Designing Verification and Enforcement for International AI Governance: Lessons from Historical Treaties

Executive summary International attention to advanced AI governance has produced summits, coalitions, and one binding regional law, but no state or bloc with significant influence has yet agreed upon or accepted a binding multilateral compliance mechanism. This brief exists to give diplomatic delegations and policy advisors some comparative evidence to judge what makes such a mechanism work before that choice is made at a negotiating table. This brief examines five binding, multilateral treaties spanning (a) environmental governance: the Montreal Protocol (1987) and the Paris Agreement (2015), and (b) arms control: the Nuclear Non-Proliferation Treaty (1968), the Biological Weapons Convention (1972), and the Chemical Weapons Convention (1993), against a three-part test assessing the rule, verification mechanism, and consequences of the doctrines. The principal finding is that political conditions, not instrument design, are the first-order determinant of whether a strong verification and consequences mechanism is achievable at all; design choices only operate within the contours of what the negotiating context — influenced by historical, geopolitical, and geoeconomic factors — allows, which means judging those conditions accurately matters as much as choosing the right template. The second key finding is that verification only works when the governed activity or object can be measured from outside. Agreements built around intent or purpose, like the Biological Weapons Convention's general purpose criterion, cannot be verified, while agreements built around a concentrated, countable input, like the Montreal Protocol's handful of chemical producers, could be verified even without on-site inspection, a distinction that maps directly onto compute as the currently observable input in AI development. What remains uncertain is whether either finding generalizes beyond binding treaties; the case set contains no plurilateral coalition and no voluntary or norms-based arrangement, so this brief cannot say whether political conditions play the same first-order role, or whether measurability sets the same ceiling on verification, outside the treaty form examined here. Further study would extend this work to non-treaty governance pathways, and in the near term would have negotiators begin building the technical capacity to measure and verify compute now, while it remains concentrated, and bring frontier developers into the design of any verification regime from the outset rather than presenting them with one after the fact.