Can the World Agree on AI? The Case for an International AI Agency

On a crisp November morning in 2023, twenty-nine nations signed the Bletchley Declaration at the historic code-breaking estate in Buckinghamshire. It was the first time world leaders had gathered specifically to address the risks of frontier AI. The declaration acknowledged that advanced AI poses risks of "serious, even catastrophic" harm. It committed signatories to cooperation. It made global headlines.

It changed nothing.

The Bletchley Declaration is not a treaty. It carries no binding obligations. It established no enforcement mechanism. It did not create an institution. In the 30 months since that symbolic moment, AI capabilities have advanced by multiple generations, autonomous AI agents have entered commercial deployment, and the governance gap between what AI can do and what humanity has agreed to do about it has widened into a chasm.

The question is not whether the world needs a global AI governance framework. That much is beyond dispute. The question is whether the international system — designed in the aftermath of World War II and barely modified since — is capable of producing one.

The Diplomatic Landscape: From Declarations to Frameworks

Since Bletchley, the international community has produced an impressive volume of diplomatic activity and an almost total absence of binding governance.

The Seoul AI Safety Summit (May 2024) expanded participation to 47 nations. It produced the Frontier AI Safety Commitments, voluntary pledges by leading AI companies to conduct pre-deployment safety testing. Voluntary. The commitments have no verification mechanism, no penalties for non-compliance, and no independent monitoring.

The AI Action Summit in Paris (February 2025) brought together 1,000 delegates from 100 countries. It launched the International Network of AI Safety Institutes, connecting national evaluation bodies across the US, UK, EU, Japan, Singapore, and others. This was genuinely useful infrastructure — shared evaluation protocols and information exchange between safety institutes. But it remains a network, not an authority.

The UN General Assembly adopted a non-binding resolution on AI in March 2024 and established the Global Dialogue on AI Governance in 2025 as part of the Global Digital Compact. The Global Dialogue aims to foster interoperability between national governance approaches. It has no legislative authority.

The G7 Hiroshima AI Process produced a code of conduct for AI developers. Voluntary.

The OECD AI Principles (updated 2024) remain the most widely endorsed international AI governance framework. Voluntary.

The pattern is unmistakable: declarations of intent without mechanisms of enforcement. The international community has achieved consensus that AI is important. It has not achieved consensus on what to do about it.

The Council of Europe Framework Convention: The First Binding Treaty

Amidst this landscape of voluntary commitments, one instrument stands apart: the Council of Europe Framework Convention on Artificial Intelligence, Human Rights, Democracy, and the Rule of Law.

Adopted in May 2024 and opened for signature in September 2024, this is the world's first legally binding international treaty specifically governing AI. It is not a declaration. It is not a code of conduct. It is a treaty with ratification requirements, legal obligations, and the institutional backing of the Council of Europe.

As of May 2026, the treaty has 46 signatories, including the European Union (which formally ratified on 15 May 2026), the United States, the United Kingdom, Canada, Japan, and Israel. It requires ratification by at least five parties — including three Council of Europe member states — to enter into force.

The Convention establishes principles-based obligations:

• AI systems must respect human rights, democracy, and the rule of law throughout their lifecycle
• Parties must implement risk identification and management frameworks
• Individuals must have access to legal remedies when adversely affected by AI-driven decisions
• Transparency obligations apply to AI systems interacting with humans
• National oversight mechanisms must be established

Twenty-nine nations signed the Bletchley Declaration. It acknowledged catastrophic risk. It committed to cooperation. It changed nothing.

The Convention is significant for three reasons. First, it demonstrates that binding international AI governance is legally possible. Second, it includes non-Council of Europe members (the US, Canada, Japan, Israel), establishing a precedent for global participation. Third, it provides a legal floor upon which more specific instruments can be built.

But the Convention also reveals the limitations of the current approach. Its obligations are principles-based and intentionally technology-neutral — designed to survive technological change but necessarily vague in operational terms. It exempts national security activities and makes private sector application optional for signatory states. A treaty that allows governments to exempt their most sensitive AI applications and permits private sector opt-outs is a treaty that governs the least dangerous uses while exempting the most dangerous ones.

The IAEA Model: Could It Work for AI?

The most frequently proposed model for global AI governance is an "IAEA for AI" — an international agency modelled on the International Atomic Energy Agency, which has overseen nuclear technology since 1957.

The nuclear analogy is seductive. Both technologies are dual-use, immensely powerful, and potentially catastrophic. The IAEA provides a precedent for international cooperation on existential risk. It has inspectors, verification mechanisms, and (through the Non-Proliferation Treaty) the authority to condition access to technology on compliance with safety obligations.

But the analogy breaks under examination:

Nuclear technology is physical. Weapons-grade uranium requires specific isotope enrichment using specialised centrifuges. These facilities are large, expensive, energy-intensive, and detectable by satellite. AI development requires computers and data — resources that are ubiquitous, commercially available, and can be distributed across jurisdictions. You can inspect a uranium enrichment facility. You cannot inspect a neural network.

Nuclear proliferation has clear thresholds. The difference between a research reactor and a weapons programme is measurable and verifiable. The difference between a large language model and a dangerous AI system is not. There is no agreed definition of what constitutes an "AI weapon" or at what capability threshold a model becomes dangerous.

The IAEA was established before proliferation. The NPT was negotiated when only five states had nuclear weapons. There are currently hundreds of organisations developing frontier AI models across dozens of countries. The proliferation has already occurred.

Political economy differs fundamentally. States agreed to nuclear oversight partly because nuclear materials are scarce and controlled. AI compute and data are the most valuable commercial commodities in the world economy. The nations leading AI development — the United States and China — view AI supremacy as a matter of national survival. Asking them to submit to international oversight is asking them to constrain the technology they consider most vital to their geopolitical future.

An alternative model, sometimes called the "CERN for AI" approach, envisions an international research institution that conducts shared safety research, develops evaluation methodologies, and builds shared infrastructure for alignment testing. This is more politically feasible — cooperation on safety research is less threatening than inspections — but provides no enforcement mechanism.

The Geopolitical Fault Lines

The single greatest obstacle to global AI governance is not technical. It is geopolitical.

The US-China Divide

The United States and China together account for the vast majority of frontier AI development, AI investment, and AI compute capacity. Any global governance framework that does not include both is irrelevant by definition.

But the two nations operate from fundamentally incompatible premises:

The United States views AI governance through the lens of national competitiveness. The Biden administration's Executive Order 14110 represented the most significant federal AI safety action in US history. The Trump administration revoked it entirely in January 2025, replacing it with "AI freedom" principles that prioritise innovation over regulation. The US position in international forums consistently opposes binding multilateral governance in favour of voluntary, industry-led standards.

China views AI governance through the lens of social stability and state control. China's AI regulations — issued rapidly through the Cyberspace Administration (CAC) — mandate that AI outputs reflect "core socialist values." China advocates for UN-centric governance frameworks, partly from genuine multilateralist conviction and partly as a counterweight to US-dominated institutional structures.

Neither nation has an incentive to submit to binding international oversight that might constrain its AI development trajectory. The governance gap at the global level is not an accident. It is a feature of great power competition.

You can inspect a uranium enrichment facility. You cannot inspect a neural network. The nuclear analogy breaks under examination.

The Global South

Developing nations face a different challenge: they are subject to AI systems built, trained, and deployed by companies and governments in the Global North, with little voice in the governance structures that might constrain those systems.

At the UN Global Dialogue on AI Governance, developing nations have pushed back against what they perceive as "club governance" — frameworks negotiated among wealthy nations and presented as global standards. They demand equitable access to AI infrastructure, capacity building, and meaningful participation in governance design.

The tension is real and unresolved. A global AI governance framework built by the G7 and imposed on the rest of the world lacks legitimacy. A framework negotiated through the UN General Assembly, where 193 nations each have one vote, may achieve legitimacy at the cost of decisiveness.

The Sovereignty Paradox

At the core of the international AI governance challenge is a paradox that existing diplomatic frameworks are not designed to resolve.

Traditional international governance assumes sovereign states as the primary actors. Treaties are agreements between states. Enforcement mechanisms apply to states. The international system was built on the Westphalian principle: sovereignty resides in the nation-state.

But AI systems don't respect borders. A model trained in San Francisco and deployed through cloud infrastructure in Singapore can affect citizens in Nigeria, make decisions about loans in Brazil, and influence elections in the Philippines. The entities building these systems are multinational corporations whose power, in some domains, exceeds that of the states ostensibly governing them. OpenAI's computing budget exceeds the entire technology investment of most nation-states.

Global AI governance requires constraining entities that are more powerful than many of the states doing the constraining, using institutional structures designed for a world where state power was supreme.

This is the sovereignty paradox. And it has no solution within the existing international architecture.

The Missing Model: Protocol-Based Governance

The models currently on the table — treaty-based, agency-based, and voluntary commitment-based — all share a common assumption: governance is something that humans do to technology from the outside.

But the most successful global coordination mechanisms of the digital age don't work this way. The internet itself is governed not by a treaty but by protocols — TCP/IP, HTTP, DNS, BGP. These are governance-as-infrastructure: rules that are embedded in the technology rather than applied to it from above. No nation signed a treaty agreeing to use TCP/IP. They adopted it because it worked, and because the cost of non-adoption was exclusion from the network.

Society OS's Global Alliance Protocol (GAP) represents this approach applied to AI governance. Rather than a top-down treaty between nation-states, the GAP operates as a Layer-0 mesh network connecting sovereign Circles — communities, organisations, and individuals — through shared protocols for value exchange, knowledge sharing, and coordinated action.

The GAP architecture addresses the specific failures of existing governance models:

Where treaties are slow, protocols are immediate. Protocol updates propagate through the network in real-time. There is no 38-month legislative cycle.

Where treaties require consensus among sovereign states, protocols achieve coordination through adoption. Participants join because the protocol creates value. Non-participants are not coerced — they simply don't benefit.

Where enforcement depends on state authority, protocol compliance is embedded in the infrastructure. A Circle that violates protocol rules cannot transact with the network. Compliance is not enforced by inspectors. It is enforced by mathematics.

Where the IAEA model requires physical inspection of physical facilities, the GAP uses cryptographic verification. Post-quantum encryption, zero-knowledge proofs, and decentralised identity verification enable trust without surveillance.

The internet did not need a treaty. It needed a protocol. AI governance does not need a declaration. It needs an operating system.

The Global Alliance Protocol doesn't replace international diplomacy. It provides the infrastructure layer that diplomacy is failing to build — a coordination mechanism that works at the speed and scale of the technology it governs, rather than at the speed and scale of human institutions designed for a previous era.

The 42 Protocols: Sovereign Architecture for Global Coordination

Society OS's approach to global AI governance is not a single document or a single institution. It is the 42 Protocols — a comprehensive, modular framework spanning seven categories with six protocols each, covering every dimension of AI governance from Digital Identity Sovereignty to Geopolitical Data Governance.

The 42 Protocols achieve what no international treaty has managed: comprehensive coverage, modular adoption, and real-time adaptability. A nation or organisation can adopt individual protocols based on their needs and capabilities, without requiring wholesale adoption of the entire framework. The framework evolves through protocol updates, not legislative amendments.

This is not a replacement for international law. It is the infrastructure that international law needs to become effective — a shared operating system for coordination that works regardless of whether the signatories are nation-states, corporations, civil society organisations, or individual sovereigns.

The Council of Europe treaty provides legal principles. The OECD provides policy guidelines. The Bletchley, Seoul, and Paris summits provided political declarations.

What none of them provides is infrastructure — the actual technical mechanisms through which coordination occurs, compliance is verified, and governance operates at the speed of the technology it governs.

That infrastructure is what a Living Operating System builds.

Conclusion: The Architecture of Global Cooperation

The world cannot agree on AI governance because the world is trying to govern a technology that operates at machine speed using institutions that operate at human speed. Every treaty takes years. Every protocol update takes minutes. Every diplomatic summit produces a declaration. Every protocol deployment produces a working system.

The path to global AI governance does not run through the UN General Assembly or the Security Council. It runs through the technical infrastructure that makes cooperation the rational choice — not because nations are forced to cooperate, but because the cost of non-cooperation becomes the exclusion from the network of value.

The internet did not need a treaty. It needed a protocol.

AI governance does not need a declaration. It needs an operating system.

The 46 nations that have signed the Council of Europe Framework Convention have demonstrated that the political will exists. The question is whether the institutional architecture can match the ambition.

Currently, it cannot.

But the architecture already exists. It is the Sovereign Stack — seven layers of digital sovereignty from sovereign compute to sovereign governance — built by one person, filed as 2,052+ patent claims on 2 February 2026, and deployed as a living system at sovereignsos.com.

The world does not need permission to adopt it. It needs only the recognition that the future of governance is not another treaty.

It is a protocol.

This article is part of the Sovereign Intelligence Hub's global governance series. For the protocol architecture, see [The 42 Protocols](/hub/society-os-42-protocols). For the governance gap these treaties fail to close, see [The Governance Gap](/hub/the-governance-gap). For the technological infrastructure that makes sovereignty possible, see [The Sovereign Stack](/hub/sovereign-stack-architecture).