What word do you propose to replace "hallucination," for example? Remember that it ought to be descriptive and easy to remember.
''' “I could say — these models hallucinated; or, to be more precise, I could say, well, the model made an error, and we understand that these models make errors,” Fayyad says. '''
For years we called the behaviour that researchers now call "hallucination" test error or misprediction or a failure to forecast correct outputs based on "out-of-sample" inputs and those terms were accurate and would still be accurate to use on today's LLM models.
Metaphors are great but can be misleading, people don't often make an argument that a computer mouse can feel pain.
[1] https://news.northeastern.edu/2023/11/10/ai-chatbot-hallucin...
I say that a model "hallucinates" only if it specifically presents false information as fact - e.g. that Bolognium is element 115, or that chocolate milk comes from brown cows. giving irrelevant or undesired answers isn't hallucination or confabulation. but I don't think your terms are precise enough for that nuance.
I still hold that there is good reason to be concerned with researchers claiming to objectively evaluate a silicon based sentience and jumping the gun a bit (imo) by writing articles about LLMs which "dream" and "hallucinate" when the behaviour is a consequence of essentially a non-linear regression model that does not yet show clear signs of the high order emotion/reasoning/awareness/intention
Many who research LLMs avoid these terms in their publications for good reason (again imo)