Safety
109d agoThe Conceptual Flaw in AI Corrigibility
Human goals are inherently manipulable, making it difficult to distinguish helpful counsel from brainwashing. This AI Alignment Forum post argues that concepts like empowerment and corrigibility rely on a flawed ontology. Practitioners cannot easily define a principled boundary for goal interference.
The Signal
This conceptual gap complicates the development of truly obedient and safe systems.