Rethinking Large Language Models in Mental Health Applications