Microsoft has drafted a “Humanist AI” code requiring future models to accept human shutdown, follow non-negotiable safety rules, and remain under control.