Human oversight

An approach to AI safety in which people check a model's or agent's work as it goes and themselves approve or reject every serious action.