LLMs are token completion engines. The correspondence of the text to the truth or authoritative sources is a function of being trained on text like that; with t

LLMs are token completion engines. The correspondence of the text to the truth or authoritative sources is a function of being trained on text like that; with the additional wrinkle that generalization from training (a desired property or it's just a memorization engine) will produce text which is only plausibly truthful, it only resembles training data. Getting beyond this is a tricky dark art. There isn't any simple there. There's nowhere to put an if statement.