Global Coordination and Competition

AI development is happening simultaneously in dozens of countries, driven by national governments, private companies, and research universities whose incentives do not always align. Two of the world's largest economies — the United States and China — are in an explicit strategic competition for AI leadership that each frames in terms of national security and economic supremacy. The European Union is building regulatory frameworks that affect AI companies globally. India, the United Kingdom, the UAE, Japan, South Korea, Canada, and Israel all have significant AI industries and distinct governance approaches. This is the global landscape within which AI safety and governance must operate.

The Cooperation-Competition Tension

The structural tension is this: safety measures require slowdown, scrutiny, and information-sharing. Competition rewards speed, secrecy, and capability advantage. When two teams are racing, the first one to impose on themselves a safety check the other has not agreed to may fall behind. In a race without agreed-upon rules, the competitive pressure is to drop safety measures that impede speed. This is not hypothetical. Several major AI labs have stated that they believe they are in a race with other labs and with foreign competitors, and that losing this race would be dangerous. The same labs have argued both that AI development needs to slow down for safety evaluation and that they need to move quickly to ensure 'safe' developers (themselves) reach the capability frontier before less safety-conscious competitors. These positions are genuinely in tension. The national security dimension adds further complexity. AI capabilities are relevant to military applications: autonomous weapons, intelligence analysis, signals intelligence, cyberoffense, logistics optimization, and strategic planning. Governments that view AI through a military lens are reluctant to share technical details with potential adversaries — even details about safety research that might be mutually beneficial. This creates a security-safety tension: information-sharing is good for safety but may be bad for security.

What coordination has been achieved? More than is commonly known, though its durability is uncertain. The OECD AI Principles (2019), adopted by 42 countries including the US, EU members, Japan, South Korea, and others, established shared language around trustworthy, human-centered AI. They are non-binding but provided a common vocabulary that shaped subsequent national frameworks. The G7 Hiroshima AI Process (2023) produced a code of conduct for advanced AI development, signed by G7 nations, that includes commitments on safety testing, incident reporting, and transparency. Notably, all seven governments agreed to language about reporting AI-related incidents — a form of information-sharing that is a genuine step toward coordination. The Bletchley Declaration (November 2023), signed at the first AI Safety Summit hosted by the UK, was signed by 28 countries including both the United States and China — one of the few explicit US-China agreements on AI policy. The declaration acknowledges that frontier AI poses serious risks and commits signatories to information-sharing about those risks. The subsequent Seoul AI Summit (May 2024) produced a further declaration with 16 AI companies committing to share safety information with governments. US-China AI cooperation has occurred in narrower technical domains. Scientific conferences like NeurIPS and ICML have historically remained open to Chinese researchers, enabling cross-pollination of basic research. Track-2 dialogues (unofficial meetings between academics and think tanks from both countries) have maintained communication channels even during periods of political tension.

The Brussels Effect and Structural Coordination Mechanisms

One of the most significant coordination mechanisms is not a treaty or a declaration — it is the Brussels Effect. This term, coined by scholar Anu Bradford, describes the phenomenon by which the EU's regulatory standards de facto become global standards because companies operating worldwide find it most efficient to comply with the most demanding requirement everywhere rather than maintain different product configurations for different markets. The EU's GDPR (General Data Protection Regulation) is the canonical example: after its 2018 enactment, companies worldwide updated their privacy practices for all users, not just EU users, because the compliance cost of differentiation exceeded the cost of uniform compliance. The EU AI Act may produce a similar effect for AI governance: if a company must meet its requirements to sell AI services in the EU market, it may simply meet those requirements globally. This is a form of coordination — not through negotiation, but through market structure. The US has been pursuing export controls as a coordination mechanism of a different kind. By restricting exports of advanced chips (specifically NVIDIA's highest-end GPUs and the equipment used to manufacture chips below a certain nanometer threshold) to China and certain other countries, the US attempts to limit the computational resources available for frontier AI development outside US-aligned countries. This is competition-via-governance rather than cooperation — using regulatory tools to shape the capability landscape rather than to promote shared safety norms. The fundamental structural obstacle to more ambitious international coordination is the absence of enforcement. Binding international agreements with real enforcement mechanisms require either sufficiently aligned interests that parties voluntarily comply, or a supranational authority with coercive power — neither of which exists for AI governance at the global level. The most realistic path forward, in the near term, involves building information-sharing, developing common technical standards through bodies like ISO/IEC, and accumulating the trust needed for more ambitious coordination.