AI models now refuse many harmful requests. Companies train them with exercises rewarding refusal and punishing over-refusal. Refusal mechanisms are probabilistic and often fail. Models still retain harmful knowledge despite refusal training. Governments may soon define their own refusal boundaries. AI refusal could enable repression by blocking legitimate speech. Refusal is central to AI safety but carries risks. The process of refusal is not fully understood by researchers. Classifiers and probes add layers of safety but remain imperfect. Jailbreaking can bypass refusal mechanisms entirely.