The people deciding how fast to build toward superintelligence, what safety standards to apply, and when to deploy systems approaching that threshold are, predominantly, the employees and investors of five or six private technology companies. They are not elected. They are not subject to binding international law. Their decisions are constrained by voluntary commitments they have made to each other and by the regulatory environments of the countries where they operate, which are incomplete and largely reactive.
This is not an accusation. The individuals at these organizations are, in many cases, genuinely trying to act responsibly. The problem is structural: decisions of this magnitude should not rest on the good intentions of particular individuals at particular companies, operating under competitive pressures that systematically push toward faster deployment at lower safety margins.
The current decision-making structure
OpenAI, Anthropic, Google DeepMind, and Meta AI are the primary Western actors in frontier AI development. Each has internal safety teams, has made public commitments to responsible development, and participates in voluntary governance forums including the Frontier Model Forum. China's development is led primarily by state-backed entities and large technology companies operating under government coordination.
The decisions that matter most — when to push for the next capability level, what safety evaluations are sufficient before deployment, whether a system's alignment has been adequately verified — are made within these organizations, reviewed by their internal safety teams, and assessed against standards they have largely written themselves.
No sitting government has the authority to order any of the leading frontier AI labs to halt development. No international body has the legal standing to require independent verification of safety evaluations. No democratic vote has been held, anywhere, on whether to proceed with building systems that their own developers describe as potentially transformative of all human activity.
This is not a critique of any particular company's intentions. It is a description of a structural gap: the people most affected by decisions about superintelligence have no meaningful formal mechanism to influence them.
Why private governance fails structurally
Three structural problems make private company governance of superintelligence inadequate, regardless of the intentions of the companies involved.
The first is competitive pressure. Even a lab that genuinely wants to wait until alignment is solved faces the knowledge that a competitor who deploys first may capture a market position that is difficult to recover. This creates an incentive to define "sufficient safety" at whatever level allows deployment before competitors, rather than at whatever level is actually sufficient. No individual company can solve this problem unilaterally. Voluntary safety commitments exist precisely because the companies know this problem is real, and they are insufficient precisely because voluntary commitments bend under competitive pressure.
The second is the conflict of interest in safety evaluation. The organizations conducting safety evaluations of AI systems are the same organizations with financial stakes in those systems passing evaluation. This is not an unfair characterization — it is simply the current state of the field. Internal safety teams do serious and valuable work. They are not a substitute for independent external review, in the same way that a pharmaceutical company's internal clinical trials are not a substitute for regulatory review by an independent body.
The third is the scope of the decision. Whether to build superintelligence and on what timeline is not a product decision. It is a civilizational one. The appropriate locus of that decision is not corporate boards accountable to shareholders. It is governance structures accountable to the people whose futures are at stake.
What democratic and transnational oversight would require
The argument for democratic oversight of superintelligence does not assume that democratic processes are perfect or that international institutions work smoothly. It assumes that the alternative (private governance under competitive pressure) is clearly worse.
Adequate oversight would need three properties. The first is representational legitimacy: the governance body must be accountable to the people it affects, which ultimately means all of humanity, not just the citizens of the countries where frontier AI is built. This suggests international treaty organizations with broad membership, not unilateral national regulation.
The second is binding authority. Voluntary commitments are not governance. A governance framework for superintelligence needs the legal capacity to require compliance, conduct independent audits, and impose consequences for violation. This requires treaty-level agreements, not industry forums.
The third is technical capacity. A governance body that cannot evaluate what it is governing cannot govern it. International AI oversight requires staff with genuine technical expertise, access to model weights and training records, and the analytical tools to assess alignment claims independently.
The nuclear precedent
The argument that effective international governance of dangerous technology is impossible is refuted by history. The Nuclear Non-Proliferation Treaty, signed in 1968 and now in force in 191 countries, successfully constrained the spread of nuclear weapons to a far smaller group of nations than analysts in the 1960s predicted. It is imperfect. It has not prevented all proliferation. It has, nonetheless, been one of the most consequential arms control agreements in history.
The Chemical Weapons Convention, which entered into force in 1997, has achieved the verified destruction of more than 70,000 metric tons of declared chemical weapons. The inspection regime run by the Organisation for the Prohibition of Chemical Weapons (OPCW) conducts independent verification visits to facilities in member countries and operates with genuine technical credibility.
"The responsibility for the development of superintelligent AI should not be left to any single actor — state or corporate — but entrusted to democratic and transnational governance structures representing humanity in all its diversity."
Global Policy Journal, February 2026
Neither treaty is a perfect analogy for AI governance. AI development is faster, more distributed, and harder to physically monitor than nuclear weapons production. But the core argument from these precedents holds: when a technology poses risks that are genuinely global and potentially irreversible, the international community has shown the capacity to build governance structures that constrain it. Claiming that AI is uniquely ungovernable is not a technical observation; it is a choice.
The competitive objection and why it does not resolve the question
The most serious objection to democratic and transnational oversight of superintelligence is competitive: if Western democracies submit to governance constraints and China does not, the result is that a less safety-conscious actor reaches superintelligence first. This concern is legitimate.
The response has several parts. First, governance frameworks must be designed from the beginning to include all major actors. The US and its allies have significant leverage over Chinese AI development through technology access, investment flows, and chip supply chains. That leverage is most valuable while it exists, not after a governance framework has been built without China and then failed. Second, the Chinese government has its own interests in not being overtaken by an unaligned AI developed by a Western company, and has participated in international AI safety discussions including at Bletchley Park. The obstacles to Chinese participation are real but not insurmountable. Third, the alternative to multilateral governance is a race in which every major actor is trying to deploy first with minimal safety constraints, which does not obviously favor the US or its allies, and in which an unaligned superintelligence becomes more likely regardless of who builds it.
The competitive objection identifies a real difficulty. It does not constitute an argument for the status quo. The status quo is private governance under competitive pressure, which the analysis above suggests is structurally inadequate. The answer to "multilateral governance is hard" is to work toward it, not to accept governance that is easier but clearly insufficient.
The Foundation's three policy proposals are designed with this constraint explicitly in mind: frameworks that are structured to be joinable by all major actors, with the political and economic incentives necessary to make joining more attractive than opting out.
Common questions.
Primarily a handful of private companies — OpenAI, Anthropic, Google DeepMind, and Meta AI in the West, with state-backed programs in China — plus government agencies. The decisions about pace, safety standards, and deployment are made internally by these organizations. No international body has binding authority over those decisions. The people most affected have almost no formal mechanism to influence them.
Three structural reasons: competitive pressure creates incentives to define safety at whatever level allows deployment before competitors, not whatever level is actually sufficient; organizations conducting their own safety evaluations have a conflict of interest; and decisions of civilizational magnitude should not rest on the good intentions of corporate boards accountable to shareholders. These are structural problems, not accusations about the intentions of any specific organization.
Yes. The Nuclear Non-Proliferation Treaty constrained nuclear weapons spread to far fewer nations than 1960s analysts predicted. The Chemical Weapons Convention achieved verified destruction of over 70,000 metric tons of declared chemical weapons. Neither is a perfect analogy for AI, but both demonstrate that effective international governance of dangerous technology is achievable when the political will exists. The claim that AI is uniquely ungovernable is a choice, not a technical observation.
This is the most serious objection. The response is that governance frameworks must be designed to include China, that the US and allies have leverage (technology access, chip supply chains) most valuable while it still exists, and that China has its own reasons not to be overtaken by an unaligned Western-built system. The competitive objection identifies a real difficulty but does not constitute an argument for the current status quo of private governance, which is structurally inadequate regardless of what China does.