Human Compatible

Stuart Russell's 2019 book arguing that the standard model of AI — maximizing a fixed objective — is fundamentally unsafe, and proposing a new approach in which AI systems remain uncertain about human preferences and defer to human correction.

Proposes inverse reward design as an alternative to objective maximization — AI should not be given a fixed goal to maximize but should treat human preferences as uncertain and seek clarification — making safety a feature of the design rather than a constraint.