Why the outputs of models should be interpreted as predictions