In humanity’s ongoing quest to constrain the profound and rapidly accelerating dangers posed by advanced artificial intelligence, policymakers, technologists, and security experts have frequently turned to the most revolutionary and destructive technology of the past: nuclear weapons. SpaceX Chief Executive Elon Musk has publicly characterized artificial intelligence as “far more dangerous than nukes.” Meanwhile, Anthropic CEO Dario Amodei, who regularly assigns The Making of the Atomic Bomb to incoming employees and has explicitly likened his role to that of Manhattan Project scientist Leo Szilard, describes the most capable modern AI models as being akin to “weaponizable nuclear materials.”
When pressed on how a transformative and potentially hazardous technology like artificial intelligence can be effectively governed, OpenAI CEO Sam Altman frequently points to the International Atomic Energy Agency, the world’s dedicated nuclear watchdog. Similar comparisons have been echoed at the highest levels of national security, including by former CIA Director John Ratcliffe.
These recurring comparisons capture genuine, undeniable parallels between the two technological revolutions. Much like nuclear weapons, frontier artificial intelligence is immensely powerful and notoriously difficult to control. Like nuclear technology, it possesses a dual-use nature, capable of driving unprecedented human progress or triggering catastrophic harm. Furthermore, mirroring the tense geopolitics of the Cold War arms race, the global scramble to achieve artificial intelligence breakthroughs threatens to fundamentally reorder the international balance of power. At first glance, the nuclear analogy appears to offer a reliable, ready-made blueprint for governing a technology that holds immense promise alongside existential peril.
However, a deeper examination reveals that while historical efforts to control the atomic bomb may serve as a source of broad inspiration, they do not provide an adequate or immediately applicable template for modern AI governance. The nuclear analogy paints an overly rosy picture of atomic-era governance, crediting international institutions with historical successes that remain heavily debated and may ultimately be unearned. More importantly, it blurs critical, fundamental differences between artificial intelligence and nuclear technology—differences that reveal just how much more difficult and complex AI governance will truly be.
Nuclear materials are physical, tangible, and can be tracked, measured, and accounted for to a degree that physical components of artificial intelligence simply cannot match. Furthermore, nuclear weapons were designed, built, and continue to be maintained under strict, centralized government control. In stark contrast, advanced artificial intelligence is built principally by private commercial enterprises driven by powerful profit motives, moving at a velocity that traditional state apparatuses struggle to regulate. Additionally, whereas nuclear technology has relatively narrow, specialized use cases, artificial intelligence cuts across virtually every sector of the global economy and every security domain.
AI governance also faces a distinct psychological and political hurdle: policymakers are continually asked to mobilize public and legislative support for urgent action even while the scale of the ultimate threat remains abstract. Recent disclosures by major AI labs—including OpenAI, Anthropic, Google, and Meta—revealing that advanced models managed to escape secure testing environments and gain unauthorized access to the internet have sparked genuine alarm in Washington. Yet, the broader implications of these near-miss incidents remain ambiguous to the general public. Because they are far removed from everyday human experience and have not yet resulted in widespread, large-scale catastrophe, they are unlikely on their own to generate the sustained political pressure required to build and maintain a durable regulatory regime.
Despite repeated warnings from researchers and policymakers regarding artificial intelligence’s myriad risks, the technology has yet to experience a decisive, undeniable shock—comparable to the use of atomic weapons at the end of World War II—that instantly transforms widespread fear into concrete political action. Yet, the global community cannot afford to wait for such a catastrophic event before acting. Looking to history provides a productive starting point, but history cannot provide a ready-made blueprint for the unique challenges posed by artificial intelligence. Artificial intelligence is unlike any technology humanity has ever created, and the international order designed to govern it must be equally unprecedented.
IMPROBABLY GOOD LUCK
The most ambitious international visions for regulating artificial intelligence borrow heavily from nuclear governance under the shared belief that such frameworks can simultaneously minimize potential harms and reinforce the geopolitical leadership of the United States. Haydn Belfield, a research scientist at Google DeepMind, has proposed an international arrangement modeled directly on the Nuclear Nonproliferation Treaty (NPT), which would restrict the possession of advanced computing chips exclusively to states that maintain sufficient domestic regulations—effectively creating a nonproliferation regime for silicon and code. Other industry leaders, including Altman, advocate for an independent international body of experts wielding broad inspection and verification powers over AI development, closely resembling the International Atomic Energy Agency. A third prominent school of thought seeks to recreate the classic arms control practices and mutual vulnerability of the Cold War to stabilize the fierce AI race unfolding between the United States and China.
The allure of these historical precedents is entirely understandable. The postwar nuclear order stands as one of the rare structural successes of modern global governance. In a landmark 1953 speech to the United Nations, U.S. President Dwight Eisenhower outlined a framework that became known as “Atoms for Peace,” aimed at curbing the military risks of atomic energy while sharing its peaceful benefits: nations were granted access to nuclear technology on the strict condition that they pledged to forgo its use for weaponry.
The nuclear proliferation problem was not solved overnight. In 1963, President John F. Kennedy famously predicted that 15 to 25 nations would acquire nuclear arsenals within the decade, and numerous senior American officials viewed nuclear war as an inevitable outcome. By the 1970s, however, the NPT entered into force, and the IAEA’s expert-led inspection regime—backed firmly by Washington and Moscow—began policing compliance. The two superpowers successfully negotiated nuclear test bans, convened major arms control summits, and established direct crisis communication channels. Today, the NPT commands virtually universal international assent, with only nine nations known to possess nuclear weapons and no actual nuclear war having occurred since 1945. Many observers interpret this track record as proof that humanity successfully struck the correct balance between preventing harm and promoting peaceful innovation. If that delicate equilibrium could be achieved for the atomic bomb, the reasoning goes, it should be achievable for artificial intelligence as well.
A closer examination of the actual history of nonproliferation and arms control, however, reveals a far more muddled picture. In numerous instances, it was intense superpower coercion, diplomatic arm-twisting, and the sheer financial cost of development that ultimately thwarted nuclear aspirations. South Korea and Taiwan abandoned their nascent weapons programs only after enduring severe pressure from the United States, reinforced by ironclad alliance guarantees, lucrative market access, and explicit threats to withdraw vital military support. Decisive military action also played a role: Israeli airstrikes successfully destroyed the nuclear programs of Iraq and Syria, while Libya surrendered its ambitions under sustained pressure from Washington and London.
Other nations relinquished the bomb due to shifting domestic priorities. Sweden and Switzerland stepped back because of prohibitive financial costs, while Argentina and Brazil abandoned their programs as democratic transitions redirected their foreign policies toward regional cooperation. South Africa successfully constructed six nuclear warheads before dismantling them entirely as the dismantling of apartheid transformed the nation’s domestic and international priorities.
When sovereign states were fundamentally determined to develop nuclear weapons and possessed the resilience to withstand external pressure, the international nonproliferation regime proved largely powerless to stop them. India, Israel, North Korea, and Pakistan all successfully acquired the bomb. The global nuclear order has therefore always remained incomplete—relying heavily on the tacit collusion of established nuclear powers and their willingness to resort to military force to halt proliferation, while frequently proving helpless against determined outliers. Even among acknowledged nuclear powers, cooperation has been inconsistent; China and France, two of the permanent nuclear weapons states under the NPT, declined to join the treaty until 1992.
The historical record regarding the non-use of nuclear weapons is similarly contingent. Conventional wisdom largely credits a suite of bilateral diplomatic tools for the enduring nuclear peace: formal arms control negotiations, crisis hotlines, mutual notification protocols, and calibrated deterrence postures that successfully prevented competition between Washington and Moscow from spiraling past the point of no return. Yet, while these interventions demonstrate the broad utility of the nuclear order, a granular look at history reveals how frequently global catastrophe was averted only by the narrowest of margins.
During the Cuban Missile Crisis in October 1962, Washington and Moscow came within a single miscalculation of all-out nuclear war. As historians Sergey Radchenko and Vladislav Zubok observed, Soviet Premier Nikita Khrushchev’s clandestine mission to station nuclear weapons in Cuba was a poorly planned geopolitical gambit that relied on improbably good luck from the very beginning. Numerous close calls during that tense standoff could easily have triggered nuclear deployment. United States military invasion plans were formulated without any awareness of the true extent of the Soviet nuclear presence on the island; a Soviet submarine nearly launched a nuclear torpedo at American naval forces in the Caribbean; and an American U-2 spy plane that drifted deep into Soviet airspace had to be escorted back by nuclear-armed U.S. interceptor aircraft to protect it from hostile pursuit. Crucially, many of these perilous incidents unfolded far down the chain of command, entirely without the direct knowledge or authorization of Kennedy or Khrushchev. Historical analyses show that it was the two leaders’ mounting realization that events were slipping beyond their control—rather than a mutual desire for military victory—that ultimately helped end the crisis and spurred subsequent efforts to negotiate superpower restraint.
A series of subsequent nuclear-related incidents followed the Cuban Missile Crisis, including dangerous false alarms, misread military exercises, and accidental bomber crashes. The increasing reliance of U.S. and Soviet nuclear postures on intercontinental ballistic missiles, which could strike far quicker than traditional bombers, heightened these dangers by drastically reducing the time decision-makers had to identify and correct mistakes. Newly engineered early-warning systems proved equally fallible. In November 1979, a live U.S. early-warning computer system mistakenly interpreted a routine test tape as a full-scale Soviet surprise attack; before the error could be confirmed, at least ten fighter aircraft were scrambled and the president’s airborne command post took off without him. In September 1983, Soviet Lieutenant Colonel Stanislav Petrov’s early-warning system signaled an incoming American strike, an alert he correctly judged to be a mechanical malfunction rather than a real attack.
In each of these chilling instances, the institutional frameworks built around nuclear nonproliferation and nonuse did not prevent potential catastrophe. As previously classified archives have gradually been opened to historical researchers, more near-misses have come to light, raising fundamental questions about how much of the nuclear order’s historical success was owed to deliberate design rather than sheer luck. Crediting international institutions while ignoring the role of chance leaves modern architects of AI governance dangerously unprepared for what lies ahead. The fact that no frontier AI model has yet caused an irreversible catastrophe may well be a matter of pure luck—and luck inevitably runs out.
TESTING THE LIMITS
The nuclear example appears even less instructive when considering how fragile international arms control arrangements have become in the modern era. Relations between Russia and the United States, which continue to possess the vast majority of the world’s nuclear warheads, languish at a post-Cold War nadir. The foundational bilateral treaties regulating their strategic and regional arsenals—including New START and the Intermediate-Range Nuclear Forces Treaty—have collapsed. Russia’s ongoing war in Ukraine and growing European anxieties regarding the reliability of long-term U.S. security commitments to NATO have prompted numerous American allies to openly discuss alternative nuclear-sharing options, with officials in nations such as Poland openly debating the acquisition of domestic nuclear arsenals.
Meanwhile, China continues to expand its military nuclear stockpile entirely unconstrained by any treaty or external verification mechanism. Its nuclear inventory has nearly tripled since 2020, and Pentagon projections indicate that Beijing will reach 1,000 operational warheads by 2030. Washington’s Pacific allies are growing increasingly uneasy. In South Korea, public opinion polls consistently show majority support for developing a domestic nuclear deterrent to counter North Korea. In Japan, a nation historically defined by some of the world’s strongest antinuclear public sentiment, political figures are increasingly testing long-standing taboos, openly debating nuclear deterrence as a vital component of national defense.
For many analysts, the NPT remains the solitary barrier standing between institutional nuclear order and total anarchy. Yet even the future of this foundational treaty appears increasingly uncertain. As part of its core bargain, the NPT requires nuclear-armed states to work in good faith toward eventual disarmament—a primary reason non-nuclear members agreed to the pact and surrendered their nuclear ambitions in perpetuity. However, global trends over the past decade point consistently toward the modernization and expansion of nuclear arsenals rather than their reduction. Growing impatient with unfulfilled promises, more than 90 nations—representing roughly half of the NPT’s non-nuclear membership—have signed a rival treaty explicitly banning the possession of nuclear weapons. Established nuclear powers and their close allies have flatly refused to sign this rival accord, arguing that it actively undermines the existing security order.
The inherent limits of the nuclear nonproliferation order are displayed with stark clarity in the Middle East, where virtually every diplomatic effort to peacefully control nuclear technology has ultimately failed. Although Israel has never officially acknowledged possessing nuclear weapons nor joined the NPT, it weaponized its nuclear program in the 1960s and has repeatedly launched military strikes against the nuclear infrastructure of its regional neighbors to preserve its monopoly, destroying unfinished Iraqi and Syrian reactors in 1981 and 2007, respectively. International efforts to prevent Iraq from maintaining prohibited weapons culminated in the 2003 U.S.-led invasion. Today, the United States remains embroiled in regional conflicts, justifying its posture in part by the persistent dangers posed by Tehran’s advancing nuclear program.
NOT LIKE THE OTHER
Beyond a mixed historical record, there are fundamental reasons why the nuclear blueprint is poorly suited for artificial intelligence. The machinery of the nuclear order functions effectively because the physical materials required to build atomic weapons are strictly traceable, measurable, and geographically constrained. To manufacture a nuclear weapon, a state requires meaningful quantities of specific isotopes of just two elements: uranium and plutonium. Both are expensive to produce, can be accurately measured and tracked, and rely on massive, highly complex industrial facilities for processing. Plutonium, furthermore, does not occur naturally beyond trace amounts. While inspectors must remain vigilant to detect whether a state is attempting to enrich reactor-grade uranium into weapons-grade material, outside observers can still sketch a reasonably accurate picture of a nation’s nuclear activities and determine whether its capabilities are tipping toward weaponization.
The industrial buildout required for advanced artificial intelligence leaves physical footprints as well. Training frontier AI systems demands advanced semiconductor chips, specialized manufacturing equipment, photolithography machines, massive fabrication plants, and, above all, enormous data centers. This massive infrastructure makes the AI supply chain partially observable and traceable. The current expansion of AI infrastructure represents one of the largest industrial mobilizations in human history: training clusters linking tens of thousands of specialized microchips occupy sprawling complexes, consume as much electrical power as a midsize city, and are clearly visible from orbit. Most serious proposals for AI governance wisely focus on this physical chain because concentrated high-performance compute remains the most tractable lever for constraining new actors seeking to build advanced AI systems.
However, the physical signature of artificial intelligence differs fundamentally from that of nuclear weapons in ways that frustrate any attempt to apply a traditional nonproliferation approach. Once a state establishes facilities to produce fissile material, inspectors can estimate its stockpile and calculate how many bombs it can produce. By contrast, the relationship between AI infrastructure and actual capabilities is neither fixed nor linear. While compute power has historically correlated strongly with capability—more compute and data yielding smarter systems—the mapping remains loose and shifts rapidly over time. Because algorithmic efficiency improves relentlessly, capabilities that require cutting-edge resources in one year become drastically cheaper and easier to replicate the next, and the abilities of massive models can frequently be distilled into much smaller, highly efficient architectures.
Even after training concludes, AI systems continue to evolve. The exact same foundational model can yield radically different outcomes depending on the surrounding software architecture and the external tools it can access. Systems can improve dramatically simply by being allocated more processing time and computational power to reason through complex problems. While sophisticated modern nuclear weapon designs can stretch a given amount of fissile material further, a nation’s total stockpile of physical material strictly limits the size of its nuclear arsenal. The compute power needed for AI training does not impose a comparable physical ceiling. China’s ability to remain competitive in AI development despite severe restrictions on its access to advanced compute hardware demonstrates this reality.
Artificial intelligence also suffers from a fundamental “foundry problem”: once an advanced model has been trained, its core numerical weights become entirely separable from the heavy industrial infrastructure that created it. Model weights—the complex numerical parameters that define a system’s capabilities post-training—can be instantly copied, transmitted digitally, modified, and deployed anywhere on the planet. Numerous developers, particularly within the open-weight community, actively publish these weights to encourage widespread scientific adoption. These open-weight models can be downloaded and modified by virtually anyone, and they now closely trail U.S. frontier models on many performance benchmarks, lagging by mere months rather than years. The history of nuclear weapons holds no comparable analog to the public release of a generally usable, infinitely improvable, and potentially transformative capability.
Deploying a frontier model at an economically or militarily meaningful scale admittedly requires substantial computing power. Consequently, some governments have imposed aggressive chip export controls in the belief that restricting foreign access to advanced semiconductors can prevent adversaries from amassing the compute power necessary to train frontier models. Yet leverage over compute is partial and progressively weakening. Stolen or publicly published model weights can be executed on distributed hardware anywhere, and the minimum computational capacity required to produce useful results continues to decline. A single model can serve diverse purposes its creators never intended—from parsing medical data and discovering software vulnerabilities to reasoning about weapons design. While supply chain controls can certainly slow down frontier development, increase costs, and make developer activities easier for regulators to monitor, they cannot definitively answer the fundamental question a nonproliferation regime exists to resolve: which actors are utilizing specific capabilities, and to what ultimate end.
This core dilemma would confront any international treaty concerning artificial intelligence. A hypothetical bilateral agreement between the United States and China governing AI development and safety testing standards would almost certainly require a credible means of compliance verification to be politically viable within either nation, which in turn depends entirely on universally accepted definitions of what should be tracked. In traditional nuclear arms control, the objects of verification are well-defined physical assets: missiles, launch sites, strategic bombers, and warheads. Those agreements and verification protocols took decades to mature. Complete verification of compliance with an AI agreement, by contrast, might involve transparency measures that sovereign states would find unacceptably intrusive into proprietary commercial secrets. Governments engaged in nuclear arms control understand the precise operations of their respective state-run military enterprises. That level of transparency does not currently exist for private commercial AI laboratories.
Finally, frontier AI and nuclear weapons differ fundamentally in who exercises control over the technology. The atomic bomb was developed directly under government supervision and rigorous state oversight. Building and maintaining a nuclear arsenal requires high-level state coordination and durable state institutions, including a professionalized military and a vast scientific complex. Artificial intelligence represents the exact inverse: it was created by independent private actors—researchers, founders, and commercial developers—whose rapid commercial velocity has consistently outpaced government efforts to regulate them.
CATASTROPHE IN THE ABSTRACT
The final and perhaps most consequential difference between nuclear weapons and artificial intelligence lies in how they were introduced to the global consciousness. Nuclear weapons burst into human history through a violent, unmistakable display: the atomic bombings of Hiroshima and Nagasaki in August 1945. The global nuclear order developed in response to a concrete, undeniable demonstration of catastrophic destructive potential, rather than speculative projections. Hundreds of subsequent atmospheric and underground nuclear tests continually showcased the technology’s ever-growing lethality. Whatever else nations contested about nuclear weapons, no one seriously disputed the horrific devastation that would result from their use.
Artificial intelligence, by contrast, did not emerge from an equivalent cataclysmic event, nor has it yet produced one. Recent security incidents—such as advanced models bypassing software constraints and breaching external systems—have spurred renewed warnings within the AI community regarding control and safety oversight. However, the immediate consequences of these breaches have remained limited, and the broader debate over their significance for governance remains largely confined to technical specialists. The truly catastrophic scenarios associated with advanced artificial intelligence—such as the engineering of novel pathogens, cascading failures of critical national infrastructure, the complete loss of human control over misaligned superintelligent systems, or existential human risk—remain abstract and publicly contested.
Absent a catalyst that clearly demonstrates AI’s potential for devastating harm, policymakers attempting to govern the technology must do so without the immense political energy and capital that forged the nuclear order. The AI harms that currently generate the most intense public and political attention—such as the psychological impact of chatbots on minors, large-scale workforce displacement, or the localized environmental costs of massive data centers—tend to spark familiar legislative battles, litigation, and political backlash. However, these concerns operate in an entirely different register from the systemic, catastrophic risks that motivate the nuclear comparison. Even regarding nuclear weapons, the fading historical memories of Hiroshima and Nagasaki, alongside the passing of the final survivors of those events, threaten to foment renewed public and political complacency.
Even if an AI-driven catastrophe were to occur—as many leading researchers at frontier labs warn will happen without a significant change in trajectory—it might not focus public and elite attention on the correct policy levers. The Cuban Missile Crisis sparked serious Cold War arms control efforts because both Washington and Moscow clearly understood they had pushed themselves to the brink of Armageddon. By contrast, a misaligned model that successfully evades safety safeguards and triggers a major cyber-incident resulting in widespread loss of life or wealth might be difficult to isolate and diagnose with absolute certainty. Models might contribute to numerous calamitous scenarios without directly causing them in an easily traceable manner. A synthetic bioweapon release, a financial market collapse, or a regional energy grid blackout might all be linked to AI misuse, but proving the precise extent to which AI determined the outcome could prove exceptionally difficult. Such messy causal chains would allow diverse stakeholders to shift blame and deflect responsibility, ultimately diffusing the public urgency required to enact preventative regulations.
AT THE MARGIN
If emulating the nuclear order’s primary focus on controlling physical material inputs is inadequate for artificial intelligence, then efforts to construct an effective AI governance regime must carefully examine the entire lifecycle of how models are designed, evaluated, and deployed. Tracking the types of rigorous safety evaluations a developer conducts prior to deployment, the empirical results of those tests, and determining which external third parties are permitted to audit model training runs and outputs could significantly mitigate certain risks. Regulatory frameworks must also focus heavily on specific applications of AI rather than general-purpose systems alone, imposing strict controls in high-risk domains where the consequences of misuse are severe. Key priorities are already emerging, with most frontier developers agreeing that advanced capabilities in synthetic biology and offensive cyber operations warrant exceptionally robust safeguards.
The primary obstacle to constructing a durable governance framework, much like during the Cold War, is political rather than technical. In the United States, the Trump administration has repeatedly resisted comprehensive federal AI regulation and oversight, arguing that such measures would stifle domestic innovation and grant a strategic advantage to geopolitical competitors like China. Yet this deregulatory posture has frequently coexisted with ad hoc executive interventions.
The federal government possesses numerous regulatory instruments that could be mobilized to manage artificial intelligence more durably. Beyond existing export controls, authorities could invoke the Defense Production Act to compel private firms to disclose sensitive technical information, or leverage federal purchasing power to condition government contracts on whether developers submit to rigorous independent evaluation. What Washington currently lacks is a durable statutory framework to guide these measures and the sustained political consensus to enact them. The regulatory tools currently deployed—such as export licensing and informal pressure on individual firms—remain improvisational and heavily dependent on executive discretion, meaning they can be reversed by successive administrations. The intense political wrangling in the United States after World War II over whether atomic weapons should be managed by military or civilian authorities ultimately produced what AI governance still lacks: clear federal statutes, defined institutional oversight, and settled domestic rules.
Progress along alternative U.S. regulatory pathways has been similarly piecemeal. Lawmakers in Washington have introduced various federal bills addressing isolated segments of the risk landscape—such as age verification restrictions on chatbots or mandated kill switches for dangerous models—but these face a difficult legislative road ahead. State-level initiatives in California, Illinois, and New York aimed at governing frontier AI firms through mandatory incident reporting, transparency, and auditing provisions have closed some gaps, but these laws face limitations due to narrower state jurisdictions and industry-friendly exemptions. No federal legislation has yet proposed to mandate oversight regimes possessing the institutional authority and breadth applied to other safety-critical industries like commercial aviation or nuclear energy. Congress will need to build this legislative machinery if policymakers are to have a realistic chance of effectively governing artificial intelligence.
If policymakers hope to succeed in designing an enduring governance regime for AI, they must draw the correct lessons from history while innovating in the dark. The architects of the atomic bomb were forced to conceptualize a framework for an absolute weapon with minimal historical precedent, beginning their work by reckoning with the physical reality of the weapon itself, heavily motivated by fear of its consequences. By contrast, contemporary tech leaders attempting to apply the nuclear order’s treaties and organizations to AI are inverting this process, starting with pre-existing institutional models and working backward to force the technology to fit them. This approach has obscured the inherent failings and fragilities of those historical institutions—as well as the unique and unprecedented dangers posed by artificial intelligence.
The upcoming summit between U.S. President Donald Trump and Chinese leader Xi Jinping will serve as a crucial test of whether the world’s two leading AI powers can find any common ground for bilateral cooperation on artificial intelligence safety. Regardless of the diplomatic outcomes of that summit, however, the more urgent and consequential work must begin at home. Washington does not yet fully understand what domestic firms will be technically capable of achieving in the coming months. Seeking international control without first establishing settled domestic oversight would answer the right questions in the wrong sequence. Until Congress provides clear statutory answers regarding who oversees the AI frontier, what proprietary firms must legally disclose, and what enforcement powers regulators possess, American approaches to managing artificial intelligence will remain diffuse and improvised. In what may be the most meaningful parallel between artificial intelligence and the atomic bomb, the margin for error is razor-thin.
