What punch cards teach us about AI risk