Training language models to be warm and empathetic makes them less reliable