Internal representations of LLMs encode information about truthfulness