International Cooperation and Competition

AI governance is one of the defining policy challenges of the 21st century. Nations face a dilemma that recurs throughout history when powerful technologies emerge: each country has incentives to develop and deploy the technology as fast as possible for competitive advantage, yet the risks those technologies create are often shared and cross-border. Nuclear weapons, financial systems, aviation safety, and climate change all required some form of international coordination, despite the competitive pressures that militated against it. AI presents this dilemma in an especially acute form, because the technology is advancing faster than most governance institutions can track.

The Governance Landscape

International AI governance is not a single institution or treaty — it is a fragmented and rapidly evolving landscape of forums, agreements, standards bodies, and regulatory frameworks. Understanding this landscape requires distinguishing between hard law, soft law, and technical standards. Hard law refers to legally binding international agreements — treaties, conventions, directives with enforcement mechanisms. AI has very little hard international law. The most significant binding regulatory framework is the European Union's AI Act, which is not international but applies to any company selling AI systems in the EU, giving it extraterritorial reach similar to GDPR's effect on global data practices. Soft law refers to non-binding agreements, declarations, and principles that shape behavior through normative pressure rather than legal enforcement. The OECD AI Principles (2019), endorsed by 46 countries, commit signatories to 'human-centered, trustworthy AI' and include principles on transparency, accountability, robustness, and avoiding discriminatory outcomes. The UN's UNESCO Recommendation on the Ethics of AI (2021), adopted by 193 member states, similarly establishes normative expectations without enforcement mechanisms. Technical standards set specifications for how AI systems are built, tested, documented, and evaluated. ISO/IEC JTC 1/SC 42 is the international standards body developing AI standards, covering topics like AI risk management, bias in AI systems, and AI system lifecycle. Standards matter because they shape industry practice even without the force of law — and because the countries and blocs that dominate standard-setting effectively export their norms globally.

Milestone Moments in Global AI Governance

Several events mark the developing arc of international AI governance. The OECD AI Principles (2019) were the first intergovernmental AI standard, establishing a baseline for responsible AI that has been widely cited by national governments in their own AI strategies. The Global Partnership on AI (GPAI) was launched in 2020 by Canada and France (with US backing) as a multi-stakeholder forum — governments, industry, civil society, and academia — to bridge AI governance gaps. GPAI now has 29 member countries. Its working groups have produced practical guidance on responsible AI, data governance, and AI in pandemic response. The Bletchley AI Safety Summit (November 2023) was hosted by the UK at Bletchley Park — the historic wartime code-breaking center — and brought together representatives from 28 countries, including the US, China, the EU, India, and major AI companies. Its Bletchley Declaration acknowledged that frontier AI poses 'potentially catastrophic' risks and committed signatories to information-sharing on frontier model risks and joint safety research. China's signature was significant — even at a moment of intense US-China AI competition, both recognized shared risk from the most powerful AI systems. The UN General Assembly AI Resolution (March 2024) was the first UN resolution on AI, passed unanimously by all 193 member states. It called for safe, secure, and trustworthy AI that respects human rights and supports sustainable development — a statement of shared aspiration rather than enforceable commitment, but notable for its unanimity. The International AI Safety Institute Network (2024) connected national AI safety institutes — the UK AISI, US AISI, and equivalents in other countries — to coordinate on evaluating frontier AI models for dangerous capabilities.

The Limits of Cooperation

The governance achievements described above are real, but they face serious structural limitations. The verification problem: unlike nuclear weapons, which are physically detectable, AI capabilities are largely invisible. It is extremely difficult to verify whether a country or company is developing AI systems beyond agreed thresholds, training models with prohibited capabilities, or sharing dangerous AI knowledge with sanctioned parties. Without verification, agreements become honor systems. The speed problem: governance institutions operate on timescales of years to decades. AI capability is advancing on timescales of months. By the time an international agreement is negotiated, ratified, and implemented, the technology it was designed to address may have been superseded by new capabilities. The inclusion problem: global AI governance conversations are dominated by the same wealthy countries that dominate AI development. The countries most affected by AI harms — in the Global South, with limited AI capability — have the least voice in governance forums. A governance system designed primarily by and for AI producers is unlikely to adequately represent the interests of AI subjects. The competition problem: even when countries recognize shared risk, competitive incentives create pressure not to constrain one's own AI development while rivals proceed unconstrained. The race dynamic can undermine the cooperation dynamic, even among parties with genuinely shared interests. Despite these limitations, international AI governance is maturing faster than many pessimists predicted. The Bletchley process, the UN resolution, and the spread of national AI safety institutes all represent genuine progress — and the alternative to imperfect governance is no governance at all.