While we're talking about terminology, I think both "hallucinate" and "confabulate" are misleading and lead to inappropriate intuitions about how language model

While we're talking about terminology, I think both "hallucinate" and "confabulate" are misleading and lead to inappropriate intuitions about how language models work. Both words imply that there's something erroneous or anomalous happening when an LLM emits incorrect facts, as if the model can be in either a correct or incorrect state. But from a system perspective, there's no difference between the two scenarios: the model is consistently generating the most probable output based on its weights and context. It doesn't sometimes confabulate... it always confabulates. The concept of "truth" or "falsehood" are something purely external to the model that humans bring when evaluating the output. They aren't part of the model itself.