AI safety via debate

Geoffrey Irving, Paul Christiano, Dario Amodei

To make AI systems broadly useful for challenging real-world tasks, we need them to learn complex human goals and preferences. One approach to specifying complex goals asks humans…