In reality, it's the "correct" responses that are the hallucinations, not the incorrect ones. Since the vast majority of the possible outputs of an LLM are "not

In reality, it's the "correct" responses that are the hallucinations, not the incorrect ones. Since the vast majority of the possible outputs of an LLM are "not true", when we see one that aligns with reality we hallucinate the LLM "getting it right".