International Coordination measures the degree of global cooperation on AI governance and safety across nations and institutions. Higher international coordination is better—it enables collective responses to risks that transcend borders, prevents regulatory arbitrage, and reduces dangerous racing dynamics between nations. This parameter tracks treaty participation, shared standards adoption, institutional network strength, and the quality of bilateral and multilateral dialogues on AI risks.
Geopolitical dynamics, AI development trajectories, and near-miss incidents all shape whether international coordination strengthens or fragments—with major implications for collective action on AI risk. High coordination enables shared safety standards that prevent racing to the bottom; low coordination risks regulatory fragmentation and competitive dynamics that undermine safety.
This parameter underpins critical AI governance mechanisms. Strong international coordination enables shared safety standards that prevent racing dynamics where competitive pressure sacrifices safety for speed—research suggests uncoordinated development could reduce safety investment by 30-60% compared to coordinated scenarios. Technical cooperation enables faster dissemination of safety research and evaluation methods across borders, while multilateral mechanisms prove essential for coordinated responses to AI incidents with global implications. Finally, global governance of transformative technology requires international buy-in for democratic legitimacy, not unilateral action by individual powers that lacks broader consent.
Contributes to: Governance Capacity
Primary outcomes affected:
| Metric | Current Value | Historical Baseline | Trend |
|---|---|---|---|
| AISI Network countries | 11 + EU | 0 (2022) | Growing |
| Combined AISI budget | ~$150M annually | $0 | Establishing |
| Binding treaty signatories | 14 (Council of Europe) | 0 | Growing |
| Summit declaration adherence | Mixed (US/UK refused Paris 2025) | High (2023 Bletchley) | Fragmenting |
| Industry safety commitments | 16 companies (Seoul) | 0 | Stable |
Sources: AISI Network, Council of Europe AI Treaty, Bletchley Declaration
| Institution | Type | Participants | Budget | Status |
|---|---|---|---|---|
| UK AI Security Institute | National | UK | ~$65M | Operational |
| US AI Safety Institute (CAISI) | National | US | ~$10M | Refocused (2025) |
| EU AI Office | Supranational | EU-27 | ~$8M | Operational |
| International AISI Network | Multilateral | 11 countries + EU | $11M+ joint research | Building capacity |
| G7 Hiroshima Process | Multilateral | G7 | N/A | Monitoring framework active |
| UN Scientific Panel on AI | Global | Pending | TBD | Proposed Sept 2024 |
Note: AISI Network held inaugural meeting November 2024 in San Francisco, completing first joint testing exercise across US, UK, and Singapore institutes. Network announced $11M+ in commitments including $7.2M from South Korea for synthetic content research and $3M from Knight Foundation.
Healthy international coordination on AI safety would exhibit several key characteristics that enable effective global governance while respecting national sovereignty and diverse regulatory philosophies.
| Characteristic | Current Status | Gap | Trend |
|---|---|---|---|
| Binding commitments | Council of Europe treaty (14 signatories) | Large—most frameworks voluntary | Worsening (US/UK withdrew Paris 2025) |
| Technical cooperation | AISI network operational; joint evaluations begun | Medium—capacity still building | Improving (first joint tests Nov 2024) |
| Inclusive participation | US/UK diverging from broader consensus | Large—key actors withdrawing | Worsening (governance bifurcation) |
| Rapid response | No mechanism exists | Very large | Flat (no progress since Bletchley) |
| Sustained commitment | Fragile—US pivoted away in 2025 | Large—political volatility | Worsening (administration reversals) |
| Mechanism Type | Example | Enforcement | Coverage | Effectiveness Score (1-5) |
|---|---|---|---|---|
| Binding treaties | Council of Europe AI Treaty | Legal obligations | 14 countries | 2/5 (limited participation) |
| Voluntary summits | Bletchley, Seoul, Paris | Reputational pressure | 28+ countries | 2/5 (non-binding, fragmenting) |
| Technical networks | AISI Network | Peer cooperation | 11 countries + EU | 3/5 (building capacity, concrete outputs) |
| Industry commitments | Frontier AI Safety Commitments | Self-regulation | 16 companies | 2/5 (voluntary, variable compliance) |
| Regulatory extraterritoriality | EU AI Act | Legal for EU market access | Global (via Brussels Effect) | 4/5 (enforceable, broad reach) |
| Bilateral agreements | US-UK MOU, US-China dialogue | Government-to-government | Pairwise | 3/5 (limited scope but sustained) |
| UN frameworks | Global Digital Compact, proposed Scientific Panel | Norm-setting | Universal participation | 2/5 (early stage, unclear enforcement) |
Note: Effectiveness scores assess actual impact on coordination quality based on enforcement capability, coverage breadth, and sustained operation. The EU AI Act scores highest due to legal enforceability and market leverage creating de facto global standards.
| Threat | Mechanism | Evidence |
|---|---|---|
| US-China rivalry | AI seen as decisive for economic/military competition | Export controls since 2022; $150B+ Chinese AI investment |
| National security framing | Safety cooperation viewed as sharing strategic advantage | UK renamed AISI to "AI Security Institute" (Feb 2025) |
| Trust deficit | Decades of strategic competition limit information sharing | US/China dialogue constrained to "working level" |
| Threat | Mechanism | Evidence |
|---|---|---|
| Administration changes | New governments can reverse predecessors' commitments | Trump revoked Biden EO 14110 within hours of taking office |
| Innovation vs. regulation framing | Safety cooperation portrayed as competitiveness threat | Vance at Paris: "cannot and will not" accept foreign regulation |
| Industry influence | Tech companies lobby against binding international rules | $100B+ annual AI investment creates strong lobbying capacity |
AI governance faces fundamental challenges that make international coordination harder than previous technology regimes. Unlike nuclear weapons, AI capabilities cannot be physically inspected through traditional verification methods—recent research on verification methods for international AI agreements explores techniques like differential privacy and secure multi-party computation, but these remain immature compared to nuclear inspection regimes. Nearly all AI research exhibits dual-use universality with both beneficial and harmful applications, making export controls more difficult than weapons-specific technologies. The speed mismatch proves severe: AI capabilities advance weekly while international diplomacy operates on annual cycles, creating persistent gaps between technical reality and governance frameworks. Finally, distributed development across thousands of organizations globally—from major labs to academic institutions to startups—makes comprehensive monitoring far harder than tracking state-run weapons programs.
| Factor | Mechanism | Status | Evidence |
|---|---|---|---|
| Catastrophic risk consensus | 28+ countries acknowledged "potential for serious, even catastrophic harm" | Bletchley Declaration (2023) | First formal international recognition of existential AI risks |
| Near-miss incidents | AI-caused harms could motivate stronger cooperation | No major incidents yet | Academic research suggests 15-25% probability of major AI incident by 2030 |
| Scientific consensus | Growing expert agreement on risk severity | AI Safety Summit series building evidence base | UN Scientific Panel on AI proposed Sept 2024, modeled on IPCC |
| US-China dialogue | Limited technical cooperation despite broader tensions | Biden-Xi agreement Nov 2024 | Agreement to exclude AI from nuclear command/control systems |
| Factor | Mechanism | Status |
|---|---|---|
| AISI network expansion | Technical cooperation builds trust and shared methods | 11 countries + EU; $11M+ joint research funding; inaugural meeting Nov 2024 completed first joint testing exercise |
| Joint model evaluations | Practical cooperation on pre-deployment testing | US-UK-Singapore joint evaluations of Claude 3.5 Sonnet, o1; demonstrates feasible cross-border technical collaboration |
| EU AI Act extraterritoriality | "Brussels Effect" creates de facto global standards | Implementation began August 2024; prohibited practices effective Feb 2025; GPAI obligations Aug 2025 |
| UN institutional frameworks | Global governance architecture development | Scientific Panel on AI proposed Sept 2024; Global Digital Compact adopted Sept 2024; biannual intergovernmental dialogues recommended |
| Factor | Mechanism | Probability |
|---|---|---|
| Major AI incident | Catastrophic event could trigger emergency cooperation | 15-25% by 2030 |
| Capability surprise | Unexpected AI advancement could motivate precaution | 10-20% |
| International incident | AI-related conflict between states could drive agreements | 5-10% |
| Domain | Impact | Severity | Quantified Risk |
|---|---|---|---|
| Racing dynamics | Countries cut safety corners to maintain competitive advantage | Critical | 30-60% reduction in safety investment vs. coordinated scenarios |
| Regulatory arbitrage | AI development concentrates in least-regulated jurisdictions | High | Similar to tax havens; creates "safety havens" for risky development |
| Fragmented standards | Incompatible safety frameworks multiply compliance costs | High | Estimated 15-25% increase in compliance costs for multinational deployment |
| Crisis response | No mechanism for coordinated action during AI emergencies | Critical | Zero current capacity for rapid multilateral intervention |
| Democratic deficit | Global technology governed by few powerful actors | High | 2-3 countries controlling 80%+ of frontier AI development |
| Verification gaps | No credible monitoring of commitments | Critical | Unlike nuclear regime with IAEA inspections; AI lacks equivalent |
International coordination directly affects existential risk through several quantifiable mechanisms that determine whether the global community can respond effectively to advanced AI development.
Racing prevention: Without coordination, competitive dynamics between US-China or between AI labs pressure actors to deploy insufficiently tested systems. Game-theoretic modeling suggests racing conditions reduce safety investment by 30-60% compared to coordinated scenarios. Coordination mechanisms like shared safety standards administered through institutions like model registries or compute governance frameworks could prevent this "race to the bottom" by creating common compliance obligations.
Collective response capability: If dangerous AI capabilities emerge, effective response may require coordinated global action—pausing development, sharing countermeasures, or coordinating deployment restrictions. Current coordination gaps leave no rapid response mechanism for AI emergencies, despite 28 countries acknowledging catastrophic risk potential. The absence of such mechanisms increases the probability that capability surprises proceed unchecked.
Legitimacy and compliance: International frameworks provide legitimacy for domestic AI governance that purely national approaches lack, similar to how climate agreements strengthen domestic climate policy. This legitimacy increases the likelihood of sustained compliance even when politically inconvenient. Research on international organizations suggests effectiveness improves dramatically with technical levers (like ICANN's DNS control), monetary levers (IMF/WTO), or reputation mechanisms—suggesting AI governance requires similar institutional design.
| Timeframe | Key Developments | Coordination Impact |
|---|---|---|
| 2025-2026 | India hosts AI Impact Summit; CAISI mission shift; EU AI Act enforcement | Mixed—institutional building continues but US/UK divergence deepens |
| 2027-2028 | Next-gen AI systems deployed; potential incidents | Uncertain—depends on whether incidents motivate cooperation |
| 2029-2030 | Council of Europe treaty enforcement; potential new frameworks | Could crystallize into either coordination or fragmentation |
| Scenario | Probability | Outcome |
|---|---|---|
| Coordination consolidation | 20-25% | Major incident or leadership change drives renewed US engagement; binding international framework emerges by 2030 |
| Muddle through | 40-50% | Voluntary frameworks continue with mixed compliance; AISI network grows but lacks enforcement; fragmented governance persists |
| Governance bifurcation | 25-30% | US/UK pursue innovation-focused approach; EU/China/Global South develop alternative framework; AI governance splits into competing blocs |
| Coordination collapse | 5-10% | Geopolitical crisis undermines all cooperation; AI development proceeds with minimal international oversight |
The question of US-China AI cooperation represents perhaps the most critical governance uncertainty, given these two nations' dominance in AI development and their broader geopolitical rivalry.
Arguments for cooperation:
Arguments against:
The AI Safety Summit process (Bletchley 2023, Seoul 2024, Paris 2025) represents a major diplomatic investment, but its ultimate effectiveness remains contested among governance researchers.
Arguments summits are building blocks:
Arguments summits are insufficient:
International Institutional Frameworks:
US-China Cooperation Prospects:
Summit Effectiveness:
Auto-generated from the master graph. Shows key relationships.