The United States faces a new technological challenge, but the history of nuclear weapons offers fewer straightforward answers than policymakers and technology executives often assume.
Artificial intelligence is increasingly being compared with nuclear weapons as governments search for ways to manage the risks associated with rapidly advancing technology. The comparison has gained support among leading technology executives, national security officials and researchers who believe the nuclear age offers lessons for regulating powerful AI systems.
Elon Musk has described artificial intelligence as potentially more dangerous than nuclear weapons. Anthropic chief executive Dario Amodei has compared advanced AI models with sensitive nuclear materials, while OpenAI chief executive Sam Altman has repeatedly referred to the International Atomic Energy Agency as a possible institutional model for AI oversight.
CIA Director John Ratcliffe has also drawn parallels between the two technologies.
The comparison is understandable. Both nuclear technology and advanced AI offer substantial benefits while creating potentially serious security risks. Both can influence international power relations, and competition for technological superiority could intensify rivalries between major powers.
However, the central argument of the Foreign Affairs analysis is that nuclear arms control should not be treated as a ready-made framework for artificial intelligence.
The history of nuclear governance is more complicated than its reputation suggests. Moreover, AI differs from nuclear technology in its physical characteristics, commercial development, potential applications and the difficulty of determining when a system becomes dangerous.
The Nuclear Order Was Less Successful Than It Appears
Proposals for international AI regulation frequently draw inspiration from the institutions established to prevent nuclear proliferation.
Some researchers have suggested an AI arrangement resembling the Nuclear Non-Proliferation Treaty, under which access to advanced computing equipment would depend on countries meeting agreed regulatory requirements.
Others favour an international inspection organisation with powers comparable to those of the International Atomic Energy Agency.
A further approach seeks to adapt Cold War arms control practices to the technological competition between the United States and China.
These proposals reflect the perceived achievements of nuclear governance.
In 1953, President Dwight Eisenhower introduced his Atoms for Peace initiative, which sought to encourage peaceful nuclear technology while restricting the spread of nuclear weapons.
The Nuclear Non-Proliferation Treaty subsequently entered into force in 1970, supported by international inspections and verification arrangements.
Washington and Moscow also established communication channels, negotiated restrictions on nuclear testing and developed arms control agreements intended to reduce the risk of catastrophic confrontation.
Today, nine countries possess nuclear weapons, and no nuclear war has occurred since the bombings of Hiroshima and Nagasaki in 1945.
Yet this outcome cannot be attributed exclusively to international treaties and institutions.
The nuclear ambitions of several countries were abandoned because of political pressure, security guarantees, economic considerations or domestic political change.
South Korea and Taiwan discontinued emerging weapons programmes following substantial American pressure. Libya surrendered its programme after sustained engagement and pressure from the United States and Britain.
Sweden and Switzerland abandoned nuclear weapons ambitions partly because of their financial costs, while democratic transitions helped redirect the policies of Argentina and Brazil.
South Africa dismantled its nuclear arsenal as the end of apartheid transformed its domestic and international priorities.
Other countries, including India, Pakistan, Israel and North Korea, developed nuclear weapons despite the international non-proliferation system.
The historical record therefore suggests that the nuclear order has always depended on political circumstances and the willingness of powerful states to enforce its restrictions.
Nuclear Catastrophe Was Sometimes Avoided by Chance
The absence of nuclear war since 1945 is frequently presented as evidence that deterrence and arms control have worked.
However, historical investigations have revealed numerous occasions when mistakes, miscommunication or unexpected developments brought nuclear-armed states close to disaster.
During the Cuban missile crisis of October 1962, the United States and Soviet Union approached a direct nuclear confrontation.
American military planners considered invading Cuba without fully understanding the extent of Soviet nuclear deployments on the island.
A Soviet submarine nearly launched a nuclear torpedo against American naval forces, while an American reconnaissance aircraft entered Soviet airspace and created another dangerous confrontation.
Several of these incidents unfolded without the direct knowledge of President John F. Kennedy or Soviet leader Nikita Khrushchev.
Later decades brought additional warning-system failures and false alarms.
In November 1979, an American early-warning system mistakenly interpreted a training tape as evidence of a major Soviet attack.
In September 1983, Soviet officer Stanislav Petrov correctly identified an apparent incoming American missile strike as a system malfunction.
These incidents demonstrate that nuclear institutions did not eliminate the possibility of catastrophic mistakes.
The analysis argues that policymakers developing AI regulation should recognise the role of contingency and chance in nuclear history rather than assuming that existing arms control institutions provide a complete explanation for the absence of nuclear war.
The Existing Nuclear Framework Is Under Pressure
The nuclear system itself is facing renewed difficulties.
Relations between Washington and Moscow have deteriorated, while major agreements governing their arsenals have expired or collapsed.
Russia’s war in Ukraine and uncertainty over American security commitments have encouraged discussions about alternative nuclear arrangements among some European countries.
China continues to expand its arsenal without being bound by a comparable bilateral arms control framework.
In East Asia, concerns about North Korea and regional security have contributed to debates over nuclear deterrence in South Korea and Japan.
The Nuclear Non-Proliferation Treaty also faces tensions between nuclear-armed and non-nuclear states.
Under the treaty, nuclear powers are expected to pursue disarmament in good faith. Yet recent trends have involved the expansion and modernisation of nuclear arsenals.
Frustration over the slow pace of disarmament has encouraged more than 90 countries to sign a separate treaty prohibiting nuclear weapons.
Nuclear-armed states and their allies have rejected that approach, arguing that it does not adequately address existing security conditions.
These developments complicate attempts to present nuclear governance as an established model that can simply be transferred to another technology.
AI Cannot Be Monitored Like Nuclear Materials
One of the most important differences between nuclear weapons and artificial intelligence concerns the ability to track their development.
Nuclear weapons require particular materials, principally uranium and plutonium, that are expensive to produce and depend on specialised industrial facilities.
Inspectors can monitor nuclear materials, estimate production capabilities and assess whether a programme may be moving towards weapons development.
Artificial intelligence also depends on physical infrastructure, including advanced semiconductors, manufacturing equipment and large data centres.
Major AI training facilities consume substantial electricity and can be observed through their physical footprint.
This makes computing infrastructure a useful point of intervention for governments seeking to monitor advanced AI development.
However, computing capacity does not provide the same reliable indication of AI capabilities that fissile materials provide for nuclear weapons.
AI systems can become more capable through improvements in algorithms, software and training methods.
A model that requires enormous computing resources today may become substantially cheaper to reproduce in the future.
Developers can also transfer capabilities from large systems into smaller models, while additional computing resources during deployment can improve performance without changing the original training infrastructure.
Consequently, restricting access to advanced chips may slow development, increase costs and improve visibility, but it cannot fully determine which actors possess dangerous AI capabilities.
AI Models Can Be Copied and Distributed
Another challenge arises after an AI model has been trained.
Its underlying numerical parameters, known as model weights, can be copied, transferred, modified and deployed independently of the data centre in which the model was developed.
Some developers deliberately publish these weights to encourage wider use and innovation.
Such models may then be adapted for purposes that their original creators never anticipated.
A single general-purpose system could assist with medical analysis, software development, cybersecurity research or potentially dangerous technical activities.
Nuclear weapons have no close equivalent to the public distribution of a powerful, reusable system that can be modified and improved by numerous independent actors.
This makes traditional non-proliferation methods less suitable for artificial intelligence.
Even when governments restrict the supply of advanced chips, they may be unable to determine where model weights have travelled or how the resulting systems are being used.
Private Companies Complicate Government Oversight
Nuclear weapons programmes have historically operated under extensive government control.
Developing and maintaining nuclear arsenals requires state institutions, military organisations and specialised scientific infrastructure.
Advanced artificial intelligence has followed a different path.
Its development has been led largely by private companies, independent researchers and commercial laboratories.
These organisations compete for investment, market share and technological leadership, often advancing faster than governments can establish regulatory requirements.
A nuclear arms control agreement can require governments to account for weapons and facilities under their authority.
An international AI agreement would need to address companies whose systems, commercial interests and technical capabilities may not be fully understood even by their own governments.
Verification would also be difficult.
Where nuclear agreements can identify warheads, missiles, launchers and other physical equipment, AI governance must determine what should be measured: computing resources, model capabilities, training procedures, deployment conditions or access to external tools.
Reaching common definitions would be essential before countries could establish credible international compliance arrangements.
AI Regulation Lacks a Defining Historical Crisis
The political circumstances surrounding the emergence of nuclear weapons also differ sharply from those of artificial intelligence.
The atomic bombings of Hiroshima and Nagasaki provided an immediate demonstration of nuclear weapons’ destructive potential.
Subsequent nuclear tests reinforced public understanding of the technology’s dangers.
Artificial intelligence has not produced an equivalent event.
Reports of advanced AI systems bypassing safeguards or gaining unauthorised access to external environments have raised concerns among developers and policymakers.
However, the consequences of these incidents have remained limited, and their wider implications are still debated.
Potential catastrophic scenarios include the misuse of AI in biological weapons development, major cyberattacks, disruption of essential infrastructure and the loss of meaningful human control over highly advanced systems.
These risks remain difficult to communicate because they have not been demonstrated on a scale comparable to the first use of nuclear weapons.
More immediate concerns, including employment disruption, the effects of chatbots on young people and the local costs of data centres, have attracted greater public and political attention.
The challenge for policymakers is therefore to establish durable safeguards without waiting for a major disaster to generate political support.
AI Governance Must Address Development and Deployment
The analysis proposes that AI regulation should extend beyond monitoring computing infrastructure.
Governments need to examine how models are developed, tested, evaluated and deployed.
This could involve requirements for developers to disclose safety-testing procedures, report significant incidents and permit independent assessments of advanced systems.
Regulation should also consider the specific environments in which AI is used.
A general-purpose model may create different levels of risk depending on whether it is assisting with routine administrative work, operating critical infrastructure or supporting sensitive biological and cybersecurity research.
Consequently, safeguards should reflect the consequences of particular applications rather than treating all AI systems as equally dangerous.
Advanced capabilities in synthetic biology and offensive cyber operations are identified as areas requiring especially strong oversight.
Washington Needs a Consistent Regulatory Framework
The article argues that the United States already possesses several instruments that could support stronger AI governance.
Export controls can restrict access to sensitive technology.
The Defense Production Act provides mechanisms through which the federal government can require certain disclosures from companies.
Federal purchasing policies could also make independent safety assessments and transparency conditions part of government contracts.
The central problem is the absence of a durable framework governing how these instruments should be used.
The Trump administration has resisted broad AI regulation on the grounds that excessive oversight could weaken American innovation and benefit China.
At the same time, it has intervened in individual companies’ activities through export restrictions and other executive measures.
Such an approach can create uncertainty for developers while leaving the public without consistent safeguards.
Federal lawmakers have introduced proposals addressing specific risks, while states including California, Illinois and New York have pursued transparency, incident-reporting and auditing requirements.
Nevertheless, the analysis maintains that the United States has not yet established a comprehensive oversight structure comparable to those governing other safety-critical industries.
Congress would need to clarify which authorities regulate advanced AI, what information developers must disclose and what powers regulators should possess.
International Cooperation Must Begin with Domestic Accountability
The prospect of discussions between President Trump and Chinese President Xi Jinping offers an opportunity for the two major AI powers to explore areas of cooperation.
However, international agreements alone cannot resolve weaknesses in domestic regulation.
The United States must first establish clearer responsibilities for overseeing its own advanced AI industry.
Without a consistent national framework, Washington may struggle to negotiate credible international commitments or verify compliance with them.
The central lesson from nuclear history is therefore not that artificial intelligence needs identical institutions.
It is that powerful technologies require governance arrangements designed around their actual characteristics, risks and political circumstances.
Nuclear weapons and artificial intelligence share the potential to transform international security, but the mechanisms needed to manage them are fundamentally different.
The article concludes that policymakers should learn from the achievements and failures of nuclear governance while developing a regulatory system specifically suited to artificial intelligence.
