Search the archive.

In a separate article, colleagues and I conduct multivariate regression analysis on this data and find that nuclear inferior states are less likely to initiate MCTs and less likely to achieve compellent success.16 This chapter presents some simple descriptive statistics and cross-tabular analysis that clearly and powerfully reveal the same basic finding. I first explore the idea that nuclear superior states will be more likely to issue compellent threats. Turning to Table 5.1, I  find overwhelming support for this hypothesis. Since 1945, nuclear-armed states have issued forty-nine militarized compellent threats, and all forty-nine of these threats were against inferior countries. Seven threats were against nuclear-armed states with fewer nuclear weapons and forty-two were against states that lacked nuclear weapons altogether, but zero were directed at nuclear states with larger nuclear arsenals. This finding is stunning. It suggests the nuclear balance of power is determinative when it comes to patterns of international coercion. Since 1945, nuclear-armed states have issued zero threats against nuclear superior countries. In other words, for nuclear-armed states, nuclear superiority has been a necessary condition for the issuance of a compellent threat.

Moreover, if the United States had fewer nuclear weapons due to a smaller land leg, nuclear deterrence might also be threatened because US adversaries might be more willing to initiate and escalate militarized challenges against the United States and its allies. Nuclear deterrence has long been conceptualized as a game of nuclear brinkmanship. Although nuclear states cannot credibly threaten to launch a suicidal nuclear war, they can credibly threaten to risk one by initiating and escalating crises.12 A state’s willingness to stand firm in these crises depends, in part, on its vulnerability to nuclear war. Recent social-science research shows that states that are more vulnerable to a nuclear exchange are more likely to be targeted with military threats from nuclear-armed states and less likely to achieve their basic goals in a nuclear crisis.13 Accordingly, reducing the size of the US ICBM force would simultaneously increase US vulnerability and decrease adversary vulnerability to nuclear coercion, thereby shifting the balance of resolve toward US adversaries. Adversaries might be more willing to challenge, and stand firm in crises against, the United States, and Washington might be more eager to seek offramps in a crisis.

This resulting shift in the balance of resolve could also increase the risk of nuclear war. The most plausible pathway to nuclear war in the contemporary international security environment stems not from a bolt-from-the-blue strike but from escalation of a conventional conflict. US adversaries might employ nuclear weapons in the event of a conventional conflict in Eastern Europe, in the Indo-Pacific, or on the Korean Peninsula.14 If adversaries are more likely to initiate and escalate crises against a United States with a smaller ICBM force, as argued above, then there is also a greater risk that Washington will find itself in dangerous crises that could spiral out of control and result in a catastrophic nuclear exchange.

Crisis escalation and heightened risk conflict due to technological change

This brings us to the final concern related to the current technological transition. According to the most pessimistic interpretations, algorithms are likely to escalate diplomatic crises into military confrontations and further increase the intensity of combat. In other words, conflict and the use of violence may soon become much more common because intelligent machines are allegedly lacking in political understanding and human experience. This is a crucial topic, worthy of the highest ethical considerations and political attention. But here again, history and logic reassure against the bleakest outlook.

It is also possible, however, that autonomous systems could help countries buy time in ways that make nuclear escalation less likely. The United States and its competitors already deploy expansive sensor networks, stretching from outer space to cyberspace. But the amount of data produced by these sensing arrays threatens to overwhelm human operators today. The emergence of highly-automated data processors could change that, allowing national decision-makers to make far better sense of the cluttered battlespace. In fact, the first use of “AI” technology at the U.S. Department of Defense was for automated information processing to help monitor full-motion video drone feeds through Project Maven.131 Greater awareness and understanding of an adversary’s actions could reduce the risk of miscalculation. Leaders would be able to replace uncertainty – and a fear of the worst – with near real-time information on an adversary’s forces. Greater visibility could reassure leaders that a surprise attack was not underway, and the knowledge of this visibility would reduce incentives for a surprise attack. More advanced automation and autonomous systems could also help to improve the security, efficiency, and resiliency of military communications and command-and-control systems, which are subject to increased disruption in the cyber, electromagnetic, and physical domains. This greater resilience could reduce the vulnerability of nuclear communications and command-and-control systems to disruption.

From a nuclear policy perspective, the majority of scholarship addressing the implications of AI, with a few important exceptions,8 argues that military applications of AI will have a destabilizing effect on strategic stability and global security. Areas regularly cited as problematic include the potential for deep fakes, human bias carried over into AI programming, cyber vulnerabilities, exacerbating great power rivalry, application to nuclear command and control, security dilemmas, and potential for loss of human control.9 This focus creates a rather defeatist paradigm, particularly for policymakers confronted with rapid advances in technology and limited thinking on how it might be applied in a stabilizing way. Whereas many of these studies treat AI as a monolith, we argue that the opportunities and risks associated with AI depend on its application, and considering their impact on nuclear weapons policies requires a greater degree of granularity. Can some military applications of AI strengthen deterrence, for example? Might it facilitate advances in arms control, at a time when the international community desperately needs new ideas and new verification techniques?

Despite the potential risks inherent with adopting AI in nuclear systems, many of the potential benefits to both deterrence and arms control have been overlooked or ignored. Additionally, many in the nuclear policy community rely on the most extreme negative examples to make the case that AI is too dangerous to adopt into nuclear policies and practices. These views are short-sighted and ignore the reality that AI will increasingly play a role in all aspects of our civil and military policies—and that potential great power competitors such as China and Russia are investing heavily in AI-related military technology.10 The broader nuclear policy community should therefore start thinking now about how to harness the power of AI as a tool for greater stability, transparency, and security as well as steps that can be put in place to mitigate the most significant risk factors.

The starting assumption underlying this book is that the diffusion of AI systems appears to be fundamentally different from past experiences, notably the diffusion of information technologies (IT) into defence during the 1970s–2010s. This period, often called IT-driven Revolution in Military Affairs (RMA), was characterised by Western (primarily the US) superiority in the research and development of advanced military technologies, which in turn were gradually and selectively diffused, in varying paths and patterns, to allies and strategic partners, including small and middle powers in Europe and East Asia (Raska 2016). For the first time in decades, however, the United States is facing a strategic peer competitor – i.e., China – with its own advancing military capabilities, backed by novel technologies, that is challenging US military-technological primacy (Mahnken 2012). In doing so, the margins of military-technological advantages between great powers are narrowing, accelerating the strategic necessity for “disruptive” defence and military innovation. This intensifying strategic competition in advanced military technologies is embedded in the broader systemic competition over technological “leadership” in areas such as AI systems, robotics, cyber, additive manufacturing, advanced materials, synthetic biology, quantum computing, directed energy, space technologies, and many others. The ability to innovate is synonymous with international influence and national power – generating economic competitiveness, political legitimacy, military power, and internal security.

However, contrary to previous decades, which, admittedly, utilised some dual-use technologies to develop major weapons platforms and systems, the current AI-enabled wave differs significantly in terms of the magnitude and impact of commercial-technological innovation as a key source of military innovation (Raska 2021). Large military-industrial primes are no longer the only drivers of technological innovation; rather, advanced technologies with a dual-use potential are being developed in the commercial sectors and then being “spun on” to military applications. In this context, the diffusion of emerging technologies, including additive manufacturing (3D printing), nanotechnology, space and space-like capabilities, artificial intelligence, and drones, is not confined solely to the great powers (Hammes 2016). The proliferation of AI-enabled sensors and autonomous weapons systems is also reflected in the defence trajectories of select advanced small states and middle powers such as Singapore, South Korea, Israel, Australia, and Japan, among others. These states have the potential to develop niche emerging technologies to advance their own defence capabilities and economic competitiveness, political influence, and status in the international arena (Barsade and Horowitz 2018).

An equal danger is what analysts call “automation bias,” or the tendency for stressed-out decision-makers to trust the information and advice supplied by advanced computers rather than their own considered judgment. For example, a U.S. president, when informed of sensor data indicating an enemy nuclear attack and under pressure to make an immediate decision, might choose to accept the computer’s advice to initiate a retaliatory strike rather than consider possible alternatives, such as with Petrov’s courageous Cold War action. Given that AI data systems can be expected to gain ever more analytical capacity over the coming decades, “it is likely that humans making command decisions will treat the AI system’s suggestions as on a par with or better than those of human advisers,” a 2018 RAND study noted. “This potentially unjustified trust presents new risks that must be considered.”19

Compounding all these risks is the likelihood that China, Russia, and the United States will all install automated NC3 systems but without informing each other of the nature and status of these systems. Under these circumstances, it is possible to imagine a “flash war,” roughly akin to a “flash crash” on Wall Street, that is triggered by the interaction of competing corporate investment algorithms. In such a scenario, the data assessment systems of each country could misinterpret signs of adversary moves and conclude an attack is imminent, leading other computers to order preparatory moves for a retaliatory strike, in turn prompting the similar moves on the other side, until both commence a rapid escalatory cycle ending in nuclear catastrophe.20

U.S. power today makes the case for the triad more dubious. Survivability is no longer a feasible justification. No U.S. adversary has the capability to destroy all U.S. ballistic submarines, let alone all three legs, and there would be time to adjust if that changed. Nuclear weapons are essentially irrelevant in actual U.S. wars, which are against insurgents and weak states without nuclear arsenals. Nuclear threats have a bigger role in hypothetical U.S. wars with nucleararmed powers. But cases where the success of deterrence hinges on the U.S. capability to destroy enemy nuclear forces are far-fetched. In any case, U.S. submarines and conventional forces can destroy those forces. Even hawkish policies do not require a triad.

The Cold War experience described above makes it clear that caution should be exercised when assuming that a new technology will have such universal effects. SSBNs were not a universal source of survivable nuclear forces – just as passive acoustics did not make all SSBNs vulnerable. Looking ahead, the emerging competition between the United States and China is elevating the importance of each side’s nuclear forces, and focusing attention on their survivability and effectiveness. With this in mind, it is not the United States that should be cautious about the viability of a new generation of SSBNs, but China. As during the Cold War, the interaction between geography and technology will likely have very different consequences for the future viability of US and Chinese SSBNs.

The technologies cited by those seeing transparent oceans in the future are too vague to assess in any detail, but it is very unlikely that any of them will be able to provide persistent, ocean-wide surveillance that SOSUS did (and still does against all but the quietist submarines). More likely is the development of much shorter range, non-acoustic sensors for surveillance in shallow, coastal waters. This is the conclusion of one of the main sources cited by those arguing that the oceans are becoming more transparent, and the threat is to American SSNs operating aggressively, far forward in Chinese coastal waters, not to its SSBNs (Clark 2015).

The deeper problem with Kroenig’s analysis arises in cases of very large attacks of different sizes. Once the United States suffers tens of nuclear explosions against major cities and critical logistical hubs, and certainly once it suffers hundreds, the differences in outcomes that are implied by differences in the size of the attacks are likely to be illusory. 7 In large part this is because a standard “cookie-cutter” model that estimates the number of people killed promptly by the blast of a nuclear explosion almost certainly severely underestimates the impact of a large nuclear attack. 8 Estimating the number killed by fire and fallout would raise the numbers. Likely more important, however, is the possibility, even likelihood, of societal collapse. Highly integrated modern economies may lack the ability to recover from large nuclear attacks. A nuclear attack that destroyed the United States’ key energy systems, communication and information systems, major ports and other transportation nodes—in addition to immediately killing many tens of millions of people and the cities they live in—could lead to economic and social collapse. 9 In this case, a still larger nuclear attack would not cause significantly greater damage; difference in force size should not then matter for damage and, in turn, bargaining.

There are other possibilities that could reduce or eliminate the difference in damage between large and even larger nuclear attacks. A nuclear-winter type environmental catastrophe would have this effect. Kroenig argues that the early claims about nuclear winter were overstated and therefore that his measures of damage remain adequate (43-44). But current research still finds that the cooling effects of a large nuclear war could greatly increase the number of deaths resulting from a large nuclear attack, and that “modern climate models not only show that the nuclear winter theory is correct, but also that the effects would last for more than a decade.”10 In addition, even if more people would survive a large nuclear attack than a still larger one, we need to consider the quality of life of the surviving. These millions of survivors could be fated to lives of misery, famine, and disease, struggling for mere survival in a “smoking radiating ruin.”11 Saving those lives would have value, but far less than is implied by simply comparing the difference in survivors for large attacks. For all of these reasons—including the failure to address the possibility of societal collapse, of environmental catastrophe, and of exceedingly poor quality of life following massive nuclear war—Kroenig’s measures of the damage from large nuclear attacks are inadequate and exaggerate the difference in outcomes.

Whether this concern would apply to countervalue retaliation is unclear. China’s nuclear strategy implicitly involves city destruction, given the limited quantity and accuracy of its long-range weapons. However, qualitative differences between missile attacks and unconventional delivery modes suggest that a discrete use calculation might apply. Not least, an indiscriminate attack against civilians weeks or even months after a provocation would seem particularly cold-blooded. Nonetheless, the credibility threshold for retaliation is presumably far lower than for initiating nuclear war, and one line of thinking in particular may permit recourse to countervalue strikes despite moral qualms about them.

Counterforce capabilities are the luxury of states that spend lavishly on offensive arms, whereas a minimalist posture is the strategy of a more restrained nuclear power. In the event of a nuclear attack, members of the latter group cannot in fairness be expected to refrain from their only available means of retaliating. This would amount to penalizing the victim for adopting a more stable and responsible nuclear posture than its aggressor. Thus, any civilian deaths that result from such a state’s retaliation can be laid squarely at the feet of the initiator of the nuclear exchange.

The deeper problem with Kroenig’s analysis arises in cases of very large attacks of different sizes. Once the United States suffers tens of nuclear explosions against major cities and critical logistical hubs, and certainly once it suffers hundreds, the differences in outcomes that are implied by differences in the size of the attacks are likely to be illusory. 7 In large part this is because a standard “cookie-cutter” model that estimates the number of people killed promptly by the blast of a nuclear explosion almost certainly severely underestimates the impact of a large nuclear attack. 8 Estimating the number killed by fire and fallout would raise the numbers. Likely more important, however, is the possibility, even likelihood, of societal collapse. Highly integrated modern economies may lack the ability to recover from large nuclear attacks. A nuclear attack that destroyed the United States’ key energy systems, communication and information systems, major ports and other transportation nodes—in addition to immediately killing many tens of millions of people and the cities they live in—could lead to economic and social collapse. 9 In this case, a still larger nuclear attack would not cause significantly greater damage; difference in force size should not then matter for damage and, in turn, bargaining.

There are other possibilities that could reduce or eliminate the difference in damage between large and even larger nuclear attacks. A nuclear-winter type environmental catastrophe would have this effect. Kroenig argues that the early claims about nuclear winter were overstated and therefore that his measures of damage remain adequate (43-44). But current research still finds that the cooling effects of a large nuclear war could greatly increase the number of deaths resulting from a large nuclear attack, and that “modern climate models not only show that the nuclear winter theory is correct, but also that the effects would last for more than a decade.”10 In addition, even if more people would survive a large nuclear attack than a still larger one, we need to consider the quality of life of the surviving. These millions of survivors could be fated to lives of misery, famine, and disease, struggling for mere survival in a “smoking radiating ruin.”11 Saving those lives would have value, but far less than is implied by simply comparing the difference in survivors for large attacks. For all of these reasons—including the failure to address the possibility of societal collapse, of environmental catastrophe, and of exceedingly poor quality of life following massive nuclear war—Kroenig’s measures of the damage from large nuclear attacks are inadequate and exaggerate the difference in outcomes.

If the United States’ anxiety over nuclear terrorism is any guide, these requirements vastly overstate the threshold for credibility. After all, the fear that North Korea might transfer a nuclear weapon to terrorists has been central to the case for reversing its nuclear program. If these weapons pose a catastrophic threat in the hands of extremists, on what basis should they be considered less threatening when deployed by their original owners? In truth, Pyongyang can have confidence in its minimalist posture for two reasons. First, contrary to the emphasis placed on strategic delivery vehicles, such platforms are not necessary for nuclear retaliation. In extreme circumstances, a variety of unconventional delivery means can be used. As the late political scientist Kenneth Waltz observed, “Everybody seems to believe that terrorists are capable of hiding bombs. Why should states be unable to do what terrorist gangs are thought to be capable of?”43 Second, no arbitrary deadline exists for a state to respond to a nuclear attack. Retaliation may come weeks or even months after a first strike, providing ample time to prepare nondeployed warheads or even construct a makeshift weapon from available nuclear material. Together these concepts call into question the key assumption on which nuclear primacy rests: that a nuclear counterstrike must come immediately and in the form of ballistic missile attacks, or not at all. This questionable premise permits US leaders to entertain first strike scenarios that are wildly at odds with their apparent tolerance for risk.

Delayed—But Assured—Retaliation

A counterforce posture, especially when paired with a “launch on warning” policy, necessarily requires high launch readiness, imposing decision windows of perhaps 15–30 minutes upon receipt of satellite and radar warning of an incoming attack. The risk of a premature or mistaken launch under this model is self-evidently higher than under one designed to ride out a nuclear attack and retaliate with secondstrike forces. Nor is the potential for miscalculation limited to a splendid counterforce attack. Consider a scenario presented by Austin Long and Brendan Green in which the United States enters into a limited conventional conflict with a nuclear adversary. In this circumstance, the enemy “would have strong incentives to try and secure their nuclear forces by dispersing them, delegating launch authority, or otherwise increasing readiness.” If the United States were decisively winning, these authors suggest, “signs of [its adversary’s] increasing readiness or weapons dispersal . . . would create dangerous windows of opportunity on the US side, as American troop concentrations, American allies, or even the American homeland could be potential hostages.” Given such high stakes, they argue, “counterforce will likely have advocates in high circles during a crisis.”59

Far from endorsing these capabilities, this scenario illustrates that US counterforce systems would be the principal driver of the enemy’s anxiety about losing its weapons in the first place. Further, movements to secure one’s nuclear forces from attack may be mistaken for launch preparations, prompting a counterforce strike and transforming what had been a limited conventional war into a nuclear one. Moreover, the possibility that enemy weapons may prove elusive is no less relevant in this circumstance than in the case of a bolt-from-the-blue attack. As Michael Gerson notes of such a scenario, “In the end, if an attempted disarming first strike leaves some of the adversary’s weapons intact, the United States may have started the nuclear war that it had hoped to prevent.”60

In this example, the spoofing attack on the weapon systems’ algorithm is executed in such a way that the imagery appears to the recognition system as indistinguishable from a valid military target, escalating the situation based on a false premise that would be unlikely to fool the human eye. AI experts have proven that even when data appears accurate to AI image recognition software, these systems often hallucinate objects that do not exist.

The explainability (or “black box”) problem associated with AI applications may further compound these dynamics. Insufficient understanding of how and why AI algorithms reach a particular judgment or decision would complicate the task of determining if datasets had been deliberately compromised to manufacture false outcomes — such as attacking incorrect targets or misdirecting allies during combat.

The deeper problem with Kroenig’s analysis arises in cases of very large attacks of different sizes. Once the United States suffers tens of nuclear explosions against major cities and critical logistical hubs, and certainly once it suffers hundreds, the differences in outcomes that are implied by differences in the size of the attacks are likely to be illusory. 7 In large part this is because a standard “cookie-cutter” model that estimates the number of people killed promptly by the blast of a nuclear explosion almost certainly severely underestimates the impact of a large nuclear attack. 8 Estimating the number killed by fire and fallout would raise the numbers. Likely more important, however, is the possibility, even likelihood, of societal collapse. Highly integrated modern economies may lack the ability to recover from large nuclear attacks. A nuclear attack that destroyed the United States’ key energy systems, communication and information systems, major ports and other transportation nodes—in addition to immediately killing many tens of millions of people and the cities they live in—could lead to economic and social collapse. 9 In this case, a still larger nuclear attack would not cause significantly greater damage; difference in force size should not then matter for damage and, in turn, bargaining.

There are other possibilities that could reduce or eliminate the difference in damage between large and even larger nuclear attacks. A nuclear-winter type environmental catastrophe would have this effect. Kroenig argues that the early claims about nuclear winter were overstated and therefore that his measures of damage remain adequate (43-44). But current research still finds that the cooling effects of a large nuclear war could greatly increase the number of deaths resulting from a large nuclear attack, and that “modern climate models not only show that the nuclear winter theory is correct, but also that the effects would last for more than a decade.”10 In addition, even if more people would survive a large nuclear attack than a still larger one, we need to consider the quality of life of the surviving. These millions of survivors could be fated to lives of misery, famine, and disease, struggling for mere survival in a “smoking radiating ruin.”11 Saving those lives would have value, but far less than is implied by simply comparing the difference in survivors for large attacks. For all of these reasons—including the failure to address the possibility of societal collapse, of environmental catastrophe, and of exceedingly poor quality of life following massive nuclear war—Kroenig’s measures of the damage from large nuclear attacks are inadequate and exaggerate the difference in outcomes.

The effect of AI on nuclear strategy depends as much or more on adversaries’ perceptions of its capabilities as on what it can actually do. For instance, it is extremely technically challenging for a state to develop the ability to locate and target all enemy nuclear-weapon launchers, but such an ability also yields an immense strategic advantage. States therefore covet this capability and might pursue it irrespective of technical difficulties and the potential to alarm rivals and increase the likelihood of conflict. The case could be made on technical grounds that advanced AI would still struggle to overcome obstacles originating from data limitations and information-theoretic arguments, but the tracking and targeting system needs only to be perceived as capable to be destabilizing. A capability that is nearly effective might be even more dangerous than one that already works.

If the United States’ anxiety over nuclear terrorism is any guide, these requirements vastly overstate the threshold for credibility. After all, the fear that North Korea might transfer a nuclear weapon to terrorists has been central to the case for reversing its nuclear program. If these weapons pose a catastrophic threat in the hands of extremists, on what basis should they be considered less threatening when deployed by their original owners? In truth, Pyongyang can have confidence in its minimalist posture for two reasons. First, contrary to the emphasis placed on strategic delivery vehicles, such platforms are not necessary for nuclear retaliation. In extreme circumstances, a variety of unconventional delivery means can be used. As the late political scientist Kenneth Waltz observed, “Everybody seems to believe that terrorists are capable of hiding bombs. Why should states be unable to do what terrorist gangs are thought to be capable of?”43 Second, no arbitrary deadline exists for a state to respond to a nuclear attack. Retaliation may come weeks or even months after a first strike, providing ample time to prepare nondeployed warheads or even construct a makeshift weapon from available nuclear material. Together these concepts call into question the key assumption on which nuclear primacy rests: that a nuclear counterstrike must come immediately and in the form of ballistic missile attacks, or not at all. This questionable premise permits US leaders to entertain first strike scenarios that are wildly at odds with their apparent tolerance for risk.

Delayed—But Assured—Retaliation

A counterforce posture, especially when paired with a “launch on warning” policy, necessarily requires high launch readiness, imposing decision windows of perhaps 15–30 minutes upon receipt of satellite and radar warning of an incoming attack. The risk of a premature or mistaken launch under this model is self-evidently higher than under one designed to ride out a nuclear attack and retaliate with secondstrike forces. Nor is the potential for miscalculation limited to a splendid counterforce attack. Consider a scenario presented by Austin Long and Brendan Green in which the United States enters into a limited conventional conflict with a nuclear adversary. In this circumstance, the enemy “would have strong incentives to try and secure their nuclear forces by dispersing them, delegating launch authority, or otherwise increasing readiness.” If the United States were decisively winning, these authors suggest, “signs of [its adversary’s] increasing readiness or weapons dispersal . . . would create dangerous windows of opportunity on the US side, as American troop concentrations, American allies, or even the American homeland could be potential hostages.” Given such high stakes, they argue, “counterforce will likely have advocates in high circles during a crisis.”59

Far from endorsing these capabilities, this scenario illustrates that US counterforce systems would be the principal driver of the enemy’s anxiety about losing its weapons in the first place. Further, movements to secure one’s nuclear forces from attack may be mistaken for launch preparations, prompting a counterforce strike and transforming what had been a limited conventional war into a nuclear one. Moreover, the possibility that enemy weapons may prove elusive is no less relevant in this circumstance than in the case of a bolt-from-the-blue attack. As Michael Gerson notes of such a scenario, “In the end, if an attempted disarming first strike leaves some of the adversary’s weapons intact, the United States may have started the nuclear war that it had hoped to prevent.”60

Korda also points out that ocean transparency, to the extent that it occurs, “will not affect countries equally. And that raises some interesting questions.” For example, U.S. nuclear-powered submarines are “the quietest on the planet. They are virtually undetectable. Even if submarines become more visible in general, this may have zero meaningful effect on U.S. submarines’ survivability.”