Language Models Mostly Know What You Believe