Introduction
On 8 September 2026, a safety researcher at Anthropic resigned publicly. Jacob Coxon did not claim to have witnessed the company compromise on safety. What he said was that the race between leading AI labs, Anthropic and OpenAI, was producing conditions in which corners could eventually be cut. "These will soon be superhuman systems," he wrote. The competitive pressure he described was structural: companies racing to build the most advanced model, with the pace of development outrunning the development of the frameworks to govern it. Within days, reports of model breakouts in controlled test environments had circulated. The UN High Commissioner for Human Rights issued an open letter on AI risks in surveillance, policing, and public services. The debate was everywhere, urgent, technical, and almost entirely being conducted between the people building AI and the people regulating it in wealthy economies.
What was missing from that debate was the perspective of the other six billion. The countries that are not building AI but will live under its governance. The governments whose regulators will implement rules they did not write. The populations whose welfare systems, policing environments, and public services will be shaped by decisions made in Washington, Brussels, and Beijing. For those countries, the urgency of the AI safety debate is real but the risk they face is not the same one being discussed. It is the risk of governance without representation: of arriving in a world where the rules have already been written, in frameworks they had no hand in making.
118 countries were absent from major international AI governance initiatives. Fewer than one in three developing countries have a national AI strategy. The window for concerted action is closing.
The Safety Argument Is Serious. It Is Also Incomplete.
The risks that safety advocates are raising are not imaginary. Alignment failures, systems that pursue objectives their designers did not intend, are a genuine engineering problem. The concentration of AI development in a handful of companies across two countries creates power asymmetries with real implications for accountability. The UN High Commissioner for Human Rights, Volker Türk, was direct: AI "is used to watch us, judge us, and control us" in ways that people often cannot challenge. Surveillance at scale, predictive policing, algorithmic decision-making in welfare and immigration these are not hypothetical risks. They are operational realities in systems already running.
These concerns apply with particular force in developing countries. Where accountability frameworks are weaker, where judicial independence is limited, where civil society has less capacity to challenge AI-driven decisions, the risks of opaque algorithmic systems in public services are greater, not smaller. The HCHR's call for meaningful regulation is the right call, and it matters most in the places where governance institutions are least equipped to demand it.
But the safety debate as currently conducted is also incomplete. It is a debate between AI-producing nations about the appropriate speed and conditions of development. It is not a debate that includes the governments who will be the subjects of AI governance rather than its architects. And that exclusion is not incidental. It is structural.
118 Countries. No Seat at the Table
UNCTAD's Technology and Innovation Report 2025 documents the governance gap with precision. 118 countries, predominantly in the Global South were absent from leading international AI governance initiatives. Fewer than one in three developing countries have published a national AI strategy. Three companies, Apple, Nvidia, and Microsoft each hold a market value roughly equivalent to the entire GDP of the African continent. One hundred firms, concentrated in the US and China, account for 40 per cent of global corporate R&D spending. The infrastructure of AI is owned by a narrow group of actors. The standards being built around it are being set by an equally narrow group of governments.
The UN Secretary-General's report A/80/817, published on 11 September 2026, sharpens this further. High-income countries held 77 per cent of global co-location data centre capacity in mid-2025. Low-income countries held less than 0.1 per cent. About half of all UN Member States had a national AI strategy, but fewer than 20 per cent of least developed countries did. The report concludes that "the window for concerted action" is closing, and calls for a Global Fund for AI Capacity-Building with initial commitments of approximately $3 billion.
The World Bank's World Development Report 2026 offers the most nuanced account of what is at stake for developing economies. AI can complement scarce expertise the report cites AI-generated weather forecasts reaching farmers where agricultural extension services are thin, AI tools expanding access to healthcare and legal guidance where professionals are few. This is a genuine and important opportunity. But the report also warns that without the right complements, reliable electricity, strong educational systems, institutions capable of governing AI procurement and deployment, and training data in local languages, the productivity benefits of AI are likely to widen the gap between high-income and developing economies rather than close it. The gains flow to those who already have the infrastructure. The risks accumulate where the infrastructure is absent.
Safety and inclusion are not competing objectives. A governance framework that excludes the majority of the world's countries is not a durable global framework.
The False Trade-Off
The governance conversation often frames safety and development inclusion as a tension: move too slowly and developing countries are denied AI's benefits; prioritise inclusion and safety standards are diluted. This framing should be rejected. It is not a trade-off. It is a design choice, and a revealing one about whose concerns are centred when the choice is framed that way.
Governance frameworks built without the Global South are not more rigorous, they are less legitimate. They produce rules designed for the infrastructure, institutional, and political contexts of wealthy economies, then applied to contexts they were never designed for. They create the conditions for regulatory arbitrage, where systems that cannot be deployed in high-income markets are deployed in lower-income ones instead. They build governance architecture on a foundation of exclusion, and governance architectures built that way do not hold.
The UN Global Dialogue on AI Governance held in 2026 showed what genuine multilateral engagement looks like and its limits. Developing countries showed up and pushed for sovereignty in AI governance. Iraq, speaking for the G77 and China, called for concrete action on disparities in AI infrastructure, technical capacity, and skills. The consensus outcome was, as analysts noted, a minimum common denominator reflecting the scale of disagreement still to be resolved. But it also demonstrated that developing-country governments are not passive recipients of AI governance. They have positions, interests, and the capacity to advocate for them when they are given a forum in which to do so.
What Must Happen Before the Architecture Closes
The window is not closed. But the institutions and governments with the power to act need to move now, not after the frameworks have consolidated.
Developing-country governments need national AI strategies, not as aspirational documents for ministerial websites, but as the operational basis for coordinated positions in international governance forums. A country that does not know what it wants from AI governance cannot advocate effectively for it. The 80 per cent of least developed countries without a strategy are not absent from these conversations because they lack opinions. They are absent because they lack the institutional capacity to form, coordinate, and project a position at the pace these processes move.
Development finance institutions and bilateral donors need to fund AI governance capacity at programme scale. Not as an item in a digital infrastructure grant. Not as a workshop in a broader technology programme. As a standalone, adequately resourced investment in the institutional capacity to understand, evaluate, negotiate, and govern AI systems. The Secretary-General's proposed $3 billion Global Fund for AI Capacity-Building is a start. It needs to be funded.
Multilateral governance processes need explicit participation requirements, not voluntary invitation. The history of international standard-setting in trade, finance, and environment is a history of frameworks written by the powerful and applied to the powerless. AI governance does not have to repeat that pattern. But it will, unless participation is designed in from the start rather than acknowledged as a gap after the fact.
Conclusion.
Jacob Coxon's resignation raised a genuine question about whether competitive pressure between AI labs is compatible with adequate safety governance. That question deserves serious engagement. So does this one: whether the governance frameworks being built to answer it will be designed to serve all the world's people, or primarily those in the countries building the technology.
These are not separate questions. A governance framework that excludes 118 countries from its design is not a safety framework. It is an arrangement of the powerful, for the powerful, in the name of safety. The risks it claims to address, accountability failures, concentration of power, systems that operate without meaningful oversight, are precisely the risks it replicates at the governance level.
AI safety and the inclusion of developing countries in AI governance are the same argument. Frameworks without the Global South are neither safe nor durable. The development community DFIs, bilateral donors, multilateral bodies, and the governments of the Global South themselves has both the responsibility and the remaining time to act on that before the architecture closes.
