Governing systems that learn

Most conversations about governing AI start with the wrong picture. Technology sprints ahead, regulation stumbles after it, and lawmakers are always one model release behind. It’s a picture that suits everyone in it. Technologists get to look unstoppable. Governments get an excuse for being slow.

I don’t think regulation was ever a race. It’s closer to a relationship between a society and the tools it decides to live with. So the question I care about isn’t how fast a state can write rules. It’s whether anyone inside it can understand those rules, check they’re working, and change them when they aren’t.

Rules are the easy part

Writing principles for AI has become almost a genre of its own. Fairness, transparency, accountability, human oversight: the same words turn up in multilateral declarations and corporate charters alike, and hardly anyone argues with them.

The trouble starts one level down. Who in a ministry can actually judge the model a vendor is selling? Which regulator has the technical staff, the data and the legal power to investigate when an automated decision hurts someone? And what happens to a rule written for this year’s systems when next year’s behave differently?

Those are questions about how institutions are designed, and my training in development policy keeps pulling me back to them. A policy that a state lacks the capacity to implement is not a policy. It is an aspiration with a letterhead.

The view from the Global South

For a country like India, the stakes cut both ways. AI could make a real difference to public services in agriculture, health and education, exactly where the gap between what a state intends and what reaches a citizen is widest. It could also mean importing systems built somewhere else, trained on other people’s data and shaped by other people’s assumptions.

There are two easy responses. Copy the most comprehensive framework going, or regulate as little as possible and call it growth. I don’t find either convincing. What we need are institutions that can learn about these systems roughly as fast as the systems learn about us.

Where I’d start

If a government asked me where to begin, before any big piece of legislation, I’d suggest three things.

1. Technical literacy across the public sector. I don’t mean one AI agency. I mean people in every major department who know what to ask a vendor.
2. Rules that expect to be revised. Review dates, sandboxes and sunset clauses, so the rules improve because of evidence rather than because of a scandal.
3. A seat at the table. Countries like India should be helping to write international standards, not just receiving them.

None of this makes headlines. Building institutions rarely does. But my bet is that the countries that govern AI well will be the ones that treated it as a long, slow question of state capacity, and not as a sprint.

K.