Three ideas on the table
According to a report by The Information, Altman spoke to OpenAI employees last week about three proposals that people across the AI industry have been discussing. The question behind all three is the same: when AI systems become smarter than humans in most areas, who keeps them under control?
The first option is a new international body, modelled on the International Atomic Energy Agency, which watches over nuclear activity around the world. Such an agency would set safety standards for advanced AI, inspect systems, require audits, and track how much computing power and energy the big labs are using. Altman appeared to favour this approach. It is not a new position for OpenAI, whose leaders have been calling for an agency of this kind since 2023.
The second option has been nicknamed the “Demis plan”, after Google DeepMind chief Demis Hassabis. Under this idea, the major AI companies would merge into, or work through, a single shared organisation, similar in spirit to CERN, the European physics laboratory where countries pool their research. That body would then supply advanced models to partner countries, instead of each company racing on its own.
The third option is the one Altman does not support. Nicknamed the “Eliezer plan”, after AI safety theorist Eliezer Yudkowsky, it imagines one developer becoming so rich and powerful that it can deny rival companies the computing resources they need to build competing systems. Altman included it to complete the list, but made clear he is against it. Some staff reportedly saw the label as a dig at rival lab Anthropic, which has ties to Yudkowsky.
A decision within a year
Altman reportedly told employees that OpenAI needs to decide within the next year which of these paths it will back. That deadline matters, because governments are starting to move. In the United States, the Trump administration is setting up its first framework to oversee companies developing superintelligence, and there is discussion of a draft bill that would let the Commerce Department stop models judged capable of causing large-scale harm.
The big labs are also trying to organise themselves. Several leading firms are working on a new industry body called the Standards Authority for Frontier AI, which is expected to launch by early 2027. The push comes as worries grow about security weaknesses and breaches linked to powerful AI models at companies including OpenAI, Anthropic, Google and Meta.
The clock is ticking
None of the three paths is easy to build. An international agency works for nuclear material because uranium can be counted and stored in known places. AI models are software, and the computing power behind them is rented from data centres all over the world, which makes inspections far harder.
A single shared laboratory raises a different problem: rival companies and rival countries would have to trust it with their best research. And the third path, one winner controlling the supply of computing power, is exactly the kind of concentration of power that regulators usually try to prevent.
Yudkowsky, whose name is attached to that third option, has long argued that superintelligent AI could put humanity itself at risk, and that the world should slow down or stop its development altogether. His warnings are part of why this debate has become so urgent.
Why it matters
The debate sounds abstract, but the outcome will shape who controls the most powerful technology of the coming decade: governments acting together, the labs acting as one, or a single winner taking all. Altman’s comments show that even the company leading the race accepts that the current free-for-all cannot last. What is not yet clear is whether rivals and governments can agree on any of the three paths before the technology forces the question.
