[AI and New Civilization Standards Special Report] Military Challenge ①: The AI-Nuclear Nexus and the Future of Global Military Order

I. Introduction

With three major international conferences on Artificial Intelligence (AI) governance held in Seoul in 2024, AI has garnered significant attention within South Korean society. In March, the third Summit for Democracy was held under the theme 'AI and Digital Technologies and Democracy.' This was followed by the second AI Seoul Summit in May, and the second Summit on Responsible Artificial Intelligence in the Military Domain (REAIM) is scheduled for September 9-10. Hosting a series of international conferences discussing norms for advanced technologies in Korea presents an invaluable opportunity to enhance South Korea's standing and influence in the international community. However, for South Korea to continue exercising leadership as a global pivotal state, it must not only propose universal norms to control the indiscriminate military use of advanced technologies but also simultaneously address the existential threat posed by North Korea, which targets the entire territory of the Republic of Korea with preemptive nuclear strikes.

This special report examines the AI-Nuclear Nexus, where nuclear weapons and AI converge, which will have the most significant impact on Korea among the issues of AI's military use, due to the North Korean nuclear threat. First, it will provide an overview of what AI is and what changes occur when it is applied to the military domain. Second, it will examine how the combination of AI technology with nuclear weapon capabilities and strategies affects the existing Offense-Defense Balance. Third, based on this analysis, it will forecast how the AI-Nuclear Nexus will transform the global military order in the future.

II. The Concept and Military Use of AI: The Multiplier Effect of Increased Operational Speed

While there is no officially agreed-upon definition of AI, most research points to it as a machine capable of performing tasks that require 'human intelligence,' such as situational awareness, pattern recognition, conclusion drawing, prediction, planning, learning, and communication (Horowitz 2018, 40; Haenlein and Kaplan 2019, 5; US Department of Defense 2019, 5; Congressional Research Service 2020, 2). Since the Dartmouth Summer Research Project on Artificial Intelligence (DSRPAI) in 1956, supported by the Rockefeller Foundation, AI research has experienced 'AI Springs' and 'AI Summers.' However, in the early 1980s, major governments such as the US and UK, disappointed by AI performance that did not exceed the level of 'experienced amateurs,' cut off financial support, leading to an 'AI Winter' until the 1990s. Although IBM's Deep Blue briefly gained attention by defeating world chess champion Garry Kasparov in 1997, its limitations were evident as expert systems, based on specific rules, could not achieve performance across diverse fields beyond a particular domain (Haenlein and Kaplan 2019, 6-8).

AI research entered its second summer when Google's AlphaGo defeated Ke Jie and Lee Sedol in Go, a game considered far more complex than chess, in 2015 (Haenlein and Kaplan 2019, 9). Machine learning, using artificial neural networks (ANNs) that apply numerous weights and biases, known as parameters, to infer results based on input values, provided a new breakthrough (Bode et al. 2024, 3-5). In particular, since the mid-2010s, the explosion of accessible data due to the activation of social media, coupled with advancements in GPU chips (e.g., A100, H100, B100) suitable for machine learning due to parallel processing capabilities, has led to a dramatic improvement in computer processing power.

This has created the conditions for implementing AI models that were previously theoretical in real-world applications. Deep learning, which enables machines to classify and learn from raw data of any form, is now possible. We have entered the era of Generative AI (GAI), such as OpenAI's ChatGPT, which can instantly create stories, music, images, videos, code, and strategies simply by issuing a prompt. The emergence of generative AI has made it clear that artificial intelligence is a technology that 'enables' previously impossible feats across a wide range of domains. Therefore, AI is a general-purpose technology like electricity or internal combustion engines, rather than a military technology like tanks or submarines (Horowitz 2018, 41). In this context, AI technology can serve as a 'force multiplier' when combined with existing Command, Control, Communication, computer and Intelligence (C4I) systems and weapon systems (Johnson 2019, 150).

What changes does the development of general-purpose generative AI bring in terms of military security? Related research commonly points to 'speed.' As discussed earlier, AI based on machine learning and deep learning excels at analyzing vast amounts of information instantaneously and identifying the most effective course of action among available options in a given situation. Considering that military innovation in the 21st century has consistently focused on (1) speed, (2) distance, and (3) accuracy (Metz 2000, 73-81; Lieber and Press 2017; Schneider and Macdonald 2024, 174-177), the revolutionary change AI brings in 'speed' implies that its military application can provide a competitive edge in military innovation.

In particular, processing battlefield information at high speeds not only accelerates target detection, identification, and tracking but also enhances response times to the adversary's tactical movements, thereby creating a multiplier effect in terms of 'distance' and 'accuracy.' Furthermore, in the context of integrating national security capabilities across land, sea, air, space, cyber, and non-military domains, as emphasized by recent concepts like the US's 'integrated deterrence' or China's 'intelligentized warfare,' applying AI technology is crucial for rapidly calculating the most effective combinations for multi-domain operations (Kim Yang-gyu 2023; 2024). This suggests a high probability that the success or failure of a nation's military operations will be determined by 'which country possesses AI-based military capabilities before others.'

The US, currently holding a dominant position in the AI field, clearly demonstrates this aspect in its plans for the military use of AI. The subtitle of the US Department of Defense's 'Data, Analytics, and Artificial Intelligence Adoption Strategy,' published in June 2023, is 'Accelerating Decision Advantage.' The strategy emphasizes that the reason the US military must focus on AI is to 'enable leaders to make better decisions faster' (U.S. Department of Defense 2023, 3). This highlights the significant impact of AI technology on the speed of military operations.

Furthermore, the strategy emphasizes that achieving 'decision advantage' as pursued by the US requires (1) battlespace awareness and understanding, (2) adaptive force planning and application, (3) fast, precise, and resilient kill chains, (4) resilient sustainment support, and (5) efficient enterprise business operations. It stresses that AI-based human-machine collaboration and rapid information analysis processing are essential for achieving these objectives.

Specifically, the 'Data, Analytics, and Artificial Intelligence Adoption Strategy' outlines its strategic goals for the militarization of AI as (1) building 'high-quality data,' (2) enhancing 'joint warfighting' capabilities, and (3) achieving 'responsible military use of AI.' Joint warfighting here means 'addressing Joint capability gaps at the operational to strategic levels' by enhancing interoperability among organizations, thereby strengthening warfighting capabilities. Taken together, increasing decision-making speed not only provides the foundational capabilities for warfighting effectiveness, operational sustainability, and success but also creates the conditions for achieving joint warfighting capabilities.

In this regard, the implications of using AI militarily are likely to be very comprehensive. Existing research categorizes the military use of AI based on whether it is used strategically or tactically, and whether humans supervise AI or machines are granted autonomy, as shown in the table below (Lushenko 2023).

<Table 1> Types of Military Use of AI

Using AI 'tactically' refers to minimizing the time from 'sensor-to-shooter' on the battlefield by rapidly processing large amounts of information acquired through sensors and responding quickly to targets. Using AI 'strategically,' on the other hand, involves utilizing AI to identify options for waging war, combining forces, and projecting power to achieve military objectives. The level of control is distinguished by whether machines are allowed to make autonomous judgments and execute actions, or whether machines provide analytical results that assist human judgment, with the final decision made by humans. Based on these criteria, four types of AI military utilization models are possible.

'Centaur' refers to a mythical creature from Greco-Roman mythology with the upper body of a human and the lower body of a horse. In this context, it represents a model where AI is used tactically to provide options that ensure the efficiency of military operations on the battlefield, but the final decision is made by humans. 'Minotaur,' a mythical monster with the head of a beast and the body of a human, represents a model where the machine makes all judgments for operational execution on the battlefield, from troop deployment to fighter squadron formation, and humans simply follow these decisions. Autonomous Weapon Systems (AWS) that control drones fall into this category. 'Mosaic Warfare' refers to a model where AI is used at a strategic level to identify the optimal combination of friendly forces to exploit enemy vulnerabilities by predicting enemy movements, but the final decision is made by humans. The current US military AI utilization plans mentioned earlier fall under Mosaic Warfare, according to Paul Lushenko's classification. Finally, 'AI-general' represents a model where a nation entrusts all strategic decisions regarding the use of military power to AI, with no human intervention.

Lushenko conducted a survey among US military commanders regarding their preferences for these four types of AI utilization models. Respondents showed a preference order of Mosaic Warfare (1st), Centaur Warfare (2nd), Minotaur Warfare (3rd), and AI-general (4th) in terms of 'reliability.' However, when asked 'Which form of AI model do you actually want the military to adopt?', Minotaur Warfare (1st), Mosaic Warfare (2nd), Centaur Warfare (3rd), and AI-general (4th) received support. These results indicate that in a situation where the stability of AI technology when applied to the military domain is still uncertain, universal human psychology tends to favor models where humans control AI at both tactical and strategic levels.

Interestingly, regarding the specific roles expected of AI on the military stage, there is a somewhat ambivalent tendency: for 'tactical' issues involving warfare on individual battlefields, there is a preference for delegating judgment to AI to increase efficiency through the use of autonomous weapon systems, while for 'strategic' issues of national security, such as planning warfare and integrating national security capabilities, there is a preference for humans to retain control, with AI playing a supporting role in aiding human judgment. This suggests that while most military commanders do not have complete confidence in the stability of AI technology, they do not have significant reservations about using AI at the tactical level. Of course, this survey was conducted solely among US military commanders, and the criteria for selecting respondents are unknown, making it difficult to generalize hastily. Nevertheless, if military commanders actually experience increased operational speed and enhanced efficiency through AI-driven decision-making processes, the role of AI in the military domain is likely to expand further.

The following section examines how the military use of AI will affect the utility and future nuclear strategies of nuclear weapons, considered the most revolutionary weapon system developed by humankind. It will explore how the introduction of AI technology impacts the 'Offense-Defense Balance' (Jervis 1978) and whether it is more likely to favor offense or defense in the realm of nuclear strategy. Ultimately, it will also consider whether the military use of AI complements or replaces existing weapons like nuclear weapons.

III. AI and Nuclear Weapons: A Shift from Countervalue to Counterforce Targeting

1. The Nuclear Revolution and the Importance of Second-Strike Capability

Nuclear weapons are an 'overkill' weapon system whose use is difficult to justify, as they have not been used again in human history since their deployment in Hiroshima and Nagasaki during World War II. Nuclear weapons, upon detonation, generate intense heat comparable to solar radiation, extreme changes in atmospheric pressure causing hurricane-like winds, and radioactive fallout, making survival impossible within tens of kilometers of the detonation point, depending on the yield of the warhead (Wolfson and Dalnoki-Veress 2022). Due to their immense destructive power that can indiscriminately target combatants and non-combatants at high speed (Fetter 1991; Pape 1996), there can be no victor in a war involving nuclear weapons, and avoiding nuclear war by any means necessary becomes the top priority of national security strategy. Paradoxically, it is precisely for this reason that all nations in the nuclear age have pursued cautious security policies, avoiding war and guarding against accidental escalation. This has brought about a revolutionary change for humanity, as evidenced by the strategic stability during the Cold War, where direct military conflict between the US and the Soviet Union was avoided for nearly 40 years (Carnesale et al. 1983; Jervis 1989; Waltz 2003).

The most crucial concept in constructing nuclear security strategy to date has been 'second-strike capability' (Wohlstetter 1959). This refers to the ability to retaliate with nuclear weapons after being subjected to a nuclear attack, thereby retaliating against the aggressor. When both sides possess second-strike capability, leading to mutual annihilation upon nuclear use, the state is called 'Mutual Assured Destruction (MAD).' In a MAD scenario, neither nuclear-armed state can threaten the other with nuclear weapons first because both can inflict 'unacceptable damage' on each other (Jervis 1989). This situation, where each side holds the lives of the other's citizens hostage, paradoxically ensures the security of both nations. Second-strike capability is also referred to as 'countervalue attack' capability, reflecting the principle of 'an eye for an eye,' or threatening the lives of the adversary's citizens in response to a threat against one's own citizens (Kahn 1960).

Conversely, the ability to neutralize the adversary's second-strike capability is called first-strike capability, which focuses on destroying the adversary's nuclear forces and is also known as 'counterforce attack.' The history of nuclear strategy development during the US-Soviet Cold War shows that as nuclear-armed states pursued the enhancement of their second-strike capabilities, strategic stability increased, while efforts to secure first-strike capability led to increased strategic instability (Kaplan 1983; Jervis 1989; Freedman 2003). A prime example is Missile Defense (MD). Although MD appears to be a defensive measure to protect one's territory and citizens from the adversary's nuclear attack, if successful in building a 100% reliable MD system, it would neutralize the adversary's second-strike capability, leading to the collapse of MAD. This is why the Anti-Ballistic Missile (ABM) Treaty was discussed alongside the Strategic Arms Limitation Talks (SALT), the first nuclear disarmament attempt between the US and the Soviet Union, and became the first US-Soviet disarmament agreement in 1972 along with SALT I.

2. Military Use of AI and Nuclear Weapons: Strengthening First-Strike Capability and Deepening Nuclear Asymmetry

Therefore, to assess the changes brought about by the introduction of AI technology in terms of nuclear strategy, it is necessary to examine whether this technology provides greater advantages in first-strike capability (counterforce capability) or second-strike capability (countervalue capability). In nuclear strategy, first-strike capability contributes to enhancing 'offensive advantage,' while second-strike capability is a factor that guarantees 'defensive advantage.'

Before proceeding with a detailed discussion, it is important to maintain the principle that a single technological variable cannot explain changes in the offense-defense balance. As previously mentioned, the most significant advantage of military AI utilization lies in the increase in operational 'tempo.' AI can achieve overwhelming efficiency compared to human cognitive abilities in areas such as target identification, analysis and understanding of the operational environment, and calculating the optimal combination of weapon systems for target engagement, enabling accurate and rapid decision-making. The effects of this capability are likely to extend across the entire spectrum of military force employment. However, this capability itself can contribute to either offense or defense. Therefore, it is difficult to definitively conclude that the era of offensive or defensive advantage has begun solely based on the AI variable. Ultimately, what matters is not the technology itself but how it is operated (Biddle 2023).

With this in mind, let us examine how the 'AI-Nuclear Nexus' will bring about changes in the future battlefield from an offense-defense balance perspective. First, AI's contributions to offensive or counterforce capabilities include Intelligence, Surveillance, and Reconnaissance (ISR), Nuclear Command and Control (NC2), and conventional counterforce operations (Johnson 2023, 18-23; 78-84; 87-90).

First, enhanced ISR capabilities due to AI pose a significant threat to nuclear states attempting to ensure the survivability of their nuclear weapons and second-strike capabilities through methods such as dispersal, mobility, concealment, and protection. Countries like China, North Korea, and even Russia have significant limitations in their submarine capabilities within their nuclear triads. Therefore, these countries attempt to enhance the survivability of their nuclear assets by concealing or protecting them in hardened missile silos or Underground Facilities (UGF), or by using Transporter Erector Launchers (TELs) to evade enemy detection (Lieber and Press 2017). However, these tactics can be easily neutralized if the US integrates AI technology with existing ISR assets such as drones and satellites, thereby maximizing reconnaissance capabilities. For instance, even nuclear submarines, previously considered the ultimate means of securing second-strike capability due to their difficulty in detection and tracking, are becoming more vulnerable with the development of the Sea Hunter unmanned drone by the US Defense Advanced Research Projects Agency (DARPA) for submarine identification and tracking. In other words, the enhancement of identification and detection capabilities directly leads to improved precision strike capabilities, making ISR a technology that contributes to strengthening first-strike capability.

Second, Nuclear Command and Control (NC2) is likely to become vulnerable to cyber and electronic warfare attacks enhanced by AI. For example, if a nation with a superior AI technological capability launches an Advanced Persistent Threat (APT) operation against an adversary, continuous computer hacking processes are initiated against the targeted nation's command and control systems. In such cases, vulnerabilities in the adversary's cyber defense system can be exploited, and malfunctions can be induced by implanting undetectable viruses or malware. Strategies such as data poisoning or spoofing can be employed by inserting false images or target information. While these cyberattacks do not involve physical destruction of the adversary's command structure, unlike precision strikes using conventional weapons, they ultimately render NC2 non-functional, making their practical effect indistinguishable from physical attacks. Therefore, the enhancement of cyber and electronic warfare capabilities should also be considered an increase in first-strike capability. In particular, nations that perceive themselves as being at a disadvantage in cyber defense may adopt the Launch on Warning (LOW) doctrine to address these issues, significantly increasing instability.

Third, conventional counterforce operations are a typical form of first-strike capability. The introduction of AI technology can dramatically improve the accuracy of unmanned weapon systems, making it easier to penetrate enemy defenses. This is particularly effective against enemy defense systems, such as China's Anti-Access Area Denial (A2/AD) strategy, which would incur prohibitive costs to penetrate with existing manned systems. Conversely, if AI-based Automatic Target Recognition (ATR) is added to missile defense systems, detection, tracking, and interception capabilities are drastically enhanced. When combined with drone swarming, this significantly boosts deterrence by denial capabilities. As discussed earlier, while these capability enhancements may appear as defensive measures to protect one's strategic assets and territory, they effectively neutralize the adversary's second-strike capability, thus possessing an offensive attribute.

Does the introduction of AI technology necessarily lead to an absolute offensive advantage in nuclear strategy, thereby increasing instability? Not necessarily. AI technology also contributes to strengthening defensive capabilities in nuclear strategy. For example, AI technology significantly enhances real-time information collection and analysis capabilities for long-range reconnaissance and complex terrain, thereby contributing to improved early warning accuracy and enhanced situational awareness and preparedness for the defending side. Surveillance and reconnaissance capabilities that can be used to neutralize enemy defense systems can also neutralize enemy surprise attacks for the defending side.

In the context of cyber warfare against nuclear command and control capabilities, mentioned earlier, it is possible to respond by enhancing cyber defense capabilities through the integration of AI technology to identify vulnerabilities in one's own cyber defense systems and detect attempts at data poisoning and spoofing by the adversary. In the case of drone swarming, the defending side generally possesses much more information about its military assets than the attacking side. This asymmetry in data can lead to AI performance, based on deep learning, that favors defense over offense. Furthermore, some research suggests that AI-enabled drone capabilities, considering the current limitations in drone range and maneuverability, favor the defending side over the attacking side (King 2024). Therefore, AI-based nuclear asset protection capabilities provide an excellent countermeasure against AI-driven conventional counterforce operations.

Moreover, AI opens a path to overcome the fundamental limitation of 'resolve' in nuclear retaliation, which has been identified as the biggest vulnerability in establishing the credibility of existing nuclear deterrence strategies, such as Russia's 'Dead Hand' launch codes. The problem of nuclear deterrence, repeatedly pointed out in previous nuclear research, is that even if a defending nation possesses the physical capability for a second strike, doubts can arise about the stability of MAD due to the inherent irrationality of nuclear retaliation. As President Eisenhower famously said, 'the only thing worse than losing a global war was winning one,' a nation that has already suffered nuclear damage has little to gain by retaliating with nuclear weapons against the adversary.

To address this issue, Thomas Schelling proposed a response method called 'Threats That Leave Something to Chance,' which intentionally reduces nuclear control, using game models. More recent research (McDermott et al. 2017) suggests signaling nuclear retaliation resolve by seeking 'psychological satisfaction' from punishing the adversary for the suffering inflicted. However, in situations like 'Dead Hand,' where the 'enemy attacks with nuclear weapons' and 'communication with the entity controlling nuclear weapons is lost,' programming AI to 'automatically retaliate with nuclear weapons' could completely overcome the credibility issues of nuclear retaliation and second-strike capability.

Therefore, the 'AI-Nuclear Nexus' does not necessarily lead to offensive advantage or defensive advantage in nuclear strategy. It can yield entirely different results depending on which model is applied and how.

3. Military Use of AI and the Future of Nuclear Weapons: Diminished Nuclear Utility

An important additional issue to discuss is how the introduction of AI technology drastically reduces the necessity of using nuclear weapons on the battlefield. AI technology has the strongest effect of enhancing rapid response and precision strike capabilities, rendering the excessive destructive power of nuclear weapons tactically insignificant. For example, AI-based countervalue strike capabilities are more suitable for precision strikes against enemy command facilities and elimination of commanders than for mass destruction of civilians. The mention of 'the end of the Kim Jong Un regime' in the US Nuclear Posture Review (NPR) released in 2022 appears to be related to this.

AI-based counterforce capabilities also exhibit high efficiency in neutralizing the adversary's second-strike capability through precision strikes against vulnerable points, or in responding to and destroying adversary conventional forces attempting to neutralize those capabilities. In such a scenario, nuclear weapons would function merely as background variables shaping the strategic maneuvering, and the situation requiring the actual use of nuclear weapons might not arise.

Simultaneously, as seen in the offense-defense balance calculation of the AI-Nuclear Nexus discussed earlier, if both sides possess similar levels of AI capabilities, it does not grant unilateral advantage to either the attacker or the defender. However, if one side (e.g., the US) has overwhelmingly superior AI capabilities compared to the other (e.g., North Korea), the superior side is likely to possess an overwhelming first-strike capability, maximizing its own defense of second-strike capability while effectively destroying the adversary's nuclear assets. In this case, North Korea's nuclear weapons would likely hold no utility against the US's integrated AI-nuclear military power in a confrontation involving armed conflict and coercive policies between North Korea and the US.

IV. The AI-Nuclear Nexus and the Future Military Order: The Possibility of Catastrophic Outcomes in US-China Strategic Competition

Taken together, the military use of AI technology is more likely to function not as a force multiplier that amplifies 'nuclear weapon capabilities' or enhances the importance of nuclear weapons, but rather as a factor that drastically enhances the efficiency of existing 'conventional forces,' thereby fundamentally eliminating the need for nuclear use. In this context, AI-based military power is more of a substitute that significantly reduces the strategic necessity and utility of nuclear weapons, rather than a complement that strengthens them.

Considering that AI performance is influenced by the quantity and quality of data, and that computational power suitable for machine learning is essential for AI to function properly, the asymmetry in AI capabilities between the US and China is likely to widen in the long term. In terms of data collected only in actual battlefield situations, such as images, videos, and equipment performance (Horowitz 2018, 52-54), the US possesses unparalleled data. Furthermore, the US dominates 90% of the global advanced semiconductor production equipment supply chain (Allen 2023). This suggests a high probability that the existing superpower, the US, will maintain its exclusive position in terms of cutting-edge AI capabilities. Consequently, the US-led hegemonic order will continue, and the multipolarization of the world political order will not occur. Of course, if China finds entirely new development paths in computational power or military data that differ from the path the US has taken (e.g., quantum computing, theft of sales and security information from Chinese drones), a future where the US does not maintain an overwhelming advantage in competition with China cannot be ruled out.

However, this long-term US-led global military order is highly likely to pass through a period of considerable instability in the short to medium term. Regarding the military use of AI, the current technical limitations include 'brittleness,' stemming from AI's inability to perform analogical reasoning or adaptation at a human level (Johnson 2023, 12-14), and 'ethical issues' concerning machines making decisions about the use of lethal force against humans (Bode et al. 2024, 8-9). Beyond these, the US-China AI competition faces a serious problem that is difficult to escape: two dynamics that significantly increase military instability during the process.

First, the issue of 'nuclear entanglement.' This refers to the problem identified when countries including China and Russia, in the context of current US-China nuclear competition, intentionally entangle conventional and nuclear forces in the deployment and operational stages of nuclear weapons. This includes using dual-use delivery systems capable of carrying both nuclear and non-nuclear warheads, and organizationally integrating units operating nuclear weapons with units operating conventional forces, to maximize the effect of a limited number of nuclear warheads. Nuclear entanglement creates a 'use-it-or-lose-it' situation where, even if an adversary intends to attack only conventional forces, the entangled nuclear weapons also become vulnerable. This significantly increases the possibility of accidental nuclear war (Acton et al. 2017; Talmadge 2017). Limited conventional warfare easily escalates into nuclear war.

A similar problem can arise in the process of building an AI-nuclear nexus. If insufficient data is supplied, AI will have difficulty accurately distinguishing whether the adversary that has established a nuclear entanglement strategy has deployed missiles, bombers, and submarines equipped with conventional warheads or nuclear warheads. In particular, due to the black box problem inherent in the parameter-based reasoning that AI machine learning technology is based on (Johnson 2023, 17-18), even if AI is strictly controlled to only assist in information analysis and all tactical judgments are made by humans, the uncertainty arising from nuclear entanglement could be amplified by AI, leading to more severe instability. That is, if AI does not provide additional explanations for its judgments recommending specific military actions, or if it is impossible to cross-verify the truthfulness of such explanations, human commanders will inevitably face considerable psychological pressure when making decisions within a limited timeframe. In such cases, there is a high probability that the choice will ultimately be to follow the AI's judgment.

Second, the possibility of ‘unintended escalation’ significantly increases when the black box problem of AI technology is combined with the uncertainty of the nuclear entanglement battlefield environment. James Johnson provides a vivid scenario of how mutual deterrence actions at the cyber warfare level between the US and China in the Taiwan Strait could escalate into nuclear war if China conducts a show of force in response to a visit by a high-ranking US official to Taiwan, given that both the US and China possess AI-based defense postures (Johnson 2023, 1-3). The key turning point where US-China strategic competition moves beyond the threshold of war is when AI-based surveillance, reconnaissance, and strategic decision-making systems recommend a pre-emptive strike for ‘early escalation dominance.’ Particularly from China's perspective, which has adopted a nuclear entanglement strategy due to possessing fewer nuclear warheads than the US, it would be considerably difficult to ignore a recommendation from its AI-based defense system that it must launch a pre-emptive strike promptly because the US is targeting not only China's conventional forces but also its nuclear weapons to neutralize China's nuclear deterrence capabilities against the US. And from the US perspective, which is well aware of China's situation, if its AI defense system predicts that China's next response will be an attack on US space-based assets or a hypersonic missile attack on key US assets in Guam, the US is highly likely to launch a pre-emptive counter-strike. All of these conditions provide fertile ground for war to easily break out in the Indo-Pacific region.

V. Conclusion

Therefore, while the US is highly likely to gain a long-term advantage in the AI-nuclear nexus competition, there is a possibility of serious armed conflict between the US and China in the short to medium term due to nuclear entanglement and unintended escalation. The more China feels it is impossible to catch up with the US in the long run, the greater the pressure it will feel to choose war at the present moment. If we do not properly prepare for these possibilities starting now, it is difficult to rule out the possibility of mutual annihilation for humanity, not just in the Indo-Pacific region. This is precisely why the establishment of universal norms for the military use of AI is urgently needed. It is necessary to conduct AI strategic dialogues focusing on agendas where the US and China are more likely to reach an agreement compared to other issues, such as the question of whether to involve AI in decisions regarding the use of nuclear weapons, and discussions on ‘nuclear entanglement’ and ‘unintended escalation,’ which predict catastrophic pathways. ■

References

Acton, James, et al. 2017. Entanglement: Chinese and Russian Perspective on Non-nuclear Weapons and Nuclear Risks. Washington, DC: Carnegie Endowment for International Peace.

Alarcon, Nefi. 2020. “OpenAI Presents GPT-3, a 175 Billion Parameters Language Model.” July 7. https://developer.nvidia.com/blog/openai-presents-gpt-3-a-175-billion-parameters-language-model/. (Accessed: November 20, 2023.)

Allen, Gregory C. 2023. “Blocking China’s Access to AI Chips Matters to U.S. National Security.” CSIS Commentary. July 31. https://www.csis.org/analysis/blocking-chinas-access-ai-chips-matters-us-national-security. (Accessed: January 4, 2024.)

Bai, Yunyi. 2023. “Wang-Sullivan talks last more than 12 hours; Taiwan question takes up longest time, Chinese FM official told GT.” Global Times. September 18. https://www.globaltimes.cn/page/202309/1298392.shtml. (Accessed: November 20, 2023.)

Biddle, Stephen. 2023. “Back in the Trenches: Why New Technology Hasn’t Revolutionized Warfare in Ukraine.” Foreign Affairs. August 10. https://www.foreignaffairs.com/ukraine/back-trenches-technology-warfare (Accessed: January 4, 2024.)

Bode, Ingvild, et al. 2024. “Algorithmic Warfare: Taking Stock of a Research Programme.” Global Society 38(1): 1–23.

Bremmer, Ian and Mustafa Suleyman. 2023. “The Al Power Paradox: Can States Learn to Govern Artificial Intelligence - before It’s Too Late?” Foreign Affairs 102: 26.

Carnesale, Albert, et al. 1983. Living with Nuclear Weapons. New York: Bantam Books.

Congressional Research Service. 2020. “Artificial Intelligence and National Security.” November 10. https://sgp.fas.org/crs/natsec/R45178.pdf. (Accessed: January 16, 2024.)

Elimian, Godfrey. 2023. “ChatGPT costs $700,000 to run daily, OpenAI may go bankrupt in 2024- report.” TechNext. August 14. https://technext24.com/2023/08/14/chatgpt-costs-700000-daily-openai/ (Accessed: January 16, 2024.).

Fetter, Steve. 1991. “Ballistic Missiles and Weapons of Mass Destruction: What Is the Threat? What Should be Done?” International Security 16(1): 5-41.

Freedman, Lawrence. 2003. The Evolution of Nuclear Strategy. London: Palgrave Macmillan London.

Gilpin, Robert. 1981. War and Change in World Politics. Cambridge: Cambridge University Press.

Government of the Netherlands. 2023. “REAIM Call to Action.” Summit on Responsible Artificial Intelligence in the Military Domain. February 16.

Haenlein, Michael and Andreas Kaplan. 2019. “A Brief History of Artificial Intelligence: On the Past, Present, and Future of Artificial Intelligence.” California Management Review 61(4): 5–14.

Horowitz, Michael C. 2018. “Artificial Intelligence, International Competition, and the Balance of Power.” Texas National Security Review 1(3). May.

Jervis, Robert. 1978. “Cooperation Under the Security Dilemma.” World Politics 30(2): 167-214.

______. 1989. The Meaning of the Nuclear Revolution. New York: Cornell University Press.

Johnson, James. 2019. “Artificial intelligence & future warfare: implications for international security.” Defense & Security Analysis 35(2): 147-169. https://doi.org/10.1080/14751798.2019.1600800.

Johnson, James. 2023. AI and the Bomb: Nuclear Strategy and Risk in the Digital Age. Oxford University Press.

Kaplan, Fred. 1983. The Wizards of Armageddon. New York: Simon and Schuster.

Kahn, Herman. 1960. On Thermonuclear War. Princeton University Press.

Kissinger, Henry A. and Graham Allison. 2023. “The Path to AI Arms Control.” Foreign Affairs. October 13. https://www.foreignaffairs.com/united-states/henry-kissinger-path-artificial-intelligence-arms-control

Lindsay, Jon R. 2023/2024. “War Is from Mars, AI Is from Venus: Rediscovering the Institutional Context of Military Automation.” Texas National Security Review 7(1): 29-47.

Lushenko, Paul. 2023. “AI and the future of warfare: The troubling evidence from the US military.” Bulletin of the Atomic Scientists. November 29.

Metz, Steven. 2000. Armed Conflict in the 21st Century: The Information Revolution and Post-Modern Warfare. Carlisle Barracks: Strategic Studies Institute.

Modelski, G., Thompson, W.R. 1988. “The Long Cycle of World Leadership.” In Seapower in Global Politics, 1494–1993. London: Palgrave Macmillan. https://doi.org/10.1007/978-1-349-09154-6_5.

Pape, Robert A. 1996. Bombing to Win: Air Power and Coercion in War. Ithaca: Cornell University Press.

Schneider, Jacquelyn and Julia Macdonald. 2024. “Looking back to look forward: Autonomous systerns, military revolutions, and the importance of cost.” Journal of Strategic Studies 47(2): 162–184. https://doi.org/10.1080/01402390.2022.2164570

Snyder, Jack. 1984. “Civil-Military Relations and the Cult of the Offensive, 1914 and 1984.” International Security 9(1): 108-146.

Talmadge, Caitlin. 2017. “Would China Go Nuclear? Assessing the Risk of Chinese Nuclear Escalation in a Conventional War with the United States.” International Security 41(4): 50-92.

U.S. Department of Defense. 2018. “Summary of the 2018 Department of Defense Artificial Intelligence Strategy.” https://media.defense.gov/2019/Feb/12/2002088963/-1/-1/1/SUMMARY-OF-DOD-AISTRATEGY.PDF. (Accessed: January 16, 2024).

______. 2023. “Data, Analytics, and Artificial Intelligence Adoption Strategy: Accelerating Decision Advantage.” July 27. https://media.defense.gov/2023/Nov/02/2003333300/-1/-1/1/DOD_DATA_ANALYTICS_AI_ADOPTION_STRATEGY.PDF. (Accessed: January 16, 2024)

Wagner, R. Harrison. 1991. “Nuclear Deterrence, Counterforce Strategies, and the Incentive to Strike First.” American Political Science Review 85(3): 727-750.

Waltz, Kenneth N. 2003. “Waltz Responds to Sagan.” In Scott D. Sagan and Kenneth N. Waltz, The Spread of Nuclear Weapons: A Debate Renewed. New York: W.W. Norton & Company.

Wolfson, Richard and Ferenc Dalnoki-Veress. 2022. “The Devastating Effects of Nuclear Weapons.” The MIT Press Reader. March 2.

Kim Yang-gyu. 2023. “'Space as a Battlefield' and Future Deterrence Strategy: Changes in US-China Security Strategy and Korea.” In Challenges of Future Warfare and the Aerospace Industry, edited by Shin Beom-sik. Seoul: Sahoe Pyeongron Academy.

Kim, Yanggyu. 2024. “Korea’s Security Policy in the Era of Offensive Dominance Driven by the Revolution in Precision and Transparency: The Global Military Order of 2024 and Korea.” New Year Special Commentary Series 3. November 20.https://www.eai.or.kr/new/ko/pub/view.asp?intSeq=22234 (Accessed: June 2, 2024).

■ Kim, Yanggyu, Senior Research Fellow at EAI and Lecturer at the Department of Political Science and International Relations, Seoul National University.

■ Managed and Edited by:Park, Jisoo, Research Fellow at EAI

    Inquiries and Editing: 02 2277 1683 (ext. 208) | jspark@eai.or.kr