Large Language Models Assume People are More Rational than We Really are