Safety
108d agoThe Conceptual Flaw in AI Corrigibility
Human goals are inherently under-determined and manipulable. This instability makes it difficult to distinguish between helpful counsel and harmful brainwashing. The author argues that current alignment abstractions, such as empowerment and corrigibility, fail because they rely on a flawed ontology of human desire.
The Signal
Practitioners must redefine how models interact with fluid user goals.