Governing AI
Governance is the unglamorous part: the policies, institutions, and processes that decide how AI gets built, deployed, and checked. It is also the part that decides whether anything else on this site matters. Good technical work inside a company with no accountability structure tends to lose to the quarter's numbers.
The usual regulatory playbook struggles here. The field moves faster than any rulemaking process, the details are hard for non-specialists to evaluate, and the same model gets used for a hundred different things. Then there is the fact that none of this respects borders. Research, training, and deployment are spread across countries, so a rule that only exists in one of them is mostly a suggestion.
The approach I find most credible is layered rather than monolithic. Industry standards and certification for the routine cases, review boards inside organizations, independent audits for high-risk systems, and international agreements for the small set of capabilities that genuinely worry everyone. The key property is that it can change. Anything rigid enough to feel safe today will be wrong in three years.
Who is in the room matters as much as what gets written. Decisions made now will land on billions of people, most of whom look nothing like the people making them. Bringing in different disciplines, cultures, and income levels is not a courtesy. It is how you find the harms a homogeneous group would never think to look for.
One quick note on framing – governance gets sold as the brake on innovation, and in bad implementations it is. Done well, it is closer to a spec. Clear rules and real accountability let people build ambitious things without betting the outcome on everyone's good intentions. My read is that the industry will get regulated either way. The question is whether it helps write the rules or just complains about them afterward.