AI-augmented conventional capabilities might affect strategic stability between great military powers. The nuanced, multifaceted possible intersections of this emerging technology with a range of advanced conventional weapons can compromise nuclear capabilities, thus amplifying the potentially destabilizing effects of these weapons. This article argues that a new generation of artificial intelligence–enhanced conventional capabilities will exacerbate the risk of inadvertent escalation caused by the commingling of nuclear and nonnuclear weapons. The increasing speed of warfare will also undermine strategic stability and increase the risk of nuclear confrontation. The hyperbole surrounding artificial intelligence (AI) makes it easy to overstate the opportunities and understate the challenges posed by the development and deployment of AI in the military sphere.1 Commingling and entangling nuclear and nonnuclear capabilities and the increasing speed of warfare may well undermine strategic stability.2 From what we know today about emerging technology, new iterations of AI augmented advanced conventional capabilities will compound the risk of military escalation,3 especially inadvertent and accidental escalation.4 While the potential escalation risks posed by advances in military technology have been discussed lightly in the literature, the potential of military AI to compound the risk and spark inadvertent escalation is missing.5 This article addresses how and why AI could affect strategic stability between nuclear-armed great powers (especially China and the United States) and the multifaceted possible intersections of this disruptive technology with advanced conventional capabilities.6 Toward this end, the article conceptualizes and defines military-use AI and identifies a broad portfolio of nonnuclear weapons with “strategic effects”7 along with their attendant enabling systems, including specific AI innovations that pose the greatest risks to nuclear stability.8 Rather than provide a net assessment of all of the possible ways AI could influence strategic stability, the article instead examines the possible stability enhancing and destabilizing effects in the nuclear domain using two examples: swarming autonomous weapon systems (AWS) and hypersonic weapons.9 Four core themes help conceptualize military-relevant AI.10 First, AI does not exist in a vacuum.That is, in isolation AI will unlikely be a strategic game changer. Instead, it will mutually reinforce the destabilizing effects of existing advanced capabilities, thereby increasing the speed of warfare and compressing the decision-making time frame. Second, AI’s impact on stability, deterrence, and escalation will likely be determined as much by a state’s perception of its functionality than what it is capable of doing. In the case of nuclear policy, deterrence, and strategic calculations more broadly, the perception of an adversary’s capabilities and intentions is as important as its actual capability. In addition to the importance of military force postures, capabilities, and doctrine, the effects of AI will therefore also have a strong cognitive element, increasing the risk of inadvertent escalation as a result of misperception and misunderstanding. For the foreseeable future, military AI will include a fair degree of human agency, especially in the safety-critical nuclear domain. Thus, strategic calculations on the use of force made in collaboration with machines at various levels will continue to be informed and shaped by human perceptions. Third, the increasingly competitive and contested nuclear multipolar world order will compound the destabilizing effects of AI and, in turn, increase escalation risks in future warfare between great military powers—especially China and the United States. Moreover, the potential operational and strategic advantages offered by AI-augmented capabilities could prove irresistible to nuclear-armed strategic rivals. Thus motivated, adversaries could eschew the limitations of AI, compromising safety and verification standards to protect or attempt to capture technological superiority on the future digitized battlefield.11 Finally, and related, against this inopportune geopolitical backdrop, the perceived strategic benefits of AI powered weapons will likely attract states as a means to sustain or capture the technological upper hand over rivals. The most pressing risk posed to nuclear security is, therefore, the premature adoption of unsafe, errorprone, unverified, and unreliable AI technology in the context of nuclear weapons, which could have catastrophic implications.12 Military AI applications can be broadly categorized into those that have utility at a predominately operational or strategic level of warfare.13 At the operational level, applications include autonomy14 and robotics (especially drone swarming); multi-actor interaction during red teaming and war gaming; big data–driven modeling;15 and intelligence analysis to locate and monitor mobile missiles, submarines, mines, and troops movement.16 At a strategic level, applications include (1) intelligence, surveillance, and reconnaissance (ISR) and command, control, communications, and intelligence (C3I) systems (especially in complex, adversarial, and cluttered environments);17 (2) enhanced missile defense with machine learning-augmented automatic target recognition (ATR) technology (i.e., improving target acquisition, tracking, guidance systems, and discrimination);18 conventional precision missile munitions (including but not limited to hypersonic variants) able to target strategic weapons; (3) increased speed and scope of the observation, orientation, decision, and action (OODA) loop decision-making to augment air defense and electronic warfare (especially in antiaccess/area-denial [A2/AD] environments); and (4) AI-enhanced offensive and defensive cyber capabilities (e.g., machine learning techniques to infiltrate and uncover network vulnerabilities and to manipulate, spoof, and even destroy these networks).19 While the potential strategic effects of military AI are not unique or exclusive to this technology, the confluence of several trends weighs heavily on the pessimistic side of the instability-stability ledger: the rapid technological advancements and diffusion of military AI; the inherently destabilizing characteristics of AI technology (especially heightened speed of warfare, explainability, and vulnerability to cyberattack); the multifaceted possible intersections of AI with nuclear weapons; the interplay of these intersections with strategic nonnuclear capabilities; and the backdrop of a competitive multipolar nuclear world order, which may entice states to prematurely deploy unverified, unreliable, and unsafe AI-augmented weapons into combat situations. The historical record demonstrates that security competition—motivated by the desire to control warfare—tends to be ratcheted up because of the complexity of military technology and operations over time.20 As a result, the Clausewitzian conditions of “fog and friction” will likely become a ubiquitous outcome of the uncertainties created by increasingly complex and inherently escalatory technologies. From this perspective, the acceleration of modern warfare, the shortening of the decision-making time frame, and the commingling of military systems have occurred within the broader context of the computer revolution (e.g., remote sensing, data processing, acoustic sensors, communications, and cyber capabilities).21 These overarching trends do not rely on AI and would have likely occurred whether AI were involved or not. AI is best understood, therefore, as a potentially powerful force multiplier of these developments. Put another way, military AI, and the advanced capabilities it enables, is a natural manifestation—rather than the cause or origin—of an established trend, potentially leading states to adopt destabilizing launch postures due to the increasing speed of war and commingling.22 The following three case studies ground the discussion of the core themes related to AI and the risk of inadvertent escalation to illustrate how and why military AI applications fused with nonnuclear weapons might cause or exacerbate escalation risks in future warfare. They also illuminate how these AI-augmented capabilities would work and, despite the risks associated with the deployment of these systems, why militaries might deploy them nonetheless. Because military commanders are concerned with tightly controlling the rungs on the “escalation ladder,” they should, in theory, be against delegating too much decision-making authority to machines—especially involving nuclear weapons.23 Competitive pressures between great military powers and fear that others will gain the upper hand in the development and deployment of military AI (and the advanced weapon systems AI could empower) might overwhelm these concerns, however. By way of a caveat, the cases do not assume that militaries will necessarily be able to implement these augmented weapon systems in the near term. Disagreements exist among AI researchers and analysts about the significant operational challenges faced by states in the deployment of AI-augmented weapon systems. The proliferation of a broad range of AI-augmented autonomous weapon systems (most notably drones used in swarming tactics) could have far-reaching strategic implications for nuclear security and escalation in future warfare.24 Several observers anticipate that sophisticated AIaugmented AWSs will soon be deployed for a range of ISR and strike missions.25 Even if AWSs are used only for conventional operations, their proliferation could nonetheless have destabilizing implications and increase the risk of inadvertent nuclear escalation. For example, AIaugmented drone swarms may be used in offensive sorties targeting ground-based air defenses and by nuclear-armed states to defend their strategic assets (i.e., launch facilities and their attendant C3I and earlywarning systems), exerting pressure on a weaker nuclear-armed state to respond with nuclear weapons in a use-them-or-lose-them situation. Recent advances in AI and autonomy have substantially increased the perceived operational value that military great powers attach to the development of a range of AWSs,26 potentially making the delegation of lethal authority to AWSs an increasingly irresistible and destabilizing prospect.27 That is, in an effort to defend or capture the technological upper hand in the possession of cutting-edge war-fighting assets vis-à-vis strategic rivals’ traditionally conservative militaries, states may eschew the potential risks of deploying unreliable, unverified, and unsafe AWS. Today, the main risk for stability and escalation is the technical limitations of the current iteration of AI machine learning software (i.e., brittleness, explainability, unpredictability of machine learning, vulnerability to subversion or “data poisoning,” and the fallibility of AI systems to biases).28 To be sure, immature deployments of these nascent systems in a nuclear context would have severe consequences.29 Conceptually speaking, autonomous systems will incorporate AI technologies such as visual perception, speech, facial recognition, and decisionmaking tools to execute a range of core air interdiction, amphibious ground assaults, long-range strike, and maritime operations independent of human intervention and supervision.30Currently, only a few weapon systems select and engage their targets without human intervention. Loitering attack munitions (LAM)—also known as “loitering munitions” or “suicide drones”—pursue targets (such as enemy radars, ships, or tanks) based on preprogrammed targeting criteria and launch an attack when their sensors detect an enemy’s air defense radar.31 Compared to cruise missiles (designed to fulfill a similar function), LAMs use AI technology to shoot down incoming projectiles faster than a human operator ever could and can remain in flight (or loiter) for much longer periods. This attribute could complicate the ability of states to reliably and accurately detect and attribute autonomous attacks.32 A low-cost lone-wolf unmanned aerial vehicle (UAV) would, for example, not pose a significant threat to a US F-35 stealth fighter, but hundreds of AI machine learning autonomous drones in a swarming sortie may potentially evade and overwhelm an adversary’s sophisticated defense capabilities—even in heavily defended regions such as China’s east and coastal regions.33 Moreover, stealth variants of these systems34—coupled with miniaturized electromagnetic jammers and cyberweapons—may be used to interfere with or subvert an adversary’s targeting sensors and communications systems, undermining its multilayered air defenses in preparation for drone swarms and long-range stealth bomber offensive attacks.35 In 2011, for example, MQ-1 and MQ-9 drones in the Middle East were infected with hard-to-remove malicious malware, exposing the vulnerability of US subset systems to offensive cyber.36 This threat might, however, be countered (or mitigated) by the integration of future iterations of AI technology into stealth fighters such as the F-35.37 Manned F-35 fighters will soon be able to leverage AI to control small drone swarms in close proximity to the aircraft performing sensing, reconnaissance, and targeting functions, including countermeasures against swarm attacks.38 In the future, extended endurance of UAVs and support platforms could potentially increase the ability of drone swarms to survive these kinds of countermeasures.39 Several prominent researchers have opined that, notwithstanding the remaining technical challenges as well as the legal and ethical feasibility,40 we can expect to see operational AWSs in a matter of years.41 According to former US deputy secretary of defense Robert Work, the United States “will not delegate lethal authority to a machine to make a decision” in the use of military force. 42 Work adds, however, that such self-restraint could be tested if a strategic competitor (especially China and Russia) “is more willing to delegate authority to machines than we are and, as that competition unfolds, we’ll have to make decisions on how we can best compete” (emphasis added).43 In short, pre-delegating authority to machines, and taking human judgment further out of the crisis decision-making process, might severely challenge the safety, resilience, and credibility of nuclear weapons in future warfare.44 The historical record is replete with examples of near nuclear misses, demonstrating the importance of human judgment in mitigating the risk of miscalculation and misperception (i.e., of another’s intentions, redlines, and willingness to use force) between adversaries during crises.45 Despite these historical precedents, the risks associated with unpredictable AIaugmented autonomous systems operating in dynamic, complex, and possibly a priori unknown environments remain underappreciated by global defense communities.46 Eschewing these risks, China and Russia plan to incorporate AI into unmanned aerial and undersea vehicles for swarming missions infused with AI machine learning technology.47 Chinese strategists have reportedly researched data-link technologies for “bee swarm” UAVs, particularly emphasizing network architecture, navigation, and anti-jamming military operations for targeting US aircraft carriers.48 Drones used in swarms are conceptually well suited to conduct preemptive attacks and nuclear ISR missions against an adversary’s nuclear and nonnuclear mobile missile launchers and nuclear-powered ballistic missile submarines (SSBN), along with their attendant enabling facilities (e.g., C3I and early warning systems, antennas, sensors, and air intakes).49 The Defense Advanced Research Projects Agency (DARPA), for example, is developing an autonomous surface vehicle (ASV) double outrigger, Sea Hunter, currently being tested by the US Navy to support antisubmarine warfare operations (i.e., submarine reconnaissance).50 Some observers have posited that autonomous systems like Sea Hunter may render the underwater domain transparent, thereby eroding the second-strike deterrence utility of stealthy SSBNs. The technical feasibility of this hypothesis is highly contested, however.51 On the one hand, several experts argue that deployed in large swarms, these platforms could transform antisubmarine warfare, rendering at-sea nuclear deterrence vulnerable. On the other hand, some consider such a hypothesis technically premature because (1) it is unlikely that sensors on board AWSs would be able to reliably detect deeply submerged submarines; (2) the range of these sensors (and the drones themselves) would be limited by battery power over extended ranges;52 and (3) given the vast areas traversed by SSBNs on deterrence missions, the chance of detection is negligible even if large numbers of autonomous swarms were deployed.53 Thus, significant advances in power, sensor technology, and communications would be needed before these autonomous systems have a gamechanging strategic impact on deterrence.54 However, irrespective of the veracity of this emerging capability, the mere perception that nuclear capabilities face new strategic challenges would nonetheless elicit distrust between nuclear-armed adversaries—particularly where strategic force asymmetries exist. Moreover, DARPA’s Sea Hunter demonstrates how the emerging generation of autonomous weapons is expediting the completion of the iterative targeting cycle to support joint operations, thus increasing the uncertainty about the reliability and survivability of states’ nuclear second-strike capability and potentially triggering use-them-or-lose-them situations. Conceptually speaking, the most destabilizing impact of AI on nuclear deterrence would be the synthesis of autonomy with a range of machinelearning-augmented sensors, undermining states’ confidence in the survival of their second-strike capabilities and in extremis triggering a retaliatory first strike.55 Enhanced by the exponential growth in computing performance and coupled with advances in machine learning techniques that can rapidly process data in real time, AI will empower drone swarms to perform increasingly complex missions, such as hunting hitherto hidden nuclear deterrence forces.56 In short, the ability of future iterations of AI able to predict based on the fusion of expanded and dispersed data sets and then to locate, track, and target strategic missiles such as mobile ICBM launchers in underground silos, on board stealth aircraft, and in SSBNs is set to grow.57 Combining speed, persistence, scope, coordination, and battlefield mass, AWSs will offer states attractive asymmetric options to project military power within contested A2/AD zones.73 Enhanced by sophisticated machine learning neural networks, China’s manned and unmanned drone teaming operations could potentially impede future US freedom of navigation operations in the South China Seas.74 Its air- and sea-based drones linked to sophisticated neural networks could, for example, support the People’s Liberation Army’s manned and unmanned teaming operations. Were China to infuse its cruise missiles and hypersonic glide capabilities with AI and autonomy, close-range encounters in the Taiwan Straits and the East and South China Seas would become more complicated, accident- prone, and destabilizing—at both a conventional and nuclear level.75 China is reportedly developing and deploying UUVs to bolster its underwater monitoring and antisubmarine capabilities as part of a broader goal to establish an “underwater Great Wall” to challenge US undersea military primacy. US AI-enhanced UUVs could, for example, theoretically threaten China’s nuclear ballistic and nonnuclear attack submarines.76 A new generation of AI-augmented advanced conventional capabilities will exacerbate the risk of inadvertent escalation caused by the commingling of nuclear and strategic nonnuclear weapons (or conventional counterforce weapons) and the increasing speed of warfare, thereby undermining strategic stability and increasing the risk of nuclear confrontation. This conclusion is grounded in the overarching findings that relate to how and why AI could affect strategic stability between great military powers— especially China and the United States. If a state perceives that the survivability of its nuclear forces were at risk, advanced conventional capabilities (e.g., autonomous drone swarms and hypersonic weapons) augmented with AI machine learning techniques will have a destabilizing impact at a strategic level of conflict. AI’s effect on strategic stability will likely be determined by states’ perceptions of its operational utility rather than actual capability. If an adversary underestimated the potential threat posed by nascent and especially poorly conceptualized accident-prone autonomous systems, the consequences would be severely destabilizing. Despite the speed, diverse data pools, and processing power of algorithms compared to humans, complex AI-augmented systems will still depend on the assumptions encoded into them by human engineers to simply extrapolate inferences—potentially erroneous or biased—from complexity, resulting in unintended outcomes. One of the most significant escalatory risks caused by AI is likely to be, therefore, the perceived pressure exerted on nuclear powers in the use of AI-augmented conventional capabilities to adopt unstable nuclear postures (such as launch on warning, rescinding no-first-use pledges, or nuclear war fighting), or even to exercise a preemptive first nuclear strike during a crisis. In extremis, human commanders might lose control of the outbreak, course, and termination of warfare. Further, a competitive and contested multipolar nuclear environment will likely exacerbate the potentially destabilizing influence of AI, increasing that risk of inadvertent escalation to a nuclear level of conflict between great military powers. In today’s multipolar geopolitical order, therefore, relatively low-risk and low-cost AI-augmented AWS capability—with ambiguous rules of engagement and absent a robust normative and legal framework—will become an increasingly enticing asymmetric option to erode an advanced military’s deterrence and resolve. By disrupting effective and reliable flows of information and communication between adversaries and allies and within military organizations, AI-augmented conventional weapon systems (i.e., C3I, early warning systems, and ISR) could complicate escalation management during future crisis or conflict— especially involving China and the United States. A prominent theme that runs through the scenarios in this article—and central to understanding the potential impact of AI for strategic stability and nuclear security—is the concern that AI systems operating at machine speed will push the pace of combat to a point where machine actions surpass the cognitive and physical ability of human decision-makers to control or even comprehend events. Effective deterrence depends on the clear communication of credible threats and consequence of violation between adversaries, which assumes the sender and recipient of these signals share a common context allowing for mutual interpretation.103 For now, it remains axiomatic that human decisions escalate a situation; however, military technology like AI that enables offensive capabilities to operate at higher speed, range, and lethality will move a situation more quickly up the escalation rungs, crossing thresholds that can lead to a strategic level of conflict. These escalatory dynamics would be greatly amplified by the development and deployment of AI-augmented tools functioning at machine speed. Military AI could potentially push the pace of combat to a point where the actions of machines surpass the cognitive and physical ability of human decision-makers to control (or even fully understand) future warfare. Thus, until experts can unravel some of the unpredictable, brittle, inflexible, unexplainable features of AI, this technology will continue to outpace strategy, and human error and machine error will likely compound one another—with erratic and unintended effects.