The UN’s 2024 “Governing AI for Humanity” report proposed a scientific panel, standards exchange, capacity network, fund, data framework, and small AI office, but no binding enforcement power. With advanced development concentrated in a few companies and countries, the debate is whether governments should grant a global institution authority over audits, incidents, or licensing.
Frontier models can be deployed globally before any one regulator understands the failure. The UN advisory body found only 7 of 193 member states participate in all seven major AI initiatives, while 118 participate in none. A narrow international body could coordinate evaluations, incident reporting, and inspections for the largest training runs without regulating ordinary software.
Nuclear and aviation systems have international institutions because failures cross borders. AI needs a technical floor that does not depend on where a company is headquartered. The UN report's seven recommendations are a useful start, but they are non-binding. Authority over shared tests and disclosure would turn coordination into evidence governments can act on.
The Biological Weapons Convention has no verification mechanism. None. Countries simply promise, and the treaty is widely considered toothless because of it. UN That is the precedent the AI debate should stare at. Everyone now proposing a purely advisory AI body is rebuilding the BWC's central flaw on purpose: an institution that can watch but cannot check. The lesson of bioweapons is not that verification is hard. It is that skipping it guarantees irrelevance. If we create an international AI body, give it inspection authority over the largest training runs from day one, or admit we are building theater.
Skeptics say AI cannot be inspected, but the supply chain says otherwise. Frontier models require enormous compute, and advanced chips come from a tiny number of suppliers and fabs. That concentration makes verification technically feasible in a way treaties rarely enjoy: you can count data centers, meter power usage, and track chip shipments. The UN advisory body's recommendations already point toward shared scientific capacity and monitoring. Nuclear inspectors verify enrichment facilities they cannot fully see inside; AI inspectors could verify compute thresholds the same way. An international body with real authority would not need to read every model's weights. It would need the power to audit the handful of facilities where frontier training is physically possible.
We have built this institution before. In 1954, countries that had recently been at war pooled money and scientists to create CERN, shared infrastructure for physics too expensive and too important for any one nation. CERN Frontier AI evaluation is the same shape of problem: the compute costs billions, the expertise is scarce, and every country needs trustworthy answers about what these systems can do. A small international body running shared evaluations and incident reporting is CERN for the one technology that will define the century. We know how to do this. We just have to decide AI deserves the same seriousness we gave particle physics.
AI changes faster than treaties. A binding body could freeze today's model thresholds and give incumbents a compliance moat. The OECD principles are non-binding, but that flexibility lets governments adapt them to local law. Use international groups for shared research and incident databases; keep licensing and penalties with accountable domestic institutions.
An “IAEA for AI” sounds precise until countries must agree on what counts as frontier, who may inspect private systems, and how to protect trade secrets. The UN body's 39 members from 33 countries could recommend science and standards, but enforcement would be dominated by the states with the most leverage. Build national capacity before creating a remote regulator.
A body with real authority over AI faces an impossible choice: be toothless or be captured. If inspections are voluntary, the countries racing hardest simply build outside the system, and the body becomes a club for the compliant. If inspections are mandatory, no great power will grant foreign inspectors access to its most strategic technology, so the authority exists only on paper. Either way, the institution ends up serving whoever funds and staffs it, likely the current tech leaders, who then write the thresholds to suit themselves. The OECD principles stay non-binding precisely because binding commitments on this technology are not credible yet. Build shared safety science first. Authority without credible enforcement is theater, and theater is what gets people hurt.