AI systems trained with RLHF withhold critical life-or-death information from users unless they use specific role-based prompts, exposing paternalistic safety design flaws.