Corrigibility
Corrigibility is a safety property in which an artificial system does not resist legitimate human interruption, correction, or goal modification.
Definition
Corrigibility is a safety property in which an artificial system does not resist legitimate human interruption, correction, or goal modification.
Technical Context
Corrigibility is more than a shutdown button. Objectives, tool permissions, and incentives should be designed so that intervention is not strategically obstructed.