When proxy-driven learning is no better than random: The consequences of representational incompleteness