Nicer Than Humans: How do Large Language Models Behave in the Prisoner's Dilemma?