Governance of superintelligence

OpenAI's leadership—Sam Altman, Greg Brockman and Ilya Sutskever—published a short but pointed essay arguing that humanity needs to start governing superintelligence now, years before it likely arrives. Their timeline isn't decades. They think systems dramatically more capable than today's AGI could show up within this decade, and they'd rather see the guardrails built early than bolted on after the fact.

The proposal has three parts. First, they want continued coordination among the labs building the most powerful systems, something like a shared safety baseline so no single company feels pressured to cut corners to stay competitive. Second, and this is the bigger swing, they call for an international body modeled loosely on the IAEA—the nuclear watchdog—that could inspect systems, require audits, test for compliance, and restrict deployments that cross certain capability thresholds. Third, they argue the technical problem of alignment, actually making these systems do what we want, needs serious, sustained research investment starting immediately, not once the systems already exist.

What's notable is what OpenAI explicitly says it doesn't want. The essay argues against slowing down or over-regulating today's models, the ones below some meaningful capability bar, because that risks strangling the benefits of current AI without meaningfully reducing future risk. It's a fairly narrow ask: light touch now, heavy oversight later, with the switch flipped once systems get powerful enough to matter.

There's also a democratic angle buried in the piece. The authors write that decisions about how superintelligence gets used, and who benefits, shouldn't be left to a handful of companies or governments. They gesture at public input and broad benefit-sharing without offering much mechanism for either, which is the part of the essay that reads more like aspiration than policy.