What the Global South
Wants From International
AI Governance.

When the 2023 Bletchley AI Safety Summit concluded with a joint declaration on frontier AI risks, 28 countries signed it. Absent from the signatories were India, Brazil, Indonesia, Mexico, Saudi Arabia, Nigeria, and most of the African continent. These countries represent a substantial share of the world's population, its economic output, and its political weight in international institutions. Any AI governance framework that achieves the participation of Western countries and China while excluding this group has an incomplete mandate and a legitimacy problem that will compound over time.

This is not a new problem. Every major technology governance treaty has faced it, and the treaties that achieved genuinely broad participation resolved it through explicit accommodations for the concerns of countries that were not the primary architects of the framework.

What developing countries have consistently said about technology governance

The concerns developing countries have raised about international technology governance frameworks have been consistent across multiple domains, from nuclear energy to ozone depletion to intellectual property. They are not unreasonable concerns, and they do not require agreeing with every version of them to recognize their political importance.

The equity concern runs as follows. The problems that technology governance frameworks address are primarily created by wealthy, technologically advanced countries. The restrictions these frameworks impose apply to all countries. The effect is that developing countries bear costs for problems they did not cause and give up development options that wealthy countries already exercised before the restrictions were imposed. The NPT froze the nuclear status quo that existed in 1968; non-nuclear states agreed not to develop what nuclear states already had. The Montreal Protocol required all countries to phase out CFCs while wealthy countries had already built their economies on CFC-dependent refrigeration and air conditioning. AI governance frameworks risk the same dynamic: wealthy countries that built their economies on AI applications they developed without governance constraints ask developing countries to accept restrictions that would have applied to those applications.

The access concern accompanies it. Developing countries have typically asked, in technology governance negotiations, for guaranteed access to the beneficial applications of the technology being governed. Non-nuclear NPT parties received guaranteed access to peaceful nuclear technology. Montreal Protocol parties received access to CFC alternatives at subsidized prices. Developing countries considering AI governance participation will ask what positive benefits membership provides.

The representation concern is structural. International institutions are designed by the countries that convene them. The governance frameworks they produce reflect the priorities and threat models of those countries. Developing countries have consistently argued that their priorities — access to technology, development pathways, national regulatory sovereignty — are underrepresented in frameworks designed without them at the table.

How the Montreal Protocol solved the equity problem

The Montreal Protocol's approach to developing country participation is the most directly instructive precedent for AI governance, because the technology involved was genuinely dual-use and the transition costs were concentrated on countries that had contributed least to the problem.

The protocol established differentiated obligations through Article 5, which gave developing countries consuming less than a defined per-capita threshold of ozone-depleting substances a ten-year delay in meeting the phaseout schedules required of developed countries. This recognized that countries that had contributed little to the existing stock of atmospheric CFCs should not bear the same immediate costs as countries that had contributed substantially.

The Multilateral Fund, established in 1990, provided financial resources to Article 5 countries to cover the incremental costs of complying with the protocol. It also funded technology transfer — providing developing countries with access to CFC alternatives and the technical capacity to use them. The fund has disbursed more than four billion dollars since its establishment. It is the reason the Montreal Protocol achieved near-universal ratification: countries that would otherwise have argued that compliance was beyond their financial capacity received the resources to comply.

The result was a protocol that eventually achieved ratification by 197 parties — every UN member state. Countries that initially resisted joined once the accommodation mechanisms were in place and once it became clear that the scientific consensus on ozone depletion made indefinite non-participation politically untenable.

The NPT grand bargain and its relevance

The Nuclear Non-Proliferation Treaty's participation architecture offers a different model. The NPT's bargain was explicit and reciprocal: non-nuclear states agreed not to develop weapons in exchange for guaranteed access to peaceful nuclear technology under IAEA safeguards, security assurances from nuclear states, and a commitment by nuclear states to pursue disarmament. Non-nuclear states gave up a capability and received positive benefits in return, not merely a commitment not to be threatened.

This reciprocal structure was essential to the treaty's breadth. Countries that were technically capable of developing nuclear weapons — West Germany, Japan, Sweden, Egypt, Brazil — agreed not to in part because the treaty's peaceful use provisions gave them access to nuclear energy technology they would otherwise have needed to develop independently or do without. The governance framework was also a technology-sharing arrangement.

An AI governance framework built on this model would offer developing countries guaranteed access to AI technology, training resources, and development assistance in exchange for accepting safety monitoring obligations on any frontier AI programs they develop. Countries that accepted monitoring obligations would be entitled to the benefits of the shared governance regime, including access to AI systems and infrastructure that treaty-party nations provide preferentially to other parties.

India and the participation problem

India is the most important country for global AI governance participation that is not currently engaged with the major governance frameworks. It is the world's most populous country, has a substantial and growing technology sector, has significant AI research capacity at institutions including the Indian Institutes of Technology, and did not sign the Bletchley Declaration in 2023. India has a consistent history of non-alignment in international governance frameworks it perceives as designed by and for major powers, dating to its non-participation in the NPT.

India's position on AI governance reflects several concerns. It has argued in various international forums that AI governance should focus on present harms rather than speculative future risks, reflecting a different risk assessment than the one that animates frontier AI safety advocacy. It has raised concerns about AI governance frameworks that could constrain its own technology development. And it has a principled commitment to national regulatory sovereignty that makes binding international frameworks politically difficult domestically.

Engaging India meaningfully requires addressing these concerns directly, not simply inviting it to ratify a framework designed without it. That means including Indian representatives in the design process at early stages, taking seriously its different prioritization of present harms alongside future risks, and ensuring that any governance framework offers India tangible benefits — technology access, capacity building, representation in governance institutions — that make participation a better option than remaining outside.

What a participation-inclusive AI governance framework would offer

Drawing on the Montreal Protocol and NPT models, a frontier AI governance framework designed to achieve broad participation would include several elements that current proposals largely omit.

Differentiated compliance schedules. Countries with nascent AI development programs would receive longer timelines before monitoring obligations apply to their programs, recognizing that they are not currently the source of the frontier AI risk the treaty is designed to address.

A technology access fund. Analogous to the Montreal Protocol's Multilateral Fund, a dedicated financing mechanism would provide developing countries with access to AI systems, cloud compute, and the technical capacity to participate in AI development for beneficial purposes. Access would be conditional on treaty participation, creating a positive incentive for joining.

Meaningful representation in governance institutions. A Global AI Monitoring Agency's governance structure should give developing countries votes and leadership positions proportionate to their population and economic weight, not solely to their current AI development capacity. An institution whose board is controlled by the five or six major AI-developing nations will not be perceived as legitimate by the rest of the world.

"A governance framework that wealthy nations design for themselves and then ask the rest of the world to join on those terms has a name in the history of international relations. It is called imperialism, and it does not produce durable agreements."

Naoto Nakada, Founder · Nakada Foundation to Save Humanity

The existential risk from frontier AI, if it materializes, will not be selective about which countries it harms. A governance framework that fails to achieve broad participation because it ignored the concerns of most of the world's population is not a framework that will hold long enough to matter. The technical design of AI governance is genuinely difficult. The political design — ensuring that the framework offers enough to enough countries to build the participation it needs — is equally difficult and receives far less attention. The Montreal Protocol's designers understood both problems. The frontier AI governance conversation has not yet grappled seriously with the second one.