Biasless Language Models Learn Unnaturally: How LLMs Fail to Distinguish the Possible from the Impossible