What Arms Control Teaches About Verification

In September 2026, Amodei proposed seeking narrow coordination with China against AI-assisted biological-weapons development, reasoning that the danger would threaten both sides.¹ Even if governments agreed on such a prohibition, they would immediately face a second question: how could either side tell whether the other was complying?

Arms-control history shows that verification is an institution, not a promise appended to a treaty. An academic study of prospective AI treaties describes nuclear verification as a gradual build-out. Test-ban arrangements were followed by material accountancy under the Non-Proliferation Treaty and then more intrusive on-site inspection regimes as technology and political trust permitted.² The International Atomic Energy Agency's safeguards combine tracking of declared nuclear material with inspections.³ These mechanisms provided partial verification, not perfect knowledge.

The figure contrasts nuclear verification’s gradual build-out with biological verification’s collapse, showing that procedures gain credibility only when access is both acceptable and informative.

flowchart LR
  subgraph N["Nuclear verification build-out"]
    direction TB
    N1["Test-ban era"] --> N2["Material accountancy"] --> N3["On-site inspection"]
  end
  subgraph B["Biological verification collapse"]
    direction TB
    B1["Protocol negotiations"] --> B2["United States judged draft too intrusive and too weak"] --> B3["No verification protocol"]
  end

Figure: Nuclear verification accumulated institutions; biological protocol negotiations ended after the United States judged the draft too intrusive and too weak. The Biological Weapons Convention provides the cautionary contrast. Negotiations to add a verification protocol collapsed after the United States judged the draft both too intrusive, because it could expose sensitive research and trade secrets, and too weak to reliably identify violations.⁴⁵ The convention still lacks a verification mechanism. Shared opposition to biological weapons was not enough to settle what inspectors could see or what their observations would prove.

This tension sits beneath the phrase "trust, but verify."⁶ More access can improve confidence while increasing the risk of revealing legitimate secrets. Less access protects secrets while making concealment easier. A workable regime must decide what participants declare, what records they retain, when outside access is allowed, and what follows from suspicious findings. Selective inspection cannot honestly be presented as catching every violation.

The boundary is crucial. Nuclear safeguards do not prove that an AI agreement can copy nuclear institutions, and the biological failure does not prove that AI verification is impossible. The comparison supplies a diagnostic question instead: has a proposal translated political agreement into procedures that produce useful evidence without demanding access participants will predictably refuse?