Technology

Altman and Brockman Agree Humans Must Stay in Control. They Disagree on How.

Martin HollowayPublished just now4 min readBased on 1 source
Reading level
Altman and Brockman Agree Humans Must Stay in Control. They Disagree on How.
Photo by TechCrunch / CC BY 2.0

Sam Altman and Greg Brockman both believe people need to stay in control of the future, according to remarks the two made together on stage at The New York Times DealBook Summit. TechCrunch reported their comments on 18 November 2026 (TechCrunch).

Altman described a gradual handover of decision-making authority to AI as a terrible idea. He said the two differ only on how to prevent an accidental or intentional handover to AI, and how to ensure people remain in control. On that account, the disagreement concerns mechanisms rather than the goal itself.

Brockman focused on the technical burden that comes with capability. He said more capable AI models bring increasing responsibility for alignment, safety and security. He also said OpenAI will need new monitoring technology, because chain-of-thought monitorability degrades as models become more capable. Chain-of-thought monitoring means reading a model's intermediate reasoning, the step-by-step text it produces before giving an answer, to detect problematic behavior. Brockman's point is that this signal weakens as models grow stronger.

Brockman also said OpenAI told 25% of its production engineers to stop building and work on securing AI technology after the Hugging Face breach. The statement describes a reallocation of a quarter of the production engineering group, from feature development to security work.

The broader context here is a shift in where the constraint sits. For most of the current model cycle, the main limits on deployment were compute, data and time to ship. Brockman's comments put alignment, safety and security on the same level of priority as capability, and tie that priority to a specific technical problem: the tools used to observe model reasoning may stop working as well as the models improve. If monitorability keeps degrading, labs will need interpretability and oversight methods that do not depend on reading the chain of thought. Those methods are not yet a settled part of the stack.

In my view, the most useful part of the exchange is the distinction Altman drew. Agreement on the goal of human control is easy to state and hard to dispute. The harder question is which mechanisms hold when models are more capable than the people evaluating them, and whether a shared commitment survives competing commercial pressure. Two leaders can agree on the destination and still differ on the route, and the route is where the engineering work sits.

The broader stakes cut in both directions. A failure to keep humans in the decision loop is the risk Altman named. Yet the same investment in monitoring, alignment and security is what would allow capability gains to be deployed in settings where oversight is required. If the monitoring technology Brockman described arrives in usable form, it would widen the range of tasks where people can rely on AI output, not merely limit it. Whether it arrives in time is the open question.