In two years the United Nations has assembled, almost from nothing, an architecture for governing artificial intelligence. The Global Digital Compact, adopted with the Pact for the Future in September 2024, followed the recommendations of the Secretary-General’s High-level Advisory Body. It committed the UN to an Independent International Scientific Panel on AI and a Global Dialogue on AI Governance, both later established by the General Assembly, and the Secretariat gained a new Office for Digital and Emerging Technologies. The Panel’s first report, presented by its co-chairs Yoshua Bengio and Maria Ressa, opened the first Global Dialogue in July 2026. On paper, the world now has a place where every state can speak about AI.
The harder question is whether speaking is the same as governing. My argument is that the new architecture solves a problem of voice while leaving untouched a problem of authority. Unless that gap is closed, the Global South, and Africa in particular, risks gaining a seat at the table without a say in the rules.
The asymmetry the architecture inherits
The system’s founding documents are candid about the inequality it confronts. The Advisory Body found that of seven major AI governance initiatives, seven countries, all in the G7, were party to every one, while “118 countries are parties to none (primarily in the global South)” (HLAB 2024, 8). Forty-eight of the African Group’s 54 members sat in that excluded group (ibid., 43). The Scientific Panel adds that the United States “accounts for 75% of the computing power among the world’s top 500 AI supercomputers,” with China holding 15% (Scientific Panel 2026, 8). The Secretary-General reports that low-income countries hold “less than 0.1 per cent” of global co-location data-centre capacity (A/80/817, 2), and the African Union that Africa, home to some 15% of humanity, held only 1.8% of large-scale data centres in 2023 (AU 2024, 45).
These are not merely gaps in access. As the Panel puts it, most states are left “dependent on systems they cannot build, inspect, audit or fully adapt to local context” (10), and may “gain access to AI while losing practical control over its standards, safeguards and local fit” (19). H.L.A. Hart taught that a legal order is defined less by its primary rules than by its secondary rules: the rules that confer power to make, change and adjudicate the rest. The AI divide is, at bottom, a divide in secondary rules. The question is not only who benefits from AI, but who holds the competence to make, interpret and enforce the norms that govern it.
Voice without authority
Measured against that problem, the new bodies are deliberately modest. The Scientific Panel is “policy-relevant but not policy-prescriptive.” The Global Dialogue negotiates nothing: member states speak in three-minute slots, and the outcome is a co-chairs’ summary (Annotated Agenda, 4, 13). Meanwhile, the Panel concedes that “assurance of safety largely depends on developer goodwill” (15), and the Advisory Body that “compliance often rests on voluntarism; practice belies rhetoric” (7), with remedies for harm “notable primarily for their absence” (37).
Modesty may have been the price of consensus, and the UN’s legitimacy rests, as the Advisory Body notes, on its “unique inclusiveness” (47). Fritz Scharpf’s distinction is useful here: the new architecture is strong on input legitimacy, since every state may speak, but weak on output legitimacy, since it cannot deliver binding results. Thomas Franck argued that rules command compliance through their perceived fairness and coherence; a system in which the decisions that matter, about compute, model release, safety thresholds and data, are taken by a handful of firms and states answerable to none of these forums will struggle to earn either.
The limits of the capacity frame
The system’s main response to the divide is capacity-building, and above all a proposed Global Fund for AI. The latest proposal is a fund of about $3 billion over two to four years, while private investment in AI infrastructure “exceeds half a trillion United States dollars, annualized” (A/80/817, 4). Nor is the fund’s logic purely developmental: the 2025 financing report invites firms to offer services “in exchange for access to emerging markets and talent pools” (A/79/966, 18) and predicts that neglected regions “will become new markets” (19), even as it warns that such countries risk becoming “merely consumers of, or resource providers for, AI technologies designed elsewhere” (2). Amandeep Gill, the UN’s Envoy for Digital and Emerging Technologies, has acknowledged that such public-interest efforts “don’t rule out profit-making,” calling this “the beauty and the paradox of these approaches” (Open Source Week 2025, 25). Read through Couldry and Mejias’s account of data colonialism, the paradox looks less benign. Amartya Sen’s capability approach points the same way: capacity is the real freedom to shape a technology, not access to it.
Openness is often offered as the remedy, but the evidence counsels care. The UN University found that no AI system has yet met the Digital Public Goods Standard, warning that “openness can reproduce rather than reduce dependency and capability gaps” (UNU 2026, 14). Kenya’s Philip Thigo put the underlying point plainly: “Artificial intelligence is neither artificial nor intelligent. It still relies on people” (OSPOs for Good 2024, 19). An open licence without compute, local data and skilled people is an invitation few can accept.
From inclusion to co-authorship
Africa is not waiting to be included. The African Union’s Continental AI Strategy of July 2024 insists on “sovereign AI capabilities” (AU 2024, 55), independent oversight bodies with “access to redress and remedy” (34), and values such as Ubuntu. Eight of the UN Panel’s forty experts come from African countries, and UN University Rector Tshilidzi Marwala has described digital public infrastructure as “sovereign terrain” (Open Source Week 2025, 22). The question is whether the global architecture will treat such agency as input to be consulted or as authority to be shared.
Four commitments would move it toward the latter. First, participation with consequence: the Dialogue’s outputs should feed formal, reviewable recommendations to the General Assembly rather than end in a chair’s summary. Second, shared governance of the money: beneficiary states should hold real weight in any Global Fund, financed predictably rather than only voluntarily. Third, public-interest infrastructure: compute and local-language data should be treated as public goods, as the UN University recommends. Fourth, enforceable accountability: independent evaluation and access to remedy, rather than developer goodwill, should become the global baseline.
The first two years of UN AI governance have answered the question of who may speak. The next two must answer the harder questions: who decides, by what authority, and in whose interest. A governance order that cannot answer those questions for the 118 will struggle to call itself global.
Sources
HLAB 2024: UN High-level Advisory Body on AI, Governing AI for Humanity: Final Report (September 2024).
Scientific Panel 2026: Independent International Scientific Panel on AI, Preliminary Report (July 2026). Annotated Agenda: Global Dialogue on AI Governance (July 2026).
A/80/817: UN Secretary-General, Unique challenges faced by developing countries in AI capacity-building (August 2026). A/79/966: Innovative voluntary financing options for AI capacity-building (July 2025).
AU 2024: African Union, Continental Artificial Intelligence Strategy (July 2024). UNU 2026: UNU Macau, AI Systems as Digital Public Goods (2026).
OSPOs for Good 2024 and Open Source Week 2025: RISE conference reports (2025; 2026). Also consulted: UN SG, Policy Brief 5: A Global Digital Compact (2023); GA Res. 79/258 (2024); UN OSET and ILO, Mind the AI Divide (2024); OSET and UNDP, Universal DPI Safeguards Framework (2024).
H.L.A. Hart, The Concept of Law (1961); Fritz W. Scharpf, Governing in Europe: Effective and Democratic? (1999); Thomas M. Franck, The Power of Legitimacy among Nations (1990); Amartya Sen, Development as Freedom (1999); Nick Couldry and Ulises A. Mejias, The Costs of Connection (2019).