Can AI safely reject unwanted advances?