Evil twins are not that evil: Qualitative insights into machine-generated prompts