What happens when autonomous machines have “to choose between various shades of wrong?”