Can large language models write reflectively