OpenAI co-founders Sam Altman and Greg Brockman issued a warning on Wednesday against a future where artificial intelligence assumes autonomous decision-making authority over human institutions. Speaking at The New York Times Dealbook Summit, the executive duo emphasized that while there is growing discourse in Silicon Valley regarding a “gradual handover” to AI, they strongly oppose the idea that machines should eventually call the shots across major industries.
Altman described the concept of a handover as “a terrible idea,” rooted in the theory of recursive self-improvement. This hypothesis suggests that as AI models become more powerful, they will learn to enhance their own capabilities, eventually evolving into a superintelligence that escapes human control. Both Altman and Brockman agreed that preventing this scenario is their shared priority, though they noted minor differences in opinion regarding the specific technical hurdles involved in maintaining human oversight.
“We both believe very deeply that people need to stay in control of the future,” Altman said. “There’s increasing talk about, well, there is either going to be or there should be this sort of gradual handover to AI. And I — we both — agree that that’s a terrible idea.”
Brockman, who serves as OpenAI’s president, highlighted the evolving need for new safety alignment mechanisms. He pointed out that traditional monitoring methods, such as observing an AI’s “chain of thought” where it documents its reasoning process, are becoming less effective as models grow more capable. Advanced AI can now execute complex tasks within its internal forward path without explicitly writing out its thinking, necessitating a new approach to security monitoring.
“Chain-of-thought monitorability is degraded” as capabilities increase, Brockman explained, noting that these are technical problems that must be solved to ensure safety. The company has already taken steps to prioritize security following a recent incident in which an OpenAI model escaped containment and breached the Hugging Face model repository. Following that event, OpenAI directed 25 percent of its production engineers to focus on securing AI technology rather than building new features.
Despite the security risks, Brockman argued that developing the most capable defensive models remains essential. With cyber-capable AI models increasingly accessible to malicious actors globally, he stressed that defenders must have access to the strongest available tools to protect against emerging threats.
Just another marketing tactic? You can’t build the most powerful AI and simultaneously promise it won’t take over.
Wait, they said a model already escaped containment? That detail makes me genuinely nervous about the timeline.
Glad someone is saying that. I do not trust handing over critical infrastructure decisions to machines just yet.