Pre-trained language models already possess the necessary knowledge of underlying reality before alignment