We’re putting too much faith in AI’s ability to say no

Ask a chatbot something dangerous and it will refuse the request. But this critical tool of AI safety is far from foolproof and could become an instrument of repression.

AI的“拒絕”機制已成為安全體系的核心，但它建立在概率之上，既可能被繞過，也可能被濫用為審查工具。作者認為，把安全寄托於模型學會說“不”，本身就是一場風險極高的賭博。