We’re proposing an AI safety technique which trains agents to debate topics with one another, using a human to judge who wins.