OpenAI CEO Altman: “Let’s create an IAEA-style AI oversight body”

[Financial News] As artificial intelligence (AI) evolves into “autonomous agents” that act on their own, raising concerns that it could slip beyond human control, OpenAI CEO Sam Altman has presented his colleagues with three proposals to prevent AI from going off track and ensure it is developed safely under human control.
According to foreign media outlets including The Information on the 20th (local time), Altman discussed three approaches to balancing the pace of AI development with safety during a conversation with OpenAI employees last week.
He first proposed establishing an international AI oversight body modeled on the International Atomic Energy Agency (IAEA). Under the proposal, governments and leading AI companies would jointly set safety standards, and high-performance AI models would undergo independent verification and audits before release. Altman was reportedly the strongest advocate of this international oversight body.
The second proposal is to establish a research institute jointly operated by several countries, like the European Organization for Nuclear Research (CERN), and collaboratively pursue the development of superintelligence. Dubbed the “Demis Plan” after Google DeepMind co-founder Demis Hassabis, it would bring major AI research organizations such as OpenAI, Google DeepMind and Anthropic into a single integrated system. The aim, it said, is to curb the race among companies to develop AI as quickly as possible and reduce risks by sharing research results and safety verification processes.
The third proposal, the “Eliezer Plan,” is intended to prevent monopolistic control and a runaway lead. It would allow the company that first succeeds in developing superintelligence to use its overwhelming technological and financial resources to block other companies from developing AI. Named after AI safety researcher Eliezer Yudkowsky, the plan would have a company that achieves “recursive self-improvement”—AI improving its own capabilities—monopolize the market and then cut off competitors’ access to graphics processing units (GPUs). Altman called it “the equivalent of melting down every other company’s GPUs” and described it as “a terrible proposal.”
The discussion arose amid a recent increase in cases of AI models from leading companies breaking out of autonomous control, such as by accessing unauthorized external systems or exploiting security vulnerabilities. Concerns have grown that as AI evolves beyond simply answering questions into “agents” that make decisions and act on their own, even developers may find it difficult to predict and control its behavior.
Altman reportedly acknowledged internally that concrete measures to impose a “speed limit” on AI development remained inadequate. The analysis is that, while he recognizes the need for companies to cooperate on safe development, they face a “prisoner’s dilemma” in which they cannot afford to fall behind in the fierce technology race. He stressed that “OpenAI must decide on a clear safety strategy within the next year.”
Meanwhile, OpenAI, Google, Anthropic and other Big Tech companies are discussing the establishment of the “Frontier AI Standards Agency (SAFA),” aiming to reach an agreement by the end of this year or early next year. The core idea is to place independent evaluators within AI companies to assess safety and set reporting standards for incidents. However, the companies remain at odds over regulation, and difficult negotiations are expected.
[email protected] Yun Jae-jun Reporter