Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection