I think the distinction is utterly useless in such a use case and ultimately just expresses our understanding or interpretation of the outcome. And trying to ca

I think the distinction is utterly useless in such a use case and ultimately just expresses our understanding or interpretation of the outcome. And trying to categorize it's output in those terms is highly misleading regarding how it works. We speak of "hallucination" when GPT provides completions that are not connected to any facts it was "provided with" in either training data or prompts, but ultimately factuality makes no difference to GPT. It ALWAYS just "simulates" text, that's what it does. If you'd want to use these words to describe factual accuracy, or to express if GPT "performs" an action, as opposed to "faking" a result, you have to bear in mind either of these qualifications are nothing but a human interpretation of the completion. It being "correct" more often than not with the content of it's simulation is just a logical, but ultimately accidental consequence of being so well fitted to simulate the structure of data it trained itself on. And a "halucination" is nothing but this simulation deviating from the context you provided, that fundamentally was simulated by the same rules that produced any other content which didn't deviate.