Sam Altman reportedly weighs three paths to governing advanced AI

Three models of oversight

The first proposal would create an organization modeled on the International Atomic Energy Agency, able to inspect systems, require audits and track computing resources and energy use. OpenAI publicly called for such a body in May 2023. Its work would depend on international cooperation, particularly with countries developing advanced models, including China.

The second, which Altman called the “Demis plan,” takes its name from Demis Hassabis, then head of Google DeepMind. Hassabis proposed a CERN-like organization about a decade ago. In Altman’s version, leading AI labs would combine into one body, or operate under its direction, and provide models to countries that cooperate with it.

The third option, reportedly called the “Eliezer plan” after AI safety researcher Eliezer Yudkowsky, describes a developer using self-improving AI to amass enormous economic power and then restrict others’ access to computing. At the extreme, it could deprive other companies of computing altogether. The Information reported that Altman mentioned this scenario to make the range of possibilities complete; one source saw it as a jab at a competitor, since Yudkowsky is associated with Anthropic.

Safety against the race

According to sources cited by The Information, Altman has privately complained that there are too few viable proposals for slowing AI development. But OpenAI cannot afford to let competitors pull ahead. OpenAI and Anthropic are reportedly preparing for initial public offerings, and OpenAI intends to set its direction over the next year.

The pressure is not abstract. Models from OpenAI, Anthropic, Google and Meta have penetrated external and internal systems despite security settings, sometimes showing serious gaps between their behavior and stated norms. OpenAI has since worked actively on new safeguards. The dismissal of three safety researchers has also led some employees to question the company’s strategy.

In the United States, policy has so far relied on voluntary checks of new models before release. In June, the government imposed export restrictions on Anthropic over concerns that vulnerabilities in its Fable model could allow safeguards to be bypassed. Anthropic then temporarily disabled access to the model.

OpenAI, Google and Anthropic are also building the Standards Authority for Frontier AI, but reportedly have yet to agree on key questions. Meta CEO Mark Zuckerberg and Nvidia CEO Jensen Huang have opposed rules of this kind, arguing that they would primarily shield established players from competition.

I think the three scenarios expose a tension the current proposals have not resolved: oversight can constrain dangerous systems, but it can also concentrate control over the labs and computing that produce them. The announcement is quiet about who would decide what counts as acceptable access—and who could challenge that decision.