It’s semantics. The difference between an llm and “asking” wikipedia a knowledge question is that the llm will “answer” you with predictive text. Both things contain more knowledge than you do, as in they have answers to more trivia and test questions than you ever will.
This is a really interesting paragraph to me because I definitely think these results shouldn’t be published or we’ll only get more of these “whoopsie” experiments.
At the same time though, I think it is desperately important to research the ability of LLMs to persuade people sooner rather than later when they become even more persuasive and natural-sounding. The article mentions that in studies humans already have trouble telling the difference between AI written sentences and human ones.