If AI systems are viewed as too advanced to control, the companies that build them can’t held liable for the harms they cause.