Two-faced AI language models learn to hide deception