The most dangerous answer is not an obviously bad one. It is the plausible answer that makes you stop checking. AI needs to become better at showing its uncertainty, not just hiding it behind fluent language.