How confessions can keep language models honest