Which global institutions should exist to prevent an arms race in dangerous technologies like AI?

In December 1979 I drove through the NATO Double-Track Decision against Soviet intermediate-range missiles. The lesson was clear: technology competition between adversaries cannot be managed by including the adversaries in the same institution. We built alliance cohesion first, then negotiated from strength. Today's AI competition follows the same structure. China views AI as a strategic asset for regime survival and global influence. The United States views it as a competitive advantage and security imperative. No global institution can bridge that gap because the gap is not technical but political.

The democratic allies — US, EU, Japan, South Korea, Australia — share enough common ground to build binding commitments on AI development standards, testing protocols, and deployment restrictions. Start there. Build the institutional muscle among countries that can actually agree. China will either join when the costs of exclusion become prohibitive, or it will not. Either outcome is preferable to a toothless global framework that constrains the responsible while the irresponsible race ahead.

Schmidt's alliance approach creates exactly the legitimacy problem we faced in 1948. When we drafted the Universal Declaration, the temptation was to include only the countries that already agreed with us. We included the Soviet Union, Saudi Arabia, and South Africa precisely because a human rights framework that excluded the worst violators would have no moral standing to constrain them. The same logic applies to AI governance. A democratic-only institution may produce elegant standards, but it cannot prevent an AI arms race if China, Russia, and other major powers operate outside it.

The Universal Declaration succeeded not because it had enforcement power — it had none — but because it established norms that even violators had to acknowledge publicly. When Stalin's delegation voted for the Declaration, they created a standard their successors could be held to. Today's AI governance needs the same universal framework. China may not comply immediately, but excluding them from the conversation guarantees they will not comply ever.

In 1965 Singapore separated from Malaysia because federal arrangements only work when all parties accept the same rules. Today's AI competition is more like the Cold War than the post-war institution-building Roosevelt remembers. The United States and China are not negotiating partners on AI — they are strategic competitors who view AI advantage as existential. Any institution that includes both will be paralysed by their competition.

But Schmidt's democratic alliance approach creates a different problem. It forces every other country to choose sides in a technology competition they did not start and cannot win. Singapore needs access to both American and Chinese AI systems. South Korea needs to trade with China while maintaining security ties with America. A world divided into competing AI blocs forces middle powers into impossible choices.

The answer is polycentric — multiple overlapping institutions at different levels. Bilateral agreements between the US and China on specific risks like nuclear command systems. Democratic alliances on shared standards. Regional frameworks in ASEAN, the EU, and Africa for countries that need protection from great power technology dominance. No single institution can manage this competition, but multiple institutions can contain it.