The countries that have historically determined whether international technology governance succeeds are not always the most powerful ones. The Montreal Protocol on ozone depletion was driven primarily by Canada, the Nordic countries, and the European Community — not by the United States, which was initially resistant. The Mine Ban Treaty (1997), which banned anti-personnel landmines, was negotiated entirely outside the major-power framework, driven by Canada and a coalition of civil society organizations, and was ratified by 164 countries despite the non-participation of the United States, China, and Russia. The International Criminal Court was established over sustained US opposition.
Middle powers matter in international governance. They are not using their leverage on AI governance effectively yet. Understanding what that leverage is and how it has worked historically is the starting point for changing that.
Who the relevant middle powers are
In the AI governance context, middle powers are not simply mid-sized countries. They are countries with significant AI regulatory capacity, significant AI market influence, or both — that are not the United States or China.
The European Union is the largest regulatory actor. The EU AI Act, which entered into force in 2024, creates binding obligations on AI systems deployed in EU markets and has compliance implications for any company seeking to operate across the EU single market. This is the Brussels Effect at work: EU regulatory standards become de facto global standards because companies prefer maintaining a single global compliance approach over separate standards for different markets. The EU's governance influence on AI is already substantial, regardless of what the US or China do at the treaty level.
The United Kingdom hosts the AI Safety Institute, the first national body established specifically to evaluate frontier AI systems for safety properties. The UK hosted the Bletchley AI Safety Summit in November 2023, the first major multilateral gathering specifically focused on frontier AI safety risks. Post-Brexit, the UK has positioned AI safety as a domain where it exercises international influence independently of the EU — a role it has pursued through science diplomacy and institutional innovation rather than regulatory market size.
Japan, South Korea, and Australia are major technology economies closely aligned with the United States. Japan is hosting the G7 AI governance process that produced the Hiroshima AI Process principles. South Korea co-hosted the follow-up AI Safety Summit in Seoul in 2024. Australia has been active in developing national AI safety frameworks. Together, these three countries represent significant AI market share and regulatory capacity across the Asia-Pacific and have ongoing relationships with both the US and, more cautiously, China that create dialogue channels not available to Western countries directly.
Canada has a long tradition of multilateral technology governance engagement and was central to the Ottawa Process on landmines — the most successful case in recent history of a middle-power coalition establishing binding international norms over major-power objection. Canada is active in AI governance discussions but has not yet assembled a comparable coalition-building effort around frontier AI safety.
What leverage middle powers have
Middle powers in AI governance have four specific forms of leverage, none of which depends on waiting for the US and China to agree.
Regulatory market access is the most immediate. Countries with significant AI markets can condition access on compliance with safety standards. If the UK, Canada, Japan, South Korea, and Australia aligned their national AI regulatory frameworks with each other and with the EU, they would collectively represent a market covering most of the world's advanced economies outside the US and China. Companies seeking to operate in those markets would face regulatory pressure to meet the governance standards those countries set. This is market-based governance, not treaty-based governance, but it is binding on companies through the operation of domestic law.
Coalition signaling changes the negotiating environment for US-China discussions. When Canada, the Nordic countries, and the EU were aligned on ozone governance in the 1980s, the political cost for the United States of remaining outside the Montreal Protocol increased in a measurable way. Domestic constituencies — businesses affected by non-participation, scientific communities, allied governments — raised the price of the hold-out position. A credible, aligned middle-power coalition on AI safety would impose analogous political costs on both the US and China for remaining outside a governance framework.
Canada convened the 1997 Mine Ban Treaty negotiation in Ottawa deliberately outside the Conference on Disarmament, where the US, China, and Russia had blocked progress. The resulting treaty was ratified by 164 countries and shaped military doctrine even in non-signatory states. A frontier AI safety coalition willing to proceed without major-power agreement could follow the same path.
Technical capacity building creates the institutional infrastructure that a formal verification regime would need. The UK AI Safety Institute, the US AI Safety Institute established in part after the UK model, and equivalent bodies in other countries are building evaluation methodology and institutional experience with frontier AI safety assessment. A network of national AI safety institutes with shared evaluation standards and data-sharing arrangements would function as a de facto verification infrastructure — without requiring treaty negotiation to begin. When formal governance obligations arrive, the institutional capacity to implement them would already exist.
International institutional hosting gives middle powers ongoing influence over governance bodies. The OPCW is in The Hague. The IAEA is in Vienna. The locations are not incidental: the host countries have ongoing relationships with these institutions and some influence over their development. A Global AI Monitoring Agency, if established, will need to be housed somewhere. Countries that position themselves as credible, neutral hosts for such a body gain an avenue for long-term governance influence that goes beyond their raw AI development capacity.
The Ottawa Treaty model
The Mine Ban Treaty (1997) was negotiated through a process explicitly designed to circumvent major-power obstruction. Canada's Foreign Minister Lloyd Axworthy convened a negotiating conference in Ottawa in October 1996 that deliberately bypassed the Conference on Disarmament, where the United States, China, and Russia had blocked progress on a landmine ban. The treaty was finalized in roughly a year by a coalition of willing states supported by a civil society campaign. One hundred sixty-four countries have ratified it.
The United States, China, Russia, India, and Pakistan have not joined. But the treaty established a powerful norm: the use of anti-personnel landmines became internationally stigmatized in ways that affected military doctrine and procurement even in non-signatory states. US military planning in subsequent years showed the influence of the norm in contexts where the US was not legally bound. The treaty without the major powers changed behavior by the major powers.
A frontier AI safety process modeled on Ottawa would accept non-participation by the US and China in the first instance, while establishing binding norms among willing countries sufficient to shape commercial behavior. It would not solve the frontier AI governance problem by itself, because a treaty that does not bind the organizations doing the most frontier AI development has a critical gap. But it would create institutional infrastructure, political momentum, and demonstrated feasibility that could eventually draw in the major powers — exactly as the Mine Ban Treaty created a norm environment that has progressively raised the cost of landmine use.
What a middle-power coalition could do now
Several steps are available without waiting for a comprehensive multilateral treaty, and without requiring US or Chinese participation at the outset.
A common safety evaluation standard for frontier AI systems. If the UK AI Safety Institute, the EU's AI regulatory body, and equivalent bodies in Canada, Japan, South Korea, and Australia adopted shared standards for evaluating frontier AI models, companies would need to meet those standards to operate in their markets. This creates governance through market access rather than through treaty obligation, and it is achievable in the near term without diplomatic breakthroughs.
A data-sharing arrangement among national AI safety institutes. The UK AI Safety Institute has conducted evaluations of frontier AI models with government access to findings not publicly available. A data-sharing arrangement among allied safety institutes would increase the collective intelligence about frontier AI safety properties and create a shared evidence base for governance decisions. The institutional relationships needed for this arrangement are already partially in place.
A commissioned technical working group on verification design. The hardest question in AI governance treaty design is how to verify compliance with declared obligations. A coalition of middle powers could commission authoritative technical work on verification system design, producing an analysis that would accelerate treaty negotiations when political conditions allow. This work needs to happen before negotiations begin, not during them.
Direct engagement with China through bilateral channels that do not require the US-China frame. Japan and South Korea have their own complex but sustained technology relationships with China. The EU has regulatory engagement with Chinese technology companies through the Digital Markets Act and related frameworks. These channels provide dialogue about AI governance norms that operate independently of US-China dynamics and that may produce a different kind of engagement than bilateral US-China discussions currently do.
"The countries with the most to gain from effective AI governance are not necessarily the ones with the most power in the current competition. Middle powers have both the incentive and the capacity to build the governance architecture that the major powers have not yet agreed to."
Naoto Nakada, Founder · Nakada Foundation to Save Humanity
The two-power framing of AI governance understates the available options. A world in which the US and China have not agreed on binding AI safety obligations is not a world in which governance is impossible. It is a world in which the countries between them need to do more of the work — as they have done for ozone governance, for landmines, for chemical weapons, and for the International Criminal Court. The precedents are there. The question is whether the coalition will form before the capabilities that make governance necessary are already beyond the reach of the institutions that could govern them.
Common questions.
The Brussels Effect refers to the tendency for EU regulatory standards to become de facto global standards because companies prefer a single global compliance framework over maintaining separate approaches for different markets. The EU's GDPR produced this effect in data protection. The EU AI Act is expected to produce a similar effect for AI: companies subject to EU AI regulations have incentives to apply compliant development practices globally rather than maintaining non-compliant practices for non-EU markets. This is governance by market access rather than treaty obligation.
Yes. The Ottawa Process on landmines is the clearest example: a coalition led by Canada established the 1997 Mine Ban Treaty outside the major-power framework, and 164 countries ratified it despite non-participation by the US, China, and Russia. The treaty significantly influenced military doctrine even in non-signatory states. The Kimberley Process on conflict diamonds, driven by Canada and African producer nations, established standards that major diamond trading nations eventually adopted. Both show that middle-power coalitions can establish consequential international norms without major-power agreement at the outset.
The AI Safety Institute is a UK government body established in November 2023 following the Bletchley AI Safety Summit. It conducts evaluations of frontier AI models, assessing safety properties, and publishes findings. It was the first national body created specifically to evaluate frontier AI systems, and its model has been replicated in the United States and other countries. For governance purposes, it is building the institutional capacity and evaluation methodology that a binding verification regime would need — the technical infrastructure that formal treaty obligations require but cannot themselves create.
A binding treaty without US and Chinese participation would have significant norm-setting value but would not govern the organizations responsible for the most advanced AI development. The more realistic near-term goal is a treaty ratified by the EU, UK, Canada, Japan, South Korea, and Australia, which creates regulatory pressure on AI companies in those markets, which in turn creates incentives for US and Chinese companies and governments to engage. A treaty without the major powers is a step toward a treaty with them — as the Mine Ban Treaty demonstrated — not a substitute for one.