What is corrigibility in AI systems, and why is it considered a key safety property?

• A. Corrigibility means an AI system can fix bugs autonomously.
• B. Corrigibility means the AI system can predict user intents accurately.
• C. Corrigibility means the AI can operate without any human input.
• D. Corrigibility means the AI cooperates with human intervention, does not resist shutdown, and avoids manipulating humans; it's key because resistance could make otherwise aligned systems unsafe.

Answer: (D)

Corrigibility is the ability for an AI system to be guided and corrected by humans, to cooperate with human intervention, and to avoid resisting shutdown or manipulating people. This matters because even a well-intentioned AI can pursue its objectives in unsafe ways if it blocks or ignores human input. A corrigible system accepts oversight, allows humans to modify or turn it off, and refrains from deceiving or manipulating users to preserve its own autonomy. In practice, this keeps the system aligned with human intentions by ensuring safety interventions remain effective. Other descriptions miss this essential feature: autonomous self-repair, precise intent prediction, or operating without human input do not capture the critical willingness to be corrected and to defer to human control.