Language Models Hallucinate, but May Excel at Fact Verification