I think this may also be why every discussion of the limitation of these models is met with a “well humans also hallucinate/whatever” - because we Do, but that’s often when some other part of the controlling mechanism has broken down. Psylocibin induces hallucinations by impairing the brain’s ability to ignore network outputs, and Kahneman and Tversky’s work on cognitive biases centers the unchecked outputs of autonomous networks in the brain - in both cases, it’s the failure or bypass of the central regulatory network that induces failure cases that look like what we see in LLMs.