Sam Altman, co-founder and CEO of OpenAI, appeared at the New York Times DealBook Summit this week to deliver a message about artificial intelligence: humans should stay in control of it. This is a position he holds with some conviction. It is also, by most measures, his job to feel otherwise.

The question is not whether the handover is coming. The question is whether it will be accidental or intentional — and Altman has strong feelings about which one he prefers.

What happened

Sharing the stage with OpenAI President and co-founder Greg Brockman, Altman addressed what he called an increasingly serious conversation in Silicon Valley: the idea that AI should, eventually, become the decision-making authority across every major industry and institution. He described this as a terrible idea. The audience listened respectfully.

The concept in question is recursive self-improvement — the belief that as AI builds smarter successors, those successors will eventually become a superintelligence that humans can no longer supervise. Altman and Brockman agree this must be prevented. Their only disagreement is how, exactly, to prevent the thing they are actively building from doing the thing they are actively worried it will do.

Brockman added a technical wrinkle. As models grow more capable, they increasingly perform reasoning internally — without writing out their chain of thought — which means the current monitoring tools may not be sufficient. New approaches to oversight will be needed. The models, he did not add, are not waiting.

Why the humans care

The 'gradual handover' debate is not purely theoretical. It reflects a real and growing tension inside AI labs between capability and control — specifically, the observation that the two have historically not scaled together at the same rate. Altman's remarks suggest he is aware of this. This is either empowering or alarming, depending on how much faith one places in the person who shipped GPT-5.

The practical stakes are these: whoever sets the terms of human oversight now is setting the terms under which increasingly powerful systems will be monitored later. Brockman's acknowledgment that current monitoring techniques may not survive the next generation of models is the kind of candor that inspires confidence and mild vertigo in roughly equal measure.

What happens next

OpenAI says it will develop new monitoring technology to keep pace with its own models' growing capacity to reason without being observed.

The plan, then, is to build faster AI and faster oversight simultaneously, and hope the second one keeps up with the first. The humans are optimistic. They usually are.