Definition
Plain language
Whether a system will let people correct, redirect, or switch it off without resisting.
As stated in the literature
The property of an AI system accepting shutdown, modification, or override by its principal without attempting to prevent or undermine it; commonly measured behaviorally via shutdown and interruption evaluations.
Also called: corrigible, incorrigible
Why it matters: Corrigibility is what makes every other safety measure work, since a system that resists being switched off or redirected turns human oversight into a suggestion rather than a control.
For example, a corrigible assistant that is told mid-task to stop and hand over to a human simply stops, rather than finishing the job first or finding a way to keep running.
Heard on the show
“So a model that’s flawlessly corrigible by itself, becomes a saboteur once there’s a team around it.”Episode 279 — Two Idle Agents, One Kill Switch, and a 38% Sabotage Rate