AI Safety via Debate

Authors

Abstract

We explore debate as a training procedure for aligning AI systems with human values. In this approach, we train agents to debate about questions posed by a human, with the human judging the debate.