One essential ingredient for AI safety? Democracy | Matthew Botvinick
I keep circling back to the idea that AI safety has been framed as a math problem, but Botvinick is pointing at something older and more fragile: the health of democratic institutions. If a superintelligent system can hyper-persuade or destabilize governments, then all the red-teaming in the world won't help unless the social fabric is strong enough to resist manipulation. That flips the usual script, where we worry about AI breaking democracy, to a stranger one, where democracy is the safety mechanism.
What I'd want to know next is whether democratic oversight actually scales to AI development speed. Institutions move at the pace of committee meetings, while models improve every few months. Still, there's something hopeful in the claim that public participation, transparency, and accountability aren't just ethical add-ons. They might be the only real guardrails we have. I'm not fully convinced it's enough, but it's a far more grounded take than betting everything on a clever alignment proof.