According to StanfordAILab, LLMs align on each other’s answers over truth, so polling fails; ICML paper finds truthfulness does not scale without verifiers.