The Geneva Digital Week opened July 6 with the inaugural Global Dialogue on AI Governance and the AI for Good Global Summit, where the United Nations’ new Independent International Scientific Panel on Artificial Intelligence presented to governments its first global scientific assessment of AI.
The gathering caps three years in which the U.N. has produced an impressive volume of work on AI, from the Global Digital Compact to the Governing AI for Humanity report; from UNESCO’s recommendation on the ethics of AI to the International Telecommunication Union’s annual summits. Read together, this work shares a single posture in which the U.N. treats AI as something to be received, a downstream resource to be channeled toward beneficial ends, aligned with the Sustainable Development Goals, monitored for societal effects, fitted with ethical guardrails.
This is the demand side of technology, and it’s where all the U.N.’s substantive engagement currently sits.
The supply side, or the places where frontier AI is produced, evaluated, and released, has no meaningful U.N. presence at all. There is no multilateral body with technical staff who can examine a laboratory’s work, no arrangement for evaluating training runs, no shared infrastructure for incident reporting across borders.
The governing architecture for the next several decades is consolidating right now in bilateral arrangements between frontier labs and the governments that host them, in private entities like Anthropic’s Project Glasswing and in export-control decisions by the parties hosting the technology.
Once those institutional facts are established, the path of least resistance for every subsequent decision will be to extend them rather than to build a multilateral alternative.
The pattern is visible in the news cycle. The U.S. Commerce Department recently authorized the release of Anthropic’s most-able AI model to roughly 100 American institutions, two weeks after an export-control suspension had taken it offline for everyone.
Both the suspension and the selective release were decided by the White House. The approved partners are American, but the populations affected by the technology are global.
European officials have publicly expressed frustration at this new dependence on decisions made in Washington, D.C. The authority to determine which populations receive access to a frontier-safety technology now resides, by default, in a single national administration.
What would a supply-side role for the U.N. look like? Pre-deployment evaluation of AI models is happening. Britain’s AI Security Institute and the U.S. Center for AI Standards and Innovation (formerly the U.S. AI Safety Institute), both test frontier models through voluntary agreements with the major laboratories, a practice so routine that the labs themselves cite it as evidence of responsible development.
The question that the multilateral system has so far failed to ask is why the safety assurance available to two countries should not be available to the other 191 states across the world.
The U.N. has done this level of technical work before. The International Atomic Energy Agency (IAEA) was created in 1957 to engage directly with the places where nuclear material is made, and its safeguards give it a standing technical presence in countries that have signed the Nuclear Non-Proliferation Treaty. The agency works inside the facilities where nuclear material is handled, rather than from outside the industry that handles it, and the populations of countries without nuclear programs receive assurance from a system that none of them could build on their own.
The AI frontier labs have reasons of their own to engage. They are already submitting models to voluntary evaluation, because it produces a credential they can cite and a defense against the argument that frontier development is happening without oversight.
A multilateral arrangement extends the same logic, giving the labs a guarantee they can offer to markets and governments beyond the two countries that currently provide it and protecting them from the regulatory fragmentation that country-by-country deals would otherwise produce. They have an interest in being trusted by the world rather than only by Washington and London.
The difficulty is that frontier AI production is concentrated in the U.S. and China, and neither shows much appetite for opening its labs to a multilateral presence.
The same was true of nuclear safeguards in 1953, four years before the IAEA was operational, when a similar concentration of capability in two rival powers, then Washington and Moscow, appeared to make any arrangement politically impossible. The agency was built because both sides eventually concluded that mutual visibility was preferable to mutual opacity, and the short window between 1953 and 1957 turned out to be the formative period during which the architecture was set for the next 70 years.
The UN80 reform agenda is the obvious vehicle through which supply-side oversight could emerge, since the initiative is grappling with how to refocus the U.N. Secretariat for future multilateralism, and the AI governance gap is exactly the structural absence a reform process is meant to address.
The U.N. cannot exercise authority on behalf of the billions of people it serves if it is present only on the side where the consequences of frontier technology are felt and absent from the side where those consequences are produced.
Filling that gap does not require a new agency on day one: A UN80 mandate for a small standing-evaluation unit, assembled from assets the U.N. already owns, such as the International Computing Center, and an invitation to the labs to extend to the center the voluntary agreements they already honor in London and Washington would be enough to begin. That is the work the next phase of multilateral engagement will have to take up.
A version of this essay originally appeared on PassBlue.
