Letting AI systems argue with each other may help expose when a large language model has made mistakes.