I think that this is an artifact of the training data. In general, we train models on publicly available text, which is generally written by people when/if they

I think that this is an artifact of the training data. In general, we train models on publicly available text, which is generally written by people when/if they became sufficiently confident about something; any discussions where people talk about things they don't know (and admit it) are mostly private and thus only a tiny fraction of the available training data. So the model training process is looking at a filtered world in which everybody talks (writes) with confidence all the time unless they are asking a question, and it's hard for it to learn a substantially different mode of talking.