Sam Altman Reportedly Outlined Three Possible Endgames for AI Safety

OpenAI CEO Sam Altman reportedly laid out three internal scenarios for governing advanced AI. The options range from an international watchdog modeled after the International Atomic Energy Agency, to a merged or umbrella structure for major AI labs, to a single company becoming powerful enough to restrict everyone else’s compute. The third path was reportedly included for completeness, and Altman called it “terrible.”

The first approach would create an institution modeled after the IAEA to enforce safety standards. According to the article, such a body could inspect systems, require audits, and monitor compute and energy use. The main challenge is coordination: this model would depend on international cooperation, including with countries building frontier models.

Altman reportedly called the second approach the “Demis Plan,” tied to a proposal from Demis Hassabis for an AI organization modeled after CERN. In Altman’s version, major labs would merge into a single entity or operate under its umbrella. That entity would then supply cooperating countries with models, making governance more centralized but also raising major competition and control questions.

The third approach, reportedly called the “Eliezer Plan,” imagines one AI developer gaining so much economic power through self-improving AI that it could withhold compute from others or shut down competing computing capabilities. The article says Altman raised the scenario only for completeness. Its inclusion underscores how central compute access has become in debates over AI safety, competition, and control.

Altman reportedly said there are too few viable proposals for a “speed limit” on AI development, while OpenAI also cannot afford to fall behind rivals. The article says OpenAI plans to choose a path within the next year. That decision comes as pressure builds around model safeguards, voluntary U.S. pre-approval reviews, and disagreements among major AI companies over frontier AI standards.