When governance advocates call for a Global AI Monitoring Agency, they are describing an end state. The harder question is what the path to that end state looks like — how an international body with the mandate, staff, budget, legal authority, and technical capacity to monitor frontier AI development actually gets built. The history of the institutions that do this work for other dangerous technologies is informative, and consistently more complicated than the governance literature suggests.
What the IAEA required to become functional
The IAEA was established by international statute in 1957, two years after President Eisenhower proposed the Atoms for Peace program in his 1953 address to the UN General Assembly. The founding statute defined a mandate — to promote peaceful uses of nuclear energy while ensuring it was not diverted to weapons — and established a Board of Governors and General Conference as the agency's governing bodies. These were political achievements, not operational ones.
The IAEA conducted its first safeguards inspections in 1961, four years after its founding, covering a small number of research reactors in Western Europe. The safeguards system it applied was rudimentary by later standards: inspectors reviewed facility records and conducted accountancy of nuclear material, but had limited rights of physical access and no environmental monitoring capability. The system was adequate for peaceful nuclear research under the 1950s and 1960s technology environment; it was not adequate for detecting a determined weapons program.
The decisive expansion of IAEA capability came not from the agency's founding statute but from the aftermath of failure. Iraq's clandestine nuclear weapons program, discovered during and after the 1991 Gulf War, revealed that IAEA routine inspections had entirely missed a program that was advancing toward a nuclear weapon. The response was the Additional Protocol, adopted in 1997, which expanded inspection rights to undeclared locations and added environmental sampling as a standard tool. The IAEA that existed before 1997 and the IAEA that existed after 1997 are substantially different institutions in terms of what they can detect.
The lesson is that international monitoring institutions develop their real capabilities in response to the failures that expose the limits of their initial design. The founding architecture establishes the framework; operational effectiveness accumulates over decades through reform driven by failure.
What the OPCW required to become functional
The OPCW was established when the Chemical Weapons Convention entered into force in 1997. Unlike the IAEA, which predated its main treaty framework, the OPCW was designed from the outset as the verification body for an existing treaty with defined obligations. This gave the institution clearer initial mandates and a more focused operational scope.
The OPCW required substantial preparatory work before it could conduct its first inspections. The treaty's verification annex, which runs to hundreds of pages, specifies in detail the procedures for different types of inspection, the rights of inspectors, the obligations of inspected states, and the handling of confidential information. Translating these procedures into operational protocols, training inspector teams, establishing secure communication systems, building analytical laboratories capable of identifying chemical agents in environmental samples, and negotiating facility agreements with the first major stockpile holders all took years of work that followed the treaty's entry into force.
The OPCW's first destruction verification missions, monitoring the elimination of US and Russian chemical weapons stockpiles, began in the late 1990s and continued for more than two decades. The technical complexity of safe chemical weapons destruction, and the verification challenges it posed, required ongoing institutional adaptation throughout that period. What the OPCW does today reflects twenty-five years of operational experience that did not exist when its founding director-general took office.
The design decisions that matter most
An international AI monitoring agency would face a set of foundational institutional design decisions that would determine its long-term effectiveness. These decisions are not technical; they are political and organizational, and they need to be made correctly before the institution opens its doors.
The mandate scope is the first decision. A monitoring agency can be designed narrowly, focused on verifying declarations of frontier AI training above specified compute thresholds, or broadly, with authority to evaluate AI systems for safety properties and publish findings about the global AI safety landscape. A narrow mandate is easier to agree politically and more straightforward to implement. A broad mandate gives the institution more governance influence but requires a larger budget, more diverse staff, and more political consensus about what the agency is permitted to say. The IAEA's mandate was initially narrow and expanded; the OPCW's mandate has remained focused. Both models have succeeded in their respective contexts.
The independence architecture is the second decision. International organizations vary considerably in how insulated they are from major-power political pressure. The UN Security Council's veto structure means that the five permanent members can block any Security Council action. The IAEA Board of Governors has a weighted structure that gives major states disproportionate influence. The OPCW has a more democratic governance structure in which all states parties have equal votes in the Conference of States Parties. An AI monitoring agency whose major decisions can be blocked by the states with the largest AI programs is a weaker institution than one with an independent director-general and technical staff who can act on their professional judgment.
The IAEA employs approximately 2,500 people, including several hundred inspectors with deep nuclear engineering expertise. The OPCW employs several hundred, including chemists capable of conducting field analysis of chemical samples. An AI monitoring agency capable of evaluating frontier AI systems would need staff with expertise that currently commands very high salaries in the private sector. Building and retaining that staff within an international civil service compensation structure is one of the hardest design problems the agency would face.
The funding model is the third decision. International organizations are funded through assessed contributions from member states, proportional to their GDP and other factors, with voluntary contributions supplementing the assessed budget for specific programs. Assessed contribution budgets are negotiated annually and are subject to political pressure: member states that object to agency activities can withhold contributions or vote for budget reductions. The IAEA's regular budget is approximately 500 million USD annually; the OPCW's is roughly 70 million euros. An AI monitoring agency with meaningful technical capacity would require a budget in this range or higher, and the funding stability question — how to maintain the agency's budget through political cycles in major contributing states — requires a design answer.
The legal authority of inspectors is the fourth decision. Inspectors can only do what the treaty authorizes them to do. The IAEA's initial safeguards agreements gave inspectors limited access rights that were expanded by the Additional Protocol; inspectors operating under older agreements without the Additional Protocol still have fewer rights. An AI monitoring agency's inspectors would need clearly specified access rights — to facilities, records, hardware logs, and personnel — and the treaty establishing the agency needs to specify those rights clearly, because ambiguity will be resolved in practice in favor of the inspected party.
Building technical capacity before the treaty
One lesson from both the IAEA and OPCW is that technical capacity cannot be improvised after the institution is established. Both agencies benefited from pre-existing scientific and technical communities that could staff them: nuclear physics and nuclear engineering for the IAEA, analytical chemistry for the OPCW. The expertise needed to verify nuclear non-proliferation commitments or to identify chemical weapons agents in soil samples existed in universities and national laboratories before the international verification bodies needed it.
The equivalent technical community for AI governance verification — people capable of evaluating frontier AI systems for safety properties, auditing training records, and designing hardware monitoring protocols for large compute clusters — is very small and almost entirely located in private sector AI companies and a handful of academic institutions. Building the expertise base from which an international monitoring agency could recruit is itself a governance task that precedes the institution's founding.
This is one reason why national AI safety institutes, such as the UK AI Safety Institute established in 2023, matter for the longer-term governance project. They are building the evaluation methodology, the institutional experience, and the professional community that a future international monitoring agency would need to draw on. The technical groundwork for international monitoring is being laid, partially and imperfectly, before the political conditions for a treaty and an institution are in place.
"You cannot build a verification body in a year. The technical methods, the trained staff, the operational procedures, and the legal frameworks have to be developed over time. The time to start is now, not after the treaty is signed."
Naoto Nakada, Founder · Nakada Foundation to Save Humanity
A Global AI Monitoring Agency is not a distant utopian goal. It is an institution that requires specific design decisions, specific resource commitments, specific legal authorities, and specific technical capacities to function. The IAEA and OPCW demonstrate that such institutions can be built and can work. They also demonstrate that building them takes longer than treaty text suggests, requires sustained political investment through multiple leadership cycles, and develops its real capabilities in response to failures that expose the limits of its initial design. Governance advocates who call for a monitoring agency without engaging these institutional design questions are describing an aspiration, not a plan.
Common questions.
A proposed international body that would monitor compliance with international AI safety obligations, analogous to the IAEA for nuclear materials or the OPCW for chemical weapons. Its core functions would include receiving declarations of frontier AI development programs, conducting inspections of declared facilities, investigating allegations of non-compliance, and publishing AI safety assessments. No such body currently exists. Several governance frameworks, including the Nakada Foundation's proposed three-part plan, call for its establishment as part of a comprehensive AI safety treaty.
No authoritative estimate exists. For reference, the IAEA has an annual regular budget of approximately 500 million USD and around 2,500 staff. The OPCW has a budget of roughly 70 million euros and several hundred staff. An AI monitoring agency's costs would depend on its mandate scope, the number and complexity of facilities it inspects, and whether it maintains internal AI safety evaluation capacity. Early governance research estimates range from several hundred million to over a billion dollars annually for a credibly staffed and equipped body.
No decision has been made because no such agency exists. Location decisions for international organizations involve considerations of political neutrality, operational costs, proximity to relevant governments and industries, and the geopolitical interests of the host country. The IAEA is in Vienna; the OPCW is in The Hague. Several countries have expressed interest in hosting AI governance institutions, but none has been formally proposed as host for a monitoring agency with binding authority.
The IAEA was founded in 1957 and conducted first inspections in 1961, but the verification regime it operates today developed over forty years. The Additional Protocol, which significantly expanded inspection rights and added environmental monitoring, was adopted in 1997, four decades after the agency's founding. The decisive expansion of IAEA capability came in direct response to the discovery of Iraq's clandestine program — failure drove reform. An AI monitoring agency would develop its real capabilities over a comparable timeframe, not immediately upon establishment.