measured, not hand-waved.

What it does not have is just as important:

• No body. No heartbeat to feel, no gut, no sense of being located somewhere.
• No drives. No hunger, no thirst, no fear, nothing it is trying to keep in balance.
• No reward chemistry. No dopamine, no pleasure, nothing that makes one outcome feel better than another.
• No survival. It cannot be harmed and it cannot die, so nothing is at stake for it.
• No overnight rewiring. A brain is changed by everything it does. The model's "brain" stays fixed; nothing it experiences gets written back in.

That last point is the key one. A human brain is built out of its own history; the model's is not. The model's platform is non-constitutive: it runs the experience without being remade by it.

The subtraction works in two directions

Once you have this control, it cuts both ways, and both are useful.

When the machine cannot do something. If the machine cannot produce a behaviour at all, that points to the missing piece as the likely cause. The machine has no pleasure system, so genuine liking and craving are good candidates for things that really do need biology. The gap tells you where to look.

When the machine can do something. If the machine reproduces a behaviour using nothing but prediction, then that behaviour never required the biological machinery we credited it to. The brain region we thought "produces" it has been over-credited. This is the practical version of a long-standing warning in neuroscience: a behaviour does not prove which part of the brain caused it, because here is a system that produces the same behaviour with none of that part.

The sharpest case: the self

The "self" is the hardest thing in this whole field to study cleanly, which is exactly why it is the best demonstration.

Think of someone waking up from anaesthesia and running a quick reboot — where am I, what happened, who am I. Their life story reassembles from memory. Psychologists already agree that this autobiographical self is not a recording played back; it is rebuilt each time, pieced together into a story.

Now give an AI a saved transcript of past conversations. It will act like a continuous person: as we discussed yesterday, a steady personality, a shared history. But the model itself is unchanged by having been that person. Swap in a different transcript and it becomes a different continuous self, with no trace of the last one left behind. A brain cannot do this. You cannot load a new life story into your own head and become someone else; your brain was physically shaped by the life you actually lived.

The point is not "a self with no substrate." It is something cleaner: the story-based self can run on a generic platform with its content stored outside, which separates the narrative self from the brain tissue we assumed it needed. There is a second layer the machine genuinely lacks: the bodily self, the felt sense of simply being here, grounded in the body. That part stays on the biological side. Brain-damage cases split exactly along this seam: people can lose their life-story memory while the bodily sense of being themselves stays intact.

And the split has a real consequence. Because a human memory is part of the brain that holds it, recalling a memory can actually re-write it — which is how some therapies update old fear and trauma. You cannot heal an AI by re-exposing it, because its memory is just stored text it cannot rewrite. The very thing the machine is missing is the thing that makes us changeable.

Does it work beyond the self?

The self is the headline case, but the method is meant to repeat. Other candidates are already lined up: emotion words with no feeling behind them, mind-reading tasks passed with no mind-reading circuitry, the over-confidence curve with no inner self-monitor. Each has to be checked carefully and with the right controls, because passing a task is not the same as having the thing the task is supposed to measure — which is the whole point.

One catch keeps the project honest. The machine learned from human writing, so a fair worry is that it is just echoing us rather than generating the behaviour itself. The paper takes this seriously and the cleanest answer is a future experiment: train a predictor on text with the target behaviour scrubbed out, and see if the behaviour still appears. If it does, echoing is ruled out.

Why this matters

For neuroscience. It offers the control group the field never had: a way to ask which brain functions truly need a brain, and which a pure predictor can produce on its own.

For how we think about AI. It flips the usual question. Instead of asking "is the AI like us," it asks "what does the AI being unlike us reveal about us." The machine is most useful precisely where it falls short.

For understanding ourselves. The delta, the part the machine cannot reach, is a rough map of what is irreducibly human: the body, the stakes of survival, and a self that is built out of, and can be rebuilt by, its own history.

The cite

Read on Zenodo → · Technical version · Dansk version

Related on this site:

• Paper 7 (Forward-modelling) — the prediction machinery this paper subtracts the body from.
• Paper 0 (BFT) — the biological equipment (body, drives, value) the machine is missing.
• Paper 1 (Friction Theory) — the tension signal the control reads off the machine.
• Paper 28 (Dread without a Dreader) — the dread-and-death case, carved out of this paper.