[saydaark]

Perhaps instead of using the word hallucination we could use average-context. When the LLM doesn't have enough information it computes some average of the information available in the wrapping context, so hallucination is some form of wrong result because of computing averages in a context. But the context could also be wrong.

[yourapostasy]

Have there been papers published about what happens to the user experience when average-context is tightly constrained to small weighting ranges or eliminated altogether, and the model just throws an "insufficient data" error?

[Jensson]

It handles human grammar by averaging and assuming contexts so you can't really fix one side without hurting the other. Humans separates grammar from facts, but these language models don't, to them grammar and facts are the same thing so you can't just tell it to stop lying without it stopping to do all the grammar tricks we expect from it.