What is the 'IAEA for AI' idea?

It is the proposal for an international agency that would monitor and verify compliance with rules on frontier AI development, modeled on the International Atomic Energy Agency's role in nuclear governance. The core analogy is the IAEA's founding bargain: states get access to a powerful technology in exchange for accepting international monitoring to ensure it is not misused. For AI, that would mean access to compute and vetted models in exchange for monitoring of large-scale development.

Could an AI agency really inspect private companies?

Not directly, international agencies act on states, not firms. But the IAEA model works through domestic implementation: states accept treaty obligations and then require the facilities on their territory, public or private, to submit to safeguards. An AI agency would function the same way. States party to an AI treaty would legislate inspection and reporting duties for the large compute facilities and labs operating in their jurisdiction.

What is the biggest weakness of the IAEA model for AI?

Consent and enforcement. The IAEA can only inspect where states permit, and it has no independent power to punish violations, enforcement runs through the Security Council, where vetoes often block action. North Korea simply expelled inspectors and left the treaty. Any AI agency would inherit these limits, which is why verification has to be paired with meaningful obligations and credible consequences designed into the underlying treaty.

How long did the IAEA take to become effective?

Decades. It was founded in 1957, but its safeguards system was strengthened repeatedly, most significantly with the 1997 Additional Protocol adopted after inspectors missed Iraq's covert program. The lesson for AI is twofold: institutions like this can be built and made to work, but they mature through hard experience, and ASI governance likely does not have decades, which is the case for starting to build the equivalent capacity immediately.

What an International AI Agency Could Learn From the IAEA

The IAEA was created in 1957 out of Eisenhower's 'Atoms for Peace' proposal. Its founding bargain was simple and durable: states that wanted access to peaceful nuclear technology (reactors, fuel, expertise) would accept international inspection to verify they were not diverting material into weapons. Access in exchange for accountability. That trade is the core of what any AI monitoring agency would need to offer, and it is why the IAEA is the most relevant institutional precedent on the table.

Take the slogan literally: the International Atomic Energy Agency is a real institution that has spent six decades verifying commitments about a dangerous, dual-use technology among governments that do not trust each other.

What the IAEA does that ASI governance would need

The agency's central function is safeguards: a system of material accounting, containment, surveillance, and on-site inspection designed to detect the diversion of nuclear material and to deter it by making detection likely. Inspectors visit declared facilities, verify inventories, install cameras and seals, and (under the 1997 Additional Protocol, adopted after Iraq's covert program was exposed) can demand access to undeclared sites and a much broader stream of information.

Translate that into AI and the requirements come into focus. An AI agency would need declared-facility monitoring (large compute clusters), material accounting (tracking of advanced chips), containment and surveillance analogs (audit logs, on-chip mechanisms), and the authority to investigate undeclared activity. The IAEA shows these are not utopian functions; they are the ordinary work of an established international body, refined over decades.

The three things the IAEA gets right

Verification without full disclosure. Safeguards confirm that material has not been diverted without requiring states to hand over weapon-design secrets. This scoping (verify the thing that matters, not everything) is exactly the compromise an AI regime needs to protect commercial and national-security sensitivity while still being credible.
Technical independence. The IAEA's inspectorate is professional and non-political enough that its findings carry weight even with adversaries. When the agency reports on a state's compliance, that report shapes Security Council action. An AI agency's authority would similarly rest on being trusted to call things straight.
A standing institution, not an event. Governance is continuous. The IAEA exists between crises, maintaining expertise, relationships, and monitoring infrastructure. ASI governance built only around periodic summits lacks this connective tissue.

The three things it gets wrong, or cannot do

The IAEA is also a cautionary tale. It cannot inspect what it is not allowed to inspect: North Korea withdrew from the Non-Proliferation Treaty and expelled inspectors, and the agency could do nothing to stop it. Its authority is delegated by states and evaporates when a state refuses consent. Enforcement belongs to the Security Council, where vetoes routinely paralyze response. And the agency has a dual mandate (promoting nuclear energy while policing it) that critics argue creates a conflict of interest.

For AI, these limits are instructive. A monitoring agency is only as strong as the treaty obligations behind it and the enforcement willing to back it. Verification generates the evidence; it does not supply the political will to act on it. Designing an AI agency means designing, alongside it, the obligations it verifies and the consequences for violation, the parts the IAEA cannot provide on its own.

An AI monitoring agency is the instrument that makes a treaty believable, not a substitute for a treaty. Verification is what turns a promise into a commitment.

Building on the model, not copying it

The realistic path is not to clone the IAEA but to adapt its logic. The 'Atoms for Peace' bargain becomes an 'access to compute and safe models in exchange for monitoring' bargain. Material accounting becomes chip tracking. The Additional Protocol's lesson (that declared-site inspection is not enough once a state has an incentive to cheat) becomes an argument for hardware-level assurances designed in from the start. The institution took decades to mature under Cold War pressure. ASI governance does not have decades, which is the argument for beginning to build the equivalent now, using a template that already exists.

What an International AI Agency
Could Learn From the IAEA

What the IAEA does that ASI governance would need

The three things the IAEA gets right

The three things it gets wrong, or cannot do

Building on the model, not copying it

Common questions.

Go deeper

What an International AI AgencyCould Learn From the IAEA

What the IAEA does that ASI governance would need

The three things the IAEA gets right

The three things it gets wrong, or cannot do

Building on the model, not copying it

Common questions.

Go deeper

The treaty doesn't exist yet.That's the point.

What an International AI Agency
Could Learn From the IAEA

The treaty doesn't exist yet.
That's the point.