Experts say balancing the need to govern AI for public safety with user privacy is ‘extraordinarily difficult’.