Wait a second, so I saw some LLM generated stuff and I was thinking, LLM's generate a predicted response based on the wording in the prompt, and if you call it out for a false statement, it then generates a predicted response based on that. So even the correction is just a generated based on who you word your response to the incorrect information. What are we even doing? (am I on the right track here?)