Models that have undergone alignment training tend to produce "safer" but more mediocre content in creative writing tasks, avoiding controversial metaphors, conflict, or moral gray areas.