In an internal briefing, the OpenAI chief outlined three industry-discussed models for governing advanced AI, reportedly favouring an IAEA-style global regulator as the Trump administration builds its first superintelligence oversight framework.
Â
OpenAI chief executive Sam Altman last week walked employees through three proposals for developing powerful artificial intelligence safely, The Information reported. He reportedly told staff the company would need to decide within the next year which approach it is prepared to support. The briefing came as the Trump administration sets up its first framework for overseeing developers working on superintelligence.
Three models on the table
The first proposal envisions an international body modelled on the International Atomic Energy Agency, tasked with enforcing safety standards worldwide. People familiar with the remarks said Altman appeared to lean towards this option. It also fits OpenAI’s public record: since 2023, the company’s leaders have backed such an agency, suggesting it could inspect AI systems, require audits and monitor how much computing power and energy developers consume.
The second, which Altman labelled the “Demis plan,” builds on an earlier idea from Google DeepMind chief Demis Hassabis for an organisation resembling CERN, the European particle physics laboratory. In Altman’s telling, leading AI companies would either fold into one entity or operate through it, and that entity would supply models to partner countries.
The third, dubbed the “Eliezer plan” after AI safety theorist Eliezer Yudkowsky, imagines a single developer amassing enough economic strength to deny rivals access to computing resources. Altman reportedly said he opposed this outcome and mentioned it only to make the list complete. According to the report, some employees took the comment as a jab at Anthropic, given Yudkowsky’s reported connections to that company.
Rivalry and regulation gather pace
The discussion follows a series of security breaches linked to AI, involving models from OpenAI, Anthropic, Google and Meta. Altman has also said internally that OpenAI cannot afford to lose ground to competitors, at a time when both OpenAI and Anthropic are moving towards public listings.
Work on safeguards is moving on several fronts. OpenAI, Google and Anthropic are jointly developing a US safety organisation called the Standards Authority for Frontier AI, which they hope to launch by early 2027. Separately, Anthropic has presented lawmakers with a draft bill that would let the Commerce Department halt models capable of causing large-scale harm.
Taken together, the developments show how the industry’s leading names are trying to shape the rules before regulators do. Altman’s briefing suggests OpenAI itself has yet to settle on a position, even as its leader appears to favour the international-agency route. With a government framework taking shape in Washington and rival firms pushing their own proposals, the company’s choice over the coming year could influence how the wider debate on governing advanced AI unfolds.
