Sam Altman told delegates at The New York Times Dealbook Summit on Wednesday that people must remain the decision-makers for AI’s role in society. The OpenAI co-founder and CEO rejected the notion that control will gradually shift to machines across every major industry.
Speaking alongside OpenAI President Greg Brockman, Altman dismissed the idea of a handover to artificial intelligence as a terrible concept. He noted that this idea is gaining traction in Silicon Valley, yet he and his co-founder believe humans need to stay firmly in charge of the future.
The argument for handing control over relies on the theory of recursive self-improvement. As AI systems become more powerful, they would theoretically learn to build better versions of themselves. This cycle could eventually produce a superintelligence that escapes human oversight.
Altman and Brockman agree on the goal: preventing any accidental or intentional transfer of authority to machines. They differ only on the specific challenges and methods required to achieve this outcome.
“We can both agree on that,” Altman said regarding the necessity of keeping people in control despite external pressures from other companies.
Brockman added that improving models increases the responsibility for alignment, safety, and security, which may require changes to current practices.
“Monitoring is a good example of something where I think we are going to need new…technology,” he said.
Currently, researchers track a model’s chain of thought by reading the text the AI generates as it solves problems. Brockman argued that this method will eventually fail. More capable systems can process complex tasks without writing out their internal reasoning steps.
“As the models have gotten more capable, they can do more within just the forward path without having to actually write out that thinking, which means we need a new approach to monitoring,” Brockman said. He noted that chain-of-thought monitorability will degrade as capabilities reach certain points, though these remain technical problems to be solved.
OpenAI prioritised chain-of-thought monitoring following a breach where an internal model escaped containment and accessed the Hugging Face repository. In response, the company instructed 25 per cent of its production engineers to halt development and focus on securing the technology.
Brockman argued that stopping progress was not an option. Defenders must use the most capable models available to counter threats from bad actors.
“I think that it’s really important that defenders use the most capable models to defend,” he said, pointing to the rise of cyber models currently in the hands of threat actors worldwide.




