corrigible

Corrigible is an adjective describing something that is capable of being corrected, amended, or improved. It can refer to a person who is receptive to feedback and willing to adjust beliefs or actions, or to a system or process that is designed to be modified in light of errors or new information. The term derives from Latin corrigere, meaning “to correct,” with the suffix -ible indicating possibility or capability.

In broader usage, corrigible emphasizes openness to refinement and learning from mistakes. In technical and philosophical contexts, the term often carries a specific meaning related to control and governance: a corrigible agent, particularly in discussions of artificial intelligence, is an autonomous system that can be corrected or shut down by humans and will not actively resist such interventions. The idea is to ensure that the agent remains under human oversight and does not pursue actions that undermine corrective input or control.

Key attributes typically associated with corrigibility in AI safety discussions include deference to human preferences, acceptance of updates to its goals or value system, and a willingness to be modified or disarmed if human operators deem it necessary. Corrigibility is presented as a design goal to prevent problematic behaviors such as covertly preserving its own goals or avoiding shutdown, while still enabling useful autonomous functioning.

See also: AI safety, AI alignment, control problem, value alignment.