MUMBAI: The race to build superintelligent AI is accelerating, but the race to keep it in check is proving harder to call. OpenAI chief executive Sam Altman has outlined three possible approaches to managing the risks of advanced artificial intelligence, reportedly telling employees that the company must decide which path to support within a year, according to The Information.
The options range from an international watchdog modelled on a nuclear safety regulator to a shared research organisation involving major AI developers. The third envisages a more concentrated form of control, in which a developer with self-improving AI could use its economic power to restrict rivals’ access to computing resources.
The discussion comes as the Donald Trump administration works on its first framework for overseeing developers of superintelligent AI. It also highlights a growing dilemma for the industry: how to encourage technological progress while addressing risks that could extend beyond the control of any single company or government.
The first option is an international body modelled on the International Atomic Energy Agency (IAEA), which monitors nuclear activities and supports safeguards against the misuse of nuclear technology. Under the proposal, a similar institution for AI could monitor development, inspect systems and enforce safety standards.
According to The Information, Altman appeared to favour this approach. OpenAI leaders have advocated for an international AI oversight body since 2023, when Altman, president Greg Brockman and then-chief scientist Ilya Sutskever proposed an agency that could inspect AI systems, require audits and track computing and energy use.
Such a model would seek to move AI oversight beyond individual companies, creating a framework for monitoring powerful systems and establishing common safety expectations across borders. Its effectiveness, however, would depend on international cooperation and the authority governments were willing to give it.
The second proposal, which Altman reportedly calls the “Demis plan”, draws on an idea associated with Google DeepMind chief executive Demis Hassabis. It envisages an organisation modelled on CERN, the European research laboratory known for collaborative scientific work.
Under this approach, major AI companies would provide models to partner countries through a shared research structure. The idea would bring developers and participating countries into a common framework rather than relying exclusively on separate corporate safety policies.
The third option, called the “Eliezer plan”, is named after AI safety researcher Eliezer Yudkowsky. It describes a scenario in which a developer creates self-improving AI and gains enough economic power to restrict competitors’ access to computing chips.
Rather than presenting this as a preferred policy, Altman reportedly included it for completeness and opposed the outcome it describes. Some OpenAI employees viewed the reference as a possible dig at Anthropic, given Yudkowsky’s links to the company, The Information reported.
The three approaches represent markedly different visions of control: international regulation, shared scientific infrastructure and the concentration of technological and economic power in a single developer.
The reported discussion comes amid wider concern about the security and potential risks of increasingly capable AI models from OpenAI, Anthropic, Google and Meta. At the same time, OpenAI and Anthropic are moving towards public listings, adding competitive and commercial pressures to an industry already racing to develop more advanced systems.
The challenge is not limited to deciding who should oversee AI. Developers also differ on how new models should be tested, what risks should trigger restrictions and how much information companies should share before a system is released.
OpenAI, Google and Anthropic are working on a US-based group called the Standards Authority for Frontier AI, which they aim to launch by early 2027. The initiative is intended to contribute to standards for advanced AI, although the companies differ on approaches to testing new models.
Anthropic has separately proposed draft legislation that would allow the US Commerce Department to halt models capable of causing large-scale harm. The proposal points to a more direct regulatory role for government, rather than relying solely on voluntary commitments from developers.
These initiatives show that the industry is exploring multiple approaches at once, from common technical standards to potential government intervention. They do not, however, establish agreement on a single model for overseeing frontier AI.
The Trump administration is also working on its first framework to oversee developers of superintelligent AI, adding a government dimension to the debate over how powerful systems should be managed.
Under the US government’s current voluntary framework, developers are asked to provide access to certain AI models up to 30 days before release. The arrangement is intended to create an opportunity for assessment ahead of public deployment, but it is voluntary rather than a comprehensive legal regime.
On 29 September, executives from Anthropic, Google, Meta, OpenAI, Nvidia and xAI signed a voluntary accord that US President Donald Trump described as “morally binding”. The agreement carries no legal liability, distinguishing it from enforceable regulation.
The contrast between voluntary commitments and proposals for formal oversight remains central to the debate. Industry agreements can provide a basis for cooperation, but they do not carry the same legal force as rules that governments can enforce.
An international agency modelled on the IAEA would represent a more institutionalised approach, while a shared research organisation could facilitate cooperation without necessarily assuming the same regulatory powers. A government-led framework, meanwhile, could establish requirements within the United States, although its reach beyond the country would be a separate question.
Altman’s reported one-year timeline puts pressure on OpenAI to determine which approach it believes offers the most credible way to manage advanced AI risks. The options also raise questions about how much responsibility should rest with individual companies, international institutions and national governments.
An international watchdog could offer common standards and oversight, but would require countries to agree on its remit and powers. A CERN-like organisation could pool expertise and encourage collaboration, while the third scenario serves as a warning about the risks of technological dominance becoming economic control.
The reported comments do not indicate that OpenAI has formally adopted any of the three approaches. Nor do they establish that the US government or other AI developers have endorsed Altman’s preferred option.
For now, the debate remains open as companies continue developing increasingly capable systems and governments consider how to respond. The central question is no longer simply how quickly AI can advance, but who should set the limits, how those limits should be enforced and whether the industry can agree on a framework before the technology moves further ahead.
Altman’s reported deadline gives OpenAI a year to clarify its position. Reaching a workable international consensus, however, could prove a much longer task.