Sam Altman’s remarks at the United Nations Security Council

When a leading lab CEO addresses the Security Council, the audience is not product managers — it is states wrestling with a technology that already feels compressed in timeline, upside, and risk.

Altman’s delivered remarks put two failure modes on the table: losing control of systems that outrun human intervention, and concentrating too much power in too few hands. He pairs that with a call for complementary national and international frontier AI standards, incident reporting, and secure vulnerability channels.

We’re taking this as a political and safety positioning speech, not a product launch.

Altman opens with a fork: AI can expand what people believe they can do, or it can become upheaval that outruns institutions. His preferred frame is agency — tools that help people learn, create, discover, and participate — and an explicit rejection of dystopian “optimize humans away” talk he attributes to parts of the industry. He says OpenAI has unilaterally slowed before and will again; competitive pace is not a reason for rash decisions.

On catastrophe risk, he refuses the comfort of small percentages. Whether someone puts the chance at 10%, 1%, or 0.1%, none are “remotely acceptable,” and labs should not train models they cannot make an extremely strong case for keeping under human control — especially as recursive self-improvement and automated AI R&D could accelerate progress. He says he largely agrees with Professor Bengio, then names two bad paths rather than three: loss of control, and concentrated power used to impose a worldview.

The constructive agenda is three principles: people remain at the center of decision-making (don’t automate human judgment or values); scientific and economic benefits by and for people; and individual empowerment rather than lab-as-hero. He cites a math arc ending with a claim that one of OpenAI’s models recently solved a Millennium Prize Problem — the Navier-Stokes equations — as evidence of discovery potential tied to healthcare and living standards, while insisting companies must not substitute for democratic process.

Internationally he wants shared standards for measuring capabilities, assessing risk, judging safeguards, and preserving meaningful human oversight; common evidence language so countries can compare and verify; fast incident reporting; and secure channels among governments, critical infrastructure operators, and technical experts. Standards, he says, should not lock in incumbents or favor open versus closed developers.

Useful AI deserves this kind of institutional scaffolding. Discovery tools that stay under human control and spread capability widely are the version worth advocating for. The downside of the speech itself is the gap between rhetoric and enforceable mechanism — Security Council remarks don’t create inspectors. Concentration-of-power warnings can also be read as competitive positioning. And extraordinary scientific claims (Millennium Prize–level results) will be stress-tested by independent mathematicians; we’re noting what was said, not rubber-stamping the proof.

I’m watching whether “complementary frontier standards” becomes a real multilateral workstream with verification teeth, or stays a speech that labs cite while shipping accelerates.

OpenAI Global Affairs / Safety post, Sep 23, 2026: “Remarks as delivered” by CEO Sam Altman to the United Nations Security Council on AI opportunity, human control, and international cooperation.