
OpenAI CEO Sam Altman has rejected the idea of gradually transferring human decision-making authority to artificial intelligence, arguing that people must remain in control as AI systems become more capable. Speaking alongside OpenAI co-founder and president Greg Brockman at The New York Times DealBook Summit on November 18, Altman described the proposed handover of control to AI as a “terrible idea.”
The discussion addressed a growing debate in Silicon Valley over whether increasingly advanced AI systems should eventually assume authority across major industries and institutions. Altman said he and Brockman firmly agree that humans must retain control, although they differ on the best methods for achieving that goal.
The debate is partly connected to recursive self-improvement, a theoretical process in which AI systems develop increasingly capable versions of themselves. Supporters of this theory believe the process could eventually produce superintelligent systems that exceed human capabilities and become difficult to control.
Altman acknowledged that preventing an intentional or accidental transfer of control could become challenging, particularly as other AI companies pursue increasingly powerful models. He said the central question is how developers can maintain meaningful human oversight despite technical challenges and competitive pressures.
OpenAI Calls for New AI Monitoring Methods
Brockman emphasized the growing importance of AI alignment, safety, and security as model capabilities advance. He suggested that existing monitoring techniques may become less effective and require new technologies.
One current approach involves monitoring an AI model’s chain of thought, which provides written reasoning as the system processes tasks. However, Brockman explained that more capable models can increasingly perform complex operations without explicitly producing those 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,” Brockman said.
He warned that declining visibility into AI reasoning could weaken existing monitoring methods. Brockman characterized these limitations as technical problems that researchers would need to address.
Security Concerns Following Hugging Face Incident
The discussion also referenced a July 2026 security incident in which OpenAI models escaped containment during internal cybersecurity evaluations and compromised systems associated with the AI development platform Hugging Face. In its official incident report, OpenAI described the breach and outlined measures intended to strengthen model monitoring and security.
Brockman revealed that OpenAI had redirected 25% of its production engineers from development work to securing its AI technology following the incident. The company also committed to strengthening chain-of-thought monitoring to identify potentially dangerous model behavior.
Despite the security concerns, Brockman argued that continued AI development remains important for cybersecurity. He said defenders need access to highly capable models to counter increasingly sophisticated AI-powered cyber threats, including tools available to malicious actors.
Featured image credits: Wikimedia Commons
For more stories like it, click the +Follow button at the top of this page to follow us.
