AI governance can learn from nuclear history, but it cannot simply inherit nuclear-era templates.

As concern about advanced AI grows, policymakers often reach for the language of arms control. The analogy is understandable. Nuclear arms control offers a familiar vocabulary: treaties, verification, monitoring, inspections, confidence-building measures, red lines, and strategic stability. It suggests that dangerous technologies can be restrained through international agreement rather than left to competition alone.

That analogy is useful. It is also incomplete.

AI is not a weapons system in the same way a missile, warhead, or enrichment facility is. It is a general-purpose capability embedded in software, compute, data, research institutions, commercial products, and military systems. It can support logistics, intelligence analysis, cybersecurity, targeting, autonomy, disinformation, biotechnology, and ordinary productivity tools. That makes traditional arms-control thinking relevant, but difficult to translate.

The question is not whether AI governance can learn from arms control. It is whether arms-control concepts can survive contact with a technology that is diffuse, dual-use, commercially driven, and constantly changing.

The Governance Activity Is Real

The current policy landscape is not empty. The United States and partner countries have promoted norms for responsible military AI and autonomy. The REAIM process has pushed states toward shared language around human responsibility, accountability, and responsible use in the military domain. The UN Convention on Certain Conventional Weapons continues to debate possible measures for lethal autonomous weapons systems. In the United States, lawmakers are beginning to focus on how the Pentagon acquires, tests, and deploys AI-enabled autonomous systems.

At the same time, technical researchers are working on the verification problem. Recent work on AI treaty verification explores compute monitoring, chip-level mechanisms, secured data-center oversight, whistleblower systems, and even cryptographic proof methods for frontier AI training. These are important developments because international agreements depend on more than shared concern. They depend on whether states can know, with enough confidence, whether others are complying.

But that is also where the gap appears.

The world is producing declarations, norms, draft texts, and verification research faster than it is producing enforceable control. There is movement, but not yet a mature governance system.

The Nuclear Analogy Has Real Value

Nuclear arms control matters because it shows that even rivals can sometimes identify shared interests in restraint. It also shows that agreements rarely work through trust alone. They require monitoring, verification, domestic institutions, technical expertise, and political incentives strong enough to make compliance more attractive than cheating.

AI governance needs that seriousness. If advanced AI systems create international security risks, then slogans about innovation or ethics will not be enough. States will need ways to signal restraint, evaluate compliance, manage uncertainty, and reduce incentives for reckless deployment.

The value of the nuclear analogy is not that AI should be governed like nuclear weapons. The value is that it forces AI policy to ask hard questions about verification, secrecy, strategic competition, and institutional capacity.

AI Is Harder to Define Than a Weapon

Arms control usually begins by defining the object being controlled. A treaty can count missiles, launchers, warheads, production facilities, or specific categories of weapons. Those definitions are never simple, but they are at least tied to physical systems.

AI does not fit that pattern cleanly. Is the object of control a model, a training run, a data center, a chip cluster, a military application, a threshold of compute, an autonomous function, or a deployment context?

Each answer creates different loopholes. A rule aimed at models may miss the infrastructure used to train them. A rule aimed at compute may miss algorithmic efficiency gains or distributed training. A rule aimed at military use may miss commercial systems that can be repurposed for conflict. A rule aimed at "autonomy" may collapse into disagreement over what level of human judgment is meaningful.

This definitional problem matters because vague arms-control language can create the appearance of agreement without producing a governable rule. A ban on "dangerous AI" is politically easier to say than to verify.

Dual Use Changes the Bargain

Many controlled weapons have civilian-adjacent technologies, but AI is unusually dual-use. The same underlying systems that improve logistics, translation, software development, cyber defense, scientific research, and medical discovery may also support surveillance, targeting, cyber operations, influence campaigns, or autonomous military systems.

That makes restraint harder. States may hesitate to accept limits that could weaken commercial competitiveness, scientific progress, or military readiness. Companies may resist rules that expose proprietary systems or slow product development. Researchers may worry that broad restrictions will chill open inquiry.

In nuclear arms control, the central bargain often concerns weapons and delivery systems. In AI, the bargain may concern the infrastructure of economic power itself: compute, talent, models, data, and deployment rights.

That does not make agreement impossible. It means the agreement has to be more precise about what kind of restraint it is asking for.

Verification Is Still the Core Problem

Verification is where the nuclear analogy becomes most important and most strained. Nuclear agreements can rely on physical inspections, national technical means, material accountancy, satellite imagery, and known production pathways. AI verification is less mature.

There is promising work underway. Researchers have proposed verification layers for large-scale AI development and deployment, including chip security features, monitoring devices, personnel-based mechanisms, audits, and compute oversight. Other work argues that zero-knowledge proof systems may eventually allow some training claims to be verified without revealing model details. Hardware-governance research is also mapping which compute-monitoring mechanisms are currently feasible and which remain speculative.

But the important word is eventually.

Many of the mechanisms most relevant to international treaty verification are not yet mature enough to carry the weight of a binding agreement. Some require new hardware. Some require intrusive data-center access. Some raise privacy, sovereignty, and abuse concerns. Some may work against companies but not against major states with incentives to evade them.

This does not mean AI verification is impossible. It means AI verification cannot be assumed. It has to be built before ambitious agreements depend on it.

What Is Currently Lacking

The missing piece is not activity. It is enforceability.

There are at least four gaps.

First, there is a definition gap. States still disagree about what should count as an autonomous weapon, what level of human control is required, and whether the focus should be on specific systems, functions, models, or uses.

Second, there is a verification gap. AI governance proposals increasingly rely on compute, chips, audits, and monitoring, but many of the technical tools needed for treaty-grade verification are still underdeveloped.

Third, there is an institutional gap. Existing diplomatic processes can produce principles and declarations, but they do not yet have the inspection authority, technical capacity, or enforcement structure associated with mature arms-control regimes.

Fourth, there is a private-sector gap. Frontier AI is not only a state project. Companies control models, infrastructure, deployment terms, and safety policies. That makes AI governance dependent on public-private coordination in a way traditional arms control was not.

These gaps do not make international AI governance hopeless. They define the work that must happen next.

What Should Translate Instead

The better lesson from arms control is not to copy treaty forms, but to identify what kind of restraint is actually verifiable. Some AI agreements may be more realistic if they target narrow and observable practices: human control over nuclear-use decisions, testing and evaluation requirements for military AI, incident reporting, limits on fully autonomous lethal targeting, compute-cluster transparency, or shared standards for high-risk deployments.

These measures may look less dramatic than a grand AI treaty. But they may be more governable.

Arms control has always depended on the relationship between the rule and the ability to know whether the rule is being followed. AI governance should begin from the same discipline. A rule that cannot be defined, monitored, or enforced may still have expressive value. But it should not be mistaken for control.

From Arms Control to Capacity Building

The next phase of international AI governance should be less about finding the perfect analogy and more about building the capacity that any serious agreement would require. That means technical verification research, trusted inspection models, secure information-sharing channels, public-private coordination, incident reporting systems, and institutions that can adapt as the technology changes.

Nuclear arms control did not begin fully formed. It developed through crises, failed proposals, experiments in verification, and decades of institutional learning. AI governance will need its own version of that learning process.

Arms control can discipline AI policy by forcing it to ask what can be defined, verified, and enforced. But AI policy will fail if it treats nuclear history as a template rather than a warning.

The lesson is not that AI cannot be governed internationally. The lesson is that governance has to match the technology it is trying to restrain.