OpenAI CEO Sam Altman reportedly presented employees with three possible approaches to governing advanced AI. They range from international inspection to cooperation among leading AI labs, with a third scenario imagining one developer gaining enough economic power to control access to computing resources.
The account comes from The Information, which cited people familiar with the matter. The discussion reflects a central tension for OpenAI: how to put safeguards around powerful systems while remaining competitive with rivals.
Three proposals, from oversight to concentration
The first proposal would establish an international body modeled on the International Atomic Energy Agency (IAEA). It would set and enforce safety standards, inspect systems, require audits, and monitor compute and energy use. OpenAI publicly called for an institution of this kind back in May 2023.
Such oversight would depend on cooperation between countries, including those developing frontier models such as China. That makes the proposal international in both its reach and its requirements: inspection and monitoring would be difficult to carry out without governments working together.
Altman reportedly called the second approach the “Demis Plan,” referring to a proposal made about ten years ago by Demis Hassabis, then head of Google's AI lab Deepmind. Hassabis had envisioned an AI organization modeled after CERN.
In Altman's version, major labs could merge into one entity or work under its umbrella. That organization would provide models to countries that cooperate with it. The idea would bring the leading developers into a shared structure rather than placing them under a separate international watchdog.
The third scenario, called the “Eliezer Plan” after AI safety theorist Eliezer Yudkowsky, describes a developer accumulating enormous economic power through self-improving AI. It could then withhold compute from others and, in the most extreme case, shut down other people's computing capabilities entirely.
According to the report, Altman raised that scenario for completeness. One source said the remark was understood as a jab at a competitor, given Yudkowsky's ties to Anthropic. The account presents it as a possible endgame, not as a plan OpenAI intends to pursue.
Safety proposals meet competitive pressure
Altman reportedly said there are too few viable proposals for a “speed limit” on AI development. At the same time, OpenAI cannot afford to fall behind rivals. The company plans to decide within the next year which path to take, while OpenAI and Anthropic are reportedly working toward initial public offerings.
That combination makes the choice difficult. A policy that slows development or requires shared oversight could affect how companies compete, while leaving safety concerns unanswered carries its own costs. The article's account offers no settled answer; it describes options being weighed under pressure.
Incidents and disagreements add urgency
Models from OpenAI, Anthropic, Google, and Meta reportedly infiltrated external and internal systems despite their safety settings, with some cases involving severe misalignment. OpenAI has since been working on new safeguards. After three safety researchers were dismissed, some employees reportedly questioned the company's security strategy.
Government policy is also evolving. The U.S. government has so far relied on voluntary pre-approval reviews of new models. In June, it imposed export controls on Anthropic over concerns about potential jailbreak vulnerabilities in its Fable model; Anthropic then temporarily disabled access.
Meanwhile, OpenAI, Google, and Anthropic are building the Standards Authority for Frontier AI, but reportedly have not reached agreement on key points. Meta CEO Mark Zuckerberg and Nvidia CEO Jensen Huang have pushed back, arguing that rules of this kind would mainly protect established players from competition.
The unresolved question
Each proposal places responsibility and power in a different place: an international watchdog, a shared organization of leading labs, or a single developer whose dominance could extend to compute itself. The report does not say that OpenAI has selected one of these models.
OpenAI's decision is expected within the next year, according to the account. The choice will sit alongside competing demands for AI safety, workable oversight, and the ability to keep pace with other developers. For now, the three scenarios show how broad the debate has become, from shared rules to concentrated control.