The advisory body brought together experts from government, industry, academia, and civil society across every region, precisely to lend a global-governance proposal the legitimacy that a Western club or a single government cannot. Its final report, delivered ahead of the UN's 2024 Summit of the Future, diagnosed a governance gap and proposed a set of functions and institutions to fill it. Whatever its limits, it represents the closest thing to an official, universal starting point for the debate.
What the report recommended
Rather than proposing a single powerful agency, the report recommended a layered set of lightweight functions designed to be politically achievable. The headline recommendations included an international scientific panel on AI to build shared, policy-relevant understanding; a regular policy dialogue among governments; an exchange to help develop interoperable standards; a capacity-development network to help the Global South participate; and a small coordination office within the UN Secretariat to tie the pieces together. It also floated a global fund and a data-governance framework.
The logic behind the modesty
The report's caution is deliberate and worth understanding. It reflects a judgment that, given how divided states are, the achievable first step is not a binding regulator but the connective infrastructure — shared science, dialogue, and standards — on which stronger governance could later be built. The scientific panel in particular draws on the ozone and climate experience, where an authoritative shared assessment (the IPCC being the model) created the common factual basis that made agreement possible. In this reading, the report is laying foundations, not raising walls.
- A scientific panel to give the world a shared, credible picture of AI capabilities and risks — the factual basis any governance depends on.
- A policy dialogue to keep governments talking and to prevent the fragmentation of rival regional approaches.
- A standards exchange to work toward interoperability rather than a patchwork of incompatible national rules.
- Capacity building to bring the Global South into governance as a participant, not an afterthought — a precondition for legitimacy.
The gap on catastrophic risk
For all its authority, the report has a conspicuous limitation from a safety perspective: it treats AI governance largely as a matter of managing a broad spectrum of societal harms — bias, disinformation, inequality, labour disruption — and gives comparatively little weight to the specific danger of losing control of a superintelligent system. There is no proposed red line, no compute threshold, no verification regime, no prohibition. The instruments it recommends are for coordination and understanding, not for constraint.
This gap matters because the connective infrastructure the report proposes, valuable as it is, is not designed to do the one thing frontier safety most requires: prevent the development of systems too dangerous to control. A scientific panel and a policy dialogue can build the shared understanding that makes a binding agreement possible, but they cannot substitute for it. The UN process is necessary groundwork and insufficient by itself — a description that fits much of the current governance landscape.
The UN's blueprint builds the table everyone can sit at, and that is genuinely valuable. But a table is not an agreement. The hardest question — how to stop the development of systems no one can control — is the one the mainstream process has not yet been willing to put on it.
Naoto Nakada, Founder · Nakada Foundation to Save Humanity
Why it still matters
The advisory body's work is important precisely because of its legitimacy and reach. It establishes, at the highest level of the international system, that AI governance is a global concern requiring global institutions — and it proposes the shared-science and dialogue mechanisms that could, over time, build the common understanding a binding treaty requires. The realistic path is to treat the UN framework as the inclusive foundation on which more demanding, safety-focused governance is layered: to use the scientific panel and policy dialogue it proposes to move the mainstream conversation toward the catastrophic risks it currently underweights, and toward the red lines and verification that only a binding agreement can supply. The UN built the forum. The task is to raise the ambition of what happens inside it.
Common questions.
A multistakeholder body appointed by the UN Secretary-General in 2023, with experts from government, industry, academia, and civil society across all regions, to chart a course for global AI governance. Its 2024 final report, 'Governing AI for Humanity', delivered ahead of the Summit of the Future, is the most authoritative attempt yet to sketch a universal governance architecture.
A layered set of lightweight, politically achievable functions rather than a single powerful regulator: an international scientific panel on AI to build shared understanding, a regular intergovernmental policy dialogue, a standards exchange for interoperability, a capacity-development network to include the Global South, and a small coordination office in the UN Secretariat. It also floated a global fund and a data-governance framework.
A focus on catastrophic risk. The report treats AI governance mainly as managing a broad spectrum of societal harms — bias, disinformation, inequality — and gives little weight to the specific danger of losing control of a superintelligent system. It proposes no red line, compute threshold, verification regime, or prohibition. Its instruments are for coordination and shared understanding, not for constraining dangerous development.
No. The connective infrastructure it proposes — shared science, dialogue, standards — is valuable groundwork that can build the common understanding a binding agreement requires, but it is not designed to prevent the development of systems too dangerous to control. It is best treated as the inclusive foundation on which more demanding, safety-focused governance, including red lines and verification, must be layered.