Natural Hallucinations

LLMs still hallucinate quite a bit. My bet is, one of the reasons for this is that there are hallucinations that are difficult for LLMs to recognize, even when they have all the information needed. One of the reasons for this may be that hallucinations are usually about something extra an LLM sees, but contradictory in-context information is normal to see, so doesn't necessarily allow the model to know its hallucinating.

Can we see how robust certain hallucinations are to include information/the question asked? Do these hallucinations transfer? If models are trained on these Natural Hallucinations, are these beliefs remembered more easily? If we find them for older models, do they predict hallucinations in later models?