Can we reverse-engineer psychiatric-like symptoms in an artificial cognitive system?
I’ve been thinking about a slightly unusual experiment.
What if instead of trying to simulate the whole human brain, we build a simplified artificial cognitive system with perception, memory, attention, prediction, confidence and decision-making?
Then we deliberately start breaking it.
Change the balance between sensory evidence and prior beliefs. Distort memory. Manipulate attention. Change confidence calibration or prediction errors.
Run thousands of variations and observe what happens.
The interesting part begins when the system develops stable abnormal behavior that resembles a known psychiatric symptom — persistent false beliefs, distorted perception, abnormal attribution of meaning, etc.
At that point, instead of only studying the symptom, we can go backwards.
We have something we usually don’t have with a human brain: a complete computational history of the system.
normal state → perturbation → first anomaly → feedback → reinforcement → stable pathological-like state
We could restore the system to a point before the transition and run counterfactual experiments.
Change parameter X — does the symptom still appear?
Restore X — does it disappear?
Change Y instead — what happens?
In other words, rather than asking an AI to imitate schizophrenia, psychosis or another disorder, we would be asking:
Can psychiatric-like phenomena emerge from controlled failures of specific cognitive mechanisms?
And if they do:
Can we identify the computational transition that caused them?
Obviously, an artificial neural network is not a human brain, and an LLM hallucination is not a psychiatric hallucination. Similar behavior would not prove a shared biological mechanism.
But if an artificial system repeatedly produces patterns resembling observations from computational psychiatry, that could give us something testable — hypotheses that could later be compared with neuroscience and clinical data.
Think of it less as a simulated psychiatric patient and more as a crash-test laboratory for cognition.
Break cognition in controlled ways.
Observe what emerges.
Roll it back.
Change one variable.
Run it again.
I’m curious whether anyone here has worked on something similar — especially around computational psychiatry, predictive processing, or artificial cognitive architectures.
Where does this idea break?