[Global NK Special Report] ④ Long-term Trends in Inter-Korean AI Cooperation and South Korea's Strategy: Coexistence and Anticipatory Engagement
Editor's Note
Choi In-ho, Senior Research Fellow at the Institute of International Affairs, Seoul National University, analyzes the contradictions of control and capability in North Korea's artificial intelligence (AI) strategy in the current situation where inter-Korean cooperation is severed. The author predicts that as North Korea's AI technology gap accumulates, pressures and turning points will emerge that will compel the regime to seek cooperation internally in the future. In preparation for this moment of decision, Dr. Choi proposes a strategy of anticipatory engagement, which involves maintaining a relationship of coexistence based on trust rather than hostility, while preemptively accumulating multilateral cooperation channels and capabilities as a response measure for South Korea.
■ Go directly to the original Global NK Zoom&Connect
1. Strategy During Periods of Inability to Cooperate: Anticipatory Engagement and Coexistence
The 9th Congress of the Workers' Party of Korea concluded on February 25, 2026. Less than a month later, on March 23, the 1st Session of the 15th Supreme People's Assembly amended the constitution to remove the peace unification clause, established a new territorial clause defining "territory adjacent to the Republic of Korea in the south," and codified the Chairman of the State Affairs Commission's command authority over nuclear forces (Lee Sang-sook 2026). Two years and three months after the declaration of the hostile two-state theory at the end of 2023, the permanence of division has been supported not only at the level of declarations but also at the constitutional and institutional levels. The language used by North Korean state media to refer to South Korea as "a thoroughly hostile country that has made confrontation policy a national policy" had already become commonplace by the summer of 2025 (Korean Central News Agency 2025.8.27). The absence of inter-Korean dialogue and North Korea-U.S. dialogue has exceeded five years, and the direct path to inter-Korean cooperation is effectively closed.
In such a context, discussing inter-Korean AI cooperation may seem anachronistic. However, recent discussions on new North Korea strategies call for a completely opposite sense of timing. "North Korea policy is not only necessary when North Korea is willing to cooperate, but must be prepared more intricately during periods when North Korea refuses cooperation" (Jeon Jae-sung 2026.5). This paradox, that the period of blocked cooperation is precisely the time to prepare for cooperation, serves as the starting point of this manuscript. This manuscript presents the path that South Korea's North Korea AI strategy should take, condensed into two concepts. The name of the strategy is anticipatory engagement, and the form of inter-Korean relations that this strategy should forge during the preparation period is coexistence.
The first concept is anticipatory engagement. While the path to direct cooperation is blocked by the hostile two-state theory and other structural factors, it refers to a strategy of preparing for future cooperation and maintaining a posture of engagement. This is a transfer of the idea of anticipatory governance, which builds management capabilities while emerging technologies are still manageable (Guston 2014), embodying proactivity in preparation and accumulation rather than waiting. The lessons left by past North Korea policies also point in this direction. The Sunshine Policy became a target of attack because its metaphor of shedding outer garments contained a structure of benefaction, and the United States' strategic patience was redefined as strategic neglect due to the lack of visibility on what to do during the waiting period, while the bold initiative led to an immediate dismissal of a one-way transaction structure as "do not dream a vain dream" (Kim Yo-jong's statement 2022.8.19). The name anticipatory engagement encapsulates the logic of time without relying on separate rhetoric, indicating the direction of cooperation without imposing a role of benefaction.
Anticipatory engagement is a strategy that simultaneously pursues three branches of preparation. These include designing incentives to encourage North Korea to decide on partial cooperation, creating a place for South Korea within multilateral channels connected to North Korea even during periods of blocked cooperation, and developing South Korea's own international AI cooperation capabilities to be able to move immediately at the moment of that decision. Each of the three branches targets North Korea, the international environment, and South Korea itself, and they proceed in parallel rather than in stages. The timing of their effects may differ. The specific content of the three branches will be addressed in Chapter 4.
The second concept concerns the form of relations that inter-Korean relations should take during this preparation period. This manuscript refers to it as coexistence. This term appropriates Peter Sloterdijk's concept of co-isolation, which describes a paradoxical state where the cells of a foam structure are separated while sharing a membrane, in the context of North Korea strategy (Sloterdijk 2016). It may sound paradoxical, but in a situation where the hostile two-state theory is firmly in place, maintaining a certain degree of separation while each side develops the capability to stand upright becomes a condition for long-term coexistence. The separation here does not necessarily need to be the hostile disconnection that North Korea advocates. Rather, it is a form of relationship that instills confidence that neither side will pursue hostile policies against the other through intentional separation, thus fostering trust. The goal is not to justify separation as a natural hostility, but to treat the membrane as a joint point to build stable coexistence.
In the AI era, this coexistence requires two simultaneous demands. On one hand, South Korea must secure thorough defenses and deterrent capabilities against North Korea's cyber attacks and infiltrations. On the other hand, South Korea must clearly demonstrate, through the connection of AI and cyberspace, that it does not intend to attack North Korea's political ideology or regime, while also refraining from hastily presenting efforts for unification or attempts at dialogue. Leaving unification as a future task and prioritizing stable coexistence is more realistic during this period. However, this restraint does not mean abandoning all criticism. There is no need to withdraw critical positions on points where North Korea's threats violate regional and global norms. If defense and deterrence are one aspect of separation, then the restraint of not targeting the regime proves that this separation is not hostility. Only when both aspects are present does separation become coexistence that builds trust rather than hostile disconnection.
The structure of the following composition is as follows. Chapter 2 analyzes the structures that hinder inter-Korean AI cooperation in the fields of politics, diplomacy, technology, and economics, as well as the internal contradictions of North Korea's AI strategy. Chapter 3 places these observations on a long-term trend line, addressing how the accumulation of contradictions leads to crises within North Korea. Chapter 4 specifies the three branches of preparation for anticipatory engagement. Where is North Korea's AI strategy headed, and what should South Korea prepare at that juncture? This question serves as the starting point.
2. Short-term Structures Blocking Cooperation and Internal Contradictions of North Korea's AI Strategy
2.1 The Triple Structure of Inability to Cooperate
First, there is political blockade. The 9th Party Congress and the subsequent constitutional amendments need to be read on two levels. The outward flow is an attempt to present the appearance of a socialist normal state, but in reality, the concentration of power in one person and the institutional disconnection of inter-Korean relations have progressed simultaneously. The status of being a nuclear power has become a national reality that is virtually irreversible, and the deletion of unification-related clauses and the establishment of territorial clauses have narrowed the very channels for internal discourse that justify cooperation with South Korea. The absence of cooperation can be evaluated as a situation where the results of policy choices have transcended into institutional outcomes.
Second, there is the deepening of technological cooperation with China and Russia. The comprehensive strategic partnership treaty between North Korea and Russia, signed in June 2024 and effective in December, explicitly names artificial intelligence and information technology as areas of cooperation in Article 10 (North Korea-Russia Treaty 2024). The head of the AI research institute at Kim Il Sung University stated in an interview with a Japanese media outlet in July 2025 that they are sending exchange students, interns, and researchers to countries including Russia (NK News 2025.7.22). Cooperation on AI-based surveillance equipment is ongoing along the North Korea-China border (NK Insider 2025.5). The landscape of international academic cooperation points in the same direction. From 2017 to 2023, the number of international co-authored papers involving North Korean researchers was 161, ranking 145th among 160 countries worldwide, with about 70 papers, or approximately 43 percent, co-authored with China, and the cooperating institutions concentrated in universities in the three northeastern provinces. During the same period, there was only one co-authored paper with South Korea and one with the United States (Kim 2024). Given that North Korea's technological cooperation lines are narrowly aligned with China and Russia, it seems that North Korea currently has little reason to need the South.
Third, there is the framework of sanctions regimes. UN Security Council Resolution 2321 (2016) fundamentally suspended scientific and technological cooperation with North Korea, while paragraph 7 of Resolution 2397 (2017) completely banned exports to North Korea of international harmonized commodity classification (HS) codes 84 and 85, which include computers and electronic devices, and paragraph 17 of Resolution 2270 (2016) explicitly prohibited education and training related to advanced computer simulations and computer science. The head of AI research at Kim Il Sung University has acknowledged that international cooperation in information technology and artificial intelligence is significantly constrained by sanctions (NK News 2025.7.22), indicating that this framework is clearly recognized even within North Korea. The priority of U.S. policy toward North Korea has been pushed back, and after the UN Security Council expert panel that has monitored the implementation of sanctions could not extend its term due to a Russian veto in March 2024, the monitoring and new designation functions of the sanctions regime have entered a state of paralysis, making it unlikely that space for cooperation mediated by adjustments or relaxations of sanctions will open in the short term.
2.2 The Contradiction of Technological Control and AI Capabilities
Under this triple structure, North Korea has placed AI at the top of its regime agenda. Reports from the Korean Central News Agency and Rodong Sinmun regarding the 9th Party Congress state that "efforts must be made to research and develop core technologies necessary for pioneering and operating new industrial sectors such as new energy industries, space industries, and artificial intelligence industries, which are emerging as realistic issues, especially not in the distant future" (Rodong Sinmun 2026.2.26). In July 2025, Kim Il Sung University restructured its Department of Information Science into a Department of Artificial Intelligence, establishing seven research institutes under it (Kyunghyang Shinmun 2026.4.1), and in the military sector, "various AI unmanned attack complexes" have been prioritized in the national defense development plan (Rodong Sinmun 2026.2.26). The recognition that AI is not a distant future but a current task indicates that the North Korean leadership is following global trends at the cognitive level.
The problem lies in the gap between cognition and implementation. Knowledge production in the AI era operates under conditions of open research networks, cross-border mobility of talent, large-scale data, semiconductor and cloud infrastructure, and private ecosystems. North Korea's closed system deviates from each of these conditions. South Korea's national research institute diagnosed that the reality of North Korean AI research is "low computational intensity research is mainstream," and that there are no signs of substantial utilization of the latest high-performance technologies such as diffusion-based generative models based on large language models (LLM), attributing part of the cause to "structural factors such as hardware access restrictions, limitations on overseas collaboration, and technology information blockades" (Kim Min-jeong 2025). The visible achievement known as Ryongma 1.0 is a multilingual translation program using deep learning, and forecasts suggest that the North Korean version of ChatGPT that will emerge in the future will also remain "operating within an internal intranet or limited to specific fields" (KISTI 2025).
At the root of the gap is the data issue. Research on Korean language generative AI published in the student newspaper of Kim Il Sung University conducted experiments using a dataset of 50,000 Korean language texts while citing OpenAI's GPT-4 technical report in the references (NK Economy 2026.5.4). Compared to the world's large language models that are trained on trillions of tokens, or text units used for learning, this dataset is less than one ten-thousandth, reflecting a reliance on publicly available literature from adversarial countries to follow generative AI. There are even more symbolic cases. A paper on fine dust forecasting published by researchers at Kim Chaek University of Technology used air quality data from Beijing rather than Pyongyang (Pak et al. 2024). This is a condensed snapshot of a self-censoring data environment where a regime that cannot disclose its own data is training its own models on external data.
The legal framework of control is tightening this data environment further. The Anti-Reactionary Thought and Culture Exclusion Act (2020), the Youth Culture Guarantee Act (2021), and the recently enacted Pyongyang Language Protection Act (2023) stipulate penalties ranging from six years of labor reform to death for the act of exchanging notifications and emails in "puppet language" in Article 58, and a technical enforcement device that replaces South Korean vocabulary in real-time at the level of smartphone operating systems has also been confirmed (SPN 2023; Fortune 2025.6). The legal system punishing the very use of language data and the technical demand of generative AI, which requires vast amounts of language data, are difficult to reconcile. The regime's operation of using AI as a tool of control undermines the soil necessary for making AI a foundation of capability.
Interestingly, North Korea's official discourse handles this contradiction. Rodong Sinmun praises that "ensuring the rapid and accurate movement of vast amounts of knowledge resources without limitation" is a characteristic of the digital economy while discussing the infrastructure of the digital economy, but remains silent on the fact that its closed network is precisely what blocks that unrestricted movement (Rodong Sinmun 2026.5.13). Reports where Kim Jong-un publicly criticized the digital incompetence of officials as "a heartbreaking reality without any representation of aspirations for modernization" (Rodong Sinmun 2026.5.7) and optimistic statements that "everything humans desire will be realized through artificial intelligence" (Rodong Sinmun 2026.5.11), coexist with the reality that the AI department of important universities is still in the "preparatory phase" (Rodong Sinmun 2026.5.14b). The contradictions are managed not through denial but through silence and suturing, which can be read as a signal that contradictions are accumulating without resolution. While a closed dictatorship may achieve certain results in the limited use of AI for military purposes or cyber capabilities, it is diagnosed that reaching a level of AI transformation that innovates the entire national economy is very difficult (Jeon Jae-sung 2026.5).
2.3 The Economic Foundation and Limitations of Digital Transformation
Observations in the economic sector are on the same trend line. On one hand, the digital foundation of the North Korean economy has clearly expanded. The number of mobile phone subscriptions has reached about 6.35 million, and it has been reported that 4G communication services began in the fall of 2023 (Williams 2026), with the progress of marketization coinciding with the integration of electronic payments and information device usage into daily life. The very fact that the discourse on the digital economy occupies the pages of the party's official newspaper shows the will to use digital transformation as a lever for economic reconstruction.
On the other hand, the upper limit of that transformation is low. First, the physical foundation is thin. According to estimates by the National Statistical Office, North Korea's power generation capacity was 25.5 billion kilowatt-hours as of 2021, which is only 4.4 percent of South Korea's (576.8 billion kilowatt-hours) (National Statistical Office 2022), a scale that is barely sufficient for stable operation of computing infrastructure, let alone large-scale calculations. Next, the costs of control undermine economic efficiency. As the distribution of information devices increases, the number of targets that need to be censored also rises, and North Korea is responding with the technologicalization of control, as evidenced by the confirmation of constant surveillance devices at the level of smartphone applications (Daily NK 2025.7). AI services for remote education and long-distance medical care designed within the Bright Network must also operate in isolation from external knowledge due to the conditions of the closed network (KISTI 2025).
2.4 Dependence on China and Russia as a Suturing Measure and Its Limitations
North Korea's primary measure to stitch together this contradiction is dependence on China and Russia. However, examining what this dependence actually brings reveals clear limitations.
Starting with Russia, the technology transfers confirmed as a price for the dispatch of troops to Ukraine are limited to air defense, electronic warfare, unmanned aerial vehicles, and reconnaissance satellite-related technology support, with AI not included in the list (National Intelligence Service 2025.4). An investigation by South Korea's national policy agency also assessed North Korea-Russia AI joint research as "unknown" (KISTI 2025). Considering that Russia itself is a latecomer in AI and relies on China for a significant portion of its advanced chips, the diagnosis that close ties with Russia will push North Korea deeper into a militarized survival state rather than bring about a transition to a developed state (Jeon Jae-sung 2026.5) is compelling.
In the case of China, while there is support for control technologies utilizing AI, fundamental technological cooperation that can enhance AI capabilities appears to be limited. The AI alert system deployed along the entire North Korea-China border in the spring of 2025 is a system for real-time sharing of unusual activity information at the border between the two countries (NK Insider 2025.5), and this cooperation is closer to technical support that maintains North Korea's closure rather than enhancing its AI capabilities. On the other hand, the pathways for computation and models that form the foundation of AI capabilities are blocked or narrow. AI semiconductors (GPUs), which are the basis for computation, are subject to export controls, leaving no official channels, and smuggling through Chinese intermediaries is only mentioned as a possibility (IISS 2026.3). The fact that supply lines are limited to informal procurement indicates the limitations of this route. In terms of models, what North Korea can obtain are only open-source models, which are publicly available designs that anyone can download, but China's open-source ecosystem is already a censored ecosystem. North Korea, which filters this again through internal censorship, ultimately finds itself in a structure of double filtration, receiving a subset of knowledge that has been re-censored. In fact, China's open-source model DeepSeek has been analyzed to define North Korea as an aggressor in response to queries related to North Korea and criticize its nuclear program (NK News 2025.7.7), making it difficult for North Korea to bring in even the knowledge from the easiest supply line, as it conflicts with its own narrative. In summary, the China route only adds to the capacity for control without providing the capacity for development, and the suturing measure risks resulting in a deepening of knowledge dependency rather than resolving contradictions. Ultimately, what North Korea's cooperation with China and Russia actually provides is control technology, an exit for personnel dispatch, and limited free-riding on open-source.
In this situation, the almost only area where North Korea's AI capabilities are real is the offensive cyber sector. Reports indicate that an organization dedicated to AI-based information theft has been newly established under the Reconnaissance General Bureau (Daily NK 2025.3), and the detection of cases of disguised employment using commercial AI services (Anthropic 2025.8), along with the scale of theft mentioned earlier, suggest that AI is being institutionalized as a tool of control and theft rather than as a foundation of capability. This does not resolve the contradiction between control and capability but is rather an expression of that contradiction. This abnormal situation is likely to further isolate North Korea within diverse stacks of AI networks.
This suturing measure also conflicts with North Korea's own narrative regarding its AI development. In the discourse of self-reliance, Rodong Sinmun states that "sectors and units that seek the reserves of production growth from others' help or national support will always occupy the back seat" while immediately rejecting the frame of external aid, yet in the same article, it writes that "one must thoroughly rely on science and technology to always set high goals and continuously compare with the world," encouraging movements of surpassing, learning from, and exchanging experiences among sectors, units, and regions (Rodong Sinmun 2026.5.14a). The standard of competition is set against the world, while the circuit of learning is closed off to internal exemplary units. In other words, self-reliance is a discourse that commands isolation rather than one that commands competition to win against the world. It does not prohibit the introduction of external technology, as the same article boasts that "the development and introduction of advanced technology is being carried out as the work of the masses themselves." What is limited is not the source of the materials but the attribution of achievements. It is allowed to bring in materials, but it must be narrated as an achievement of internal subjects rather than as external assistance. The problem lies in the distance between what this discourse promises and what the suturing measure actually brings. The discourse promises to reach world standards without relying on others, but as shown earlier, the North Korea-China-Russia route leads to an expansion of the gap with world standards and a deepening of dependence on specific countries. A similar structural dilemma has already been pointed out in diplomatic discourse. North Korea's discourse on multipolarity carries a structural dilemma of being a shield for maintaining sovereignty while justifying dependence on great powers (Jeon Jae-sung 2026.2). The tension observed in the self-reliance discourse can be said to manifest in the realm of AI development. At a time when sovereignty is emphasized most loudly, dependence on China and Russia deepens, and the gap with the world widens.
3. Mid-term, Accumulation of Contradictions and 'New Failing State in the AI Era'
3.1 The Trend Line of Self-Reinforcing Gaps
The contradictions and limitations of suturing discussed in Chapter 2 are not static conditions but rather a trend that deepens over time. From the perspective of the outside world, in 2025, U.S. private investment in AI reached $285.9 billion, showing a gap of about 23 times compared to China's $12.4 billion, although the same report notes that Chinese government-led funds were excluded from the tally and that the performance gap of the top models between the U.S. and China narrowed to 2.7 percent as of March 2026 according to the chatbot arena (Stanford HAI 2026). As the U.S. and China pour such resources into competition, the entry barriers become asymmetrically higher for latecomers. Moreover, the latest AI models have begun to exhibit dynamics of self-improvement, where models help improve other models, and there are warnings that the distribution of AI capabilities among countries could become a new axis of great divergence (UNDP 2025). Once a gap widens, the cost of catching up increases, and as costs rise, the temptation to evade the closed system becomes stronger, creating a self-reinforcing loop.
3.2 The Pathways of Crisis Manifestation and Short-term Backflow
The self-reinforcing AI gap implies that North Korea may face the crisis of a new failing state in the AI era. This state failure does not mean a rapid collapse of the regime, but rather indicates a state where the accumulated AI gap selectively erodes the foundational functions of social governance, such as agriculture, healthcare, and education. In other words, it is not the state that collapses, but rather the functions that the state must provide to its residents that may collapse in various sectors.
AI is not merely a technology of specific industries but a general-purpose technology that penetrates labor markets and industries as a whole (Eloundou et al. 2024), serving as a foundational technology that supports the operation of administration and statistics, production and logistics, disaster response, and social services. The OECD has compiled 200 AI use cases across 11 core government functions, including public service delivery, justice, and financial management (OECD 2025), indicating that the operation of the state itself has become an area of application for this technology. The lesson from the history of general-purpose technologies is that what determines success or failure is not the initial innovation but the capacity to spread the technology across the economy and institutions, and examples of great powers that failed to spread have been repeated since the Industrial Revolution (Ding 2024). While the world raises the efficiency of state operations with this foundational technology, a regime that fails to utilize it will fall behind in all aspects of governance, from the accuracy of administration to the productivity of the economy and the speed of crisis response, and empirical indicators show that the AI usage gap between leading and lagging countries is widening rather than narrowing (Microsoft 2026.1). Moreover, as discussed in Chapter 2, North Korea's failure to utilize AI is not an incidental delay but a structural outcome arising from the contradiction between control and capability, meaning that the gap will not end with a single loss but will accumulate. While the world moves, remaining in place does not mean staying in the same place but rather retreating relatively.
North Korea's current position is already a practical example of this proposition. In the UN e-government survey, North Korea's e-government development index dropped from 131st in 2012 to 184th out of 193 countries in 2024, with not only the ranking but also the absolute score itself declining. In the same survey, South Korea ranks 4th in the world (UN DESA 2024). The government AI readiness index excluded North Korea from the assessment due to a lack of data in the 2024 edition, but when it was first assessed in the 2025 edition, it ranked 167th out of 195 countries, receiving a score of 0 in the policy sector (Oxford Insights 2026), and the IMF's AI readiness index could not include North Korea in the assessment due to a lack of data (IMF 2024). The lag is either measurable or not even measurable. Either way, it records the gap.
This accumulating retreat will first and most visibly manifest in areas with already vulnerable baselines, such as health and education. The DTP (Diphtheria, Tetanus, Pertussis) vaccination rate for infants dropped sharply from 97 percent in 2020 during the border closure period, estimated at virtually 0 percent at one point (WHO·UNICEF joint estimate WUENIC), and while it has entered a recovery phase in 2024 with the import of vaccines and catch-up vaccinations for missed cohorts, the vulnerability of the health infrastructure remains. The proportion of stunted children, despite a long-term improvement trend, is estimated at 16.6 percent in 2024, still one in six (UNICEF·WHO·World Bank JME 2026). The failure to utilize AI translates into specific losses, as exemplified by tuberculosis. North Korea, a long-standing high-burden country for tuberculosis, has its incidence estimates classified as "under review" in the World Health Organization's 2025 report, indicating that it cannot even count how many tuberculosis patients it has (WHO 2025). Globally, as of 2024, about 2.4 million tuberculosis patients, or one in five, remain undiagnosed, with the primary bottleneck in low-resource environments being the lack of specialized personnel to interpret chest X-rays. The computer-aided detection (CAD) AI recommended by the World Health Organization for tuberculosis screening serves as a tool to fill this bottleneck, performing interpretations comparable to specialists with just a laptop and portable X-ray, without an internet connection (WHO 2021). Paradoxically, North Korea ranks among the highest in the world in terms of the number of doctors and hospital beds per capita, so what is lacking is not the number of medical personnel but diagnostic capabilities and electricity, which is precisely where AI can substitute. While other countries use such offline AI to find undiagnosed patients, a regime that cannot introduce these tools will continue to miss one in five, and untreated tuberculosis will spread through transmission and drug resistance, exacerbating losses. The same structure repeats in maternal and child diagnostics and infectious disease surveillance.
In education, AI departments are still in the preparatory stage, and in agriculture, there is a significant gap between the rhetoric of scientific farming and the reality on the ground. The thinner the foundation in these areas, the greater the gap in AI utilization translates into losses directly related to human survival and capability. The military-first policy has been pointed out as a "self-contradiction that rather leads to a weakening of security over time" and "a weakening of overall domestic capabilities," a critique that emerged during the evaluation of the 8th Party Congress, but it carries even more weight now with the added variable of AI.
Of course, the accumulation of crises does not mean imminent collapse. There are diagnoses suggesting that North Korea can endure for a considerable time by relying on a combination of nuclear weapons, military provocations, internal control, and external balance, and we must be wary of both illusions that North Korea will collapse soon or change its strategy solely due to external pressure, as well as the expectation that dialogue and cooperation alone will return to the framework of past inter-Korean relations (Jeon Jae-sung 2026.5). The crisis referred to in this manuscript is not the end of the regime but rather the accumulation of internal pressures that will compel a review of partial cooperation, with the time horizon not being short-term but rather mid-term and beyond. Moreover, in the short term, there may even be backflow. As observed, the deepening of the AI gap may push the regime toward stronger control and military dependence in the short term (Jeon Jae-sung 2026.5), and the strengthening of information control legislation and concentration on the cyber sector can be read as signs of this backflow. This backflow also implies that the period of inability to cooperate may be prolonged, so South Korea's preparations should also be designed with a long breath.
3.3 The Hypothesis of Turning Points and Competing Scenarios
Cases where closed regimes have shifted to selective openness in the face of technological gaps provide significant insights into the conditions for turning points. Looking across socialist and authoritarian states that have pursued technological capabilities under sanctions and isolation, the trigger for transition has been pulled when three conditions are met: recognition of the gap by elites, accumulation of crises, and alleviation of external hostility. The Soviet Union delayed its transition because the leadership failed to recognize the information technology gap in a timely manner in the 1970s. Vietnam, despite receiving a shocking diagnosis through secret research by diplomats that the Soviet bloc was over a decade behind the West, only moved toward the reform and opening policy of Doi Moi in 1986 when the accumulation of crises such as hyperinflation and a two-front war coincided with the normalization of relations with the U.S. (Path 2020). The lesson consistently told by comparative historians is that if the alleviation of external hostility is absent, even if the gap is recognized, the opening will be delayed or fail.
Common grammar can also be found in the methods of transition. In no case has the discourse of self-reliance or ideological superiority been discarded. China has reframed self-reliance as "self-reliance through openness," justifying the introduction of advanced global technologies as a starting point for self-reliance. Iran has increased international joint research during times of heightened sanctions (Kokabisaghi et al. 2019), and the scientific rise proclaimed by Khamenei in the 2014 science and technology policy communiqué as "scientific jihad" has been packaged in the language of religion and nationalism (Bazoobandi 2024). Cuba has narrated its concentration on healthcare and biotechnology under blockade as "medical internationalism." Closed regimes have only been able to reduce political costs when they package openness as an extension rather than a break from existing justifications. Notably, Cuba has advanced to supply its own vaccines to over 30 countries even under sanctions by using the certifications from the World Health Organization and the Pan American Health Organization as pathways for internationalization (Pentón-Arias et al. 2021), demonstrating that mediation by multilateral organizations can simultaneously satisfy the face and practical needs of closed regimes.
North Korea's own precedents point in the same direction. The cooperation between Kim Chaek University of Technology and Syracuse University (initiated in 2001, continuing until around 2012) (Thorson 2012), the operation of Pyongyang University of Science and Technology, and the experiences of inter-Korean IT cooperation in the 2000s show that North Korea has accepted partial cooperation in non-political technological fields when it is clear, can be packaged in its own way, and can control human contact under the mediation of multilateral or third countries. The most vivid recent comparison is vaccines. North Korea rejected the vaccines from the international vaccine procurement organization COVAX, which required the entry of international monitoring personnel in 2021 (NK News 2021.9), but in 2024, it accepted approximately 4.2 million doses of vaccines from a UN agency executed by the Ministry of Health without the presence of international staff (UNICEF 2024.7). During the same period, it did not respond to the South Korean Red Cross's proposal for flood relief. This contrast shows that the variable determining the decision is not cooperation itself but the level of monitoring and the political color of the pathway.
However, it should be noted that the hypothesis of turning points has competing scenarios. The two-state theory may change again if the international order changes, but there is an observation that this direction could be a resumption of an aggressive unification line (Jeon Jae-sung 2025.9), and the case of Myanmar suggests that partial openness was a product of confidence in control rather than a weakness of the regime (Jones 2014). Transitions are not only open in the direction of cooperation. However, this caveat does not lead to a theory of useless preparation. Rather, if North Korea's openness is predicated on confidence in control, the form of cooperation that South Korea should prepare must be designed in a way that North Korea can feel is controllable, that is, adjustable in terms of monitoring levels, reversible, and starting small.
4. South Korea's Strategy, Foundations of Coexistence, and Three Branches of Preparation
Based on the above trend diagnosis, South Korea's strategy consists of two layers. One is to support the form of relations that inter-Korean relations should take during the preparation period, namely coexistence, with specific policies in the AI and cyber domains (4.1). The other is to prepare three branches for future cooperation on that foundation, namely providing incentives, preemptive groundwork, and capacity enhancement (4.2-4.4). Each of the three branches targets different subjects. Providing incentives targets North Korea's decision-making, preemptive groundwork targets the international environment of sanctions regimes and multilateral systems, and capacity enhancement targets South Korea itself. The three branches proceed in parallel rather than in stages, with the timing of their effects differing.
4.1 Coexistence and Trust Through Coexistence
Coexistence does not remain an abstract relational theory but is translated into specific policies in the AI and cyber domains. As stated in the introduction, the isolation during this period requires both aspects simultaneously. One aspect is a solid defense and deterrence to block North Korea's attacks and infiltrations, while the other aspect is restraint, clearly demonstrating that South Korea does not intend to shake North Korea's regime or political ideology through AI and cyberspace connections. If the former shows that the isolation is not flimsy, the latter shows that this isolation is not hostility. Only when both aspects are present does isolation become a foundation of trust.
The need for defense and deterrence is not abstract but a response to ongoing threats. The National Intelligence Service diagnosed in its 2026 "National Cybersecurity White Paper" that North Korea has significantly strengthened its cyber attack capabilities under the hostile two-state theory, exploiting vulnerabilities in information protection software and office programs to conduct supply chain attacks and steal secrets from the defense industry, energy infrastructure, and high-tech companies (National Intelligence Service 2026), and that the North Korean-linked organization Kim Suki created forged images of South Korean military identification cards using generative AI to target defense-related institutions (Genians 2025). The fact that North Korea stole approximately $2 billion worth of cryptocurrency in just 2025, accumulating over $6.7 billion (Chainalysis 2025), and reports of the establishment of an organization dedicated to AI-based information theft under the Reconnaissance General Bureau (Daily NK 2025.3) show that the AI threat is not a one-off incident but a institutionalized permanent challenge. South Korea is building institutional thickness to respond to this. The national cybersecurity strategy and its basic plan announced by the National Security Office in 2024 set proactive defense and resilience of critical infrastructure as core tasks (National Security Office 2024), with the Cyber Operations Command of the Defense Ministry, the National Cybersecurity Center of the public sector, and the Korea Internet & Security Agency sharing roles. The South Korea-U.S. alliance adopted a strategic cybersecurity cooperation framework in 2023 to establish a common foundation for threat information sharing and deterrence against hostile forces (Ministry of Foreign Affairs 2023), and the establishment of the AI Safety Research Institute in 2024 and the government's push for a transition to AI-based cyber defense in 2026 are efforts to thicken the walls of isolation in the AI era.
The other aspect is more difficult and is still closer to the realm of recommendations. When South Korea clearly demonstrates restraint by not using its AI and cyber capabilities to target North Korea's regime itself, isolation becomes trust rather than hostility. There are precedents for this restraint. The current government voluntarily suspended loudspeaker broadcasts toward North Korea in June 2025 and requested the cessation of private leaflet distribution toward North Korea (Ministry of Unification 2025), which continues the spirit of the 2018 Panmunjom Declaration and military agreements that formalized the mutual cessation of hostile acts and psychological warfare means. The remaining task is to transfer the same principle of restraint to the AI and cyber domains. Specifically, it involves not mobilizing AI-generated disinformation or deepfakes targeting North Korean leadership as tools of influence operations, and clearly distinguishing defensive proactivity from offensive operations targeting the regime. The fact that the international cyber norms agreed upon with South Korea prohibit attacks on the critical infrastructure of other countries and interference in internal affairs (UN GGE 2015) supports that this restraint is not an isolated concession but an extension of universal norms. However, since there are currently no binding international norms directly regulating regime-targeting operations using AI, South Korea's restraint in this area takes on the character of a voluntary choice that precedes the norms.
Restraint does not mean abandoning criticism. Political attacks targeting the legitimacy of the regime and criticisms based on universal norms must be distinguished. What South Korea should withdraw is the former, while what it should not withdraw is the latter. The area where this distinction is most clearly established is precisely in AI and cyber. Since 2015, the international community has reached consensus on cyber norms for responsible state behavior through the UN Group of Governmental Experts, with the core being not to engage in cyber activities that damage the critical infrastructure of other countries and not to deliberately allow one's territory to be used for international illegal acts (UN GGE 2015). The official position of the South Korean government is that international law and the UN Charter apply equally to cyberspace, so these norms universally affect North Korea, which did not participate in the agreement.
North Korea has systematically repeated acts that directly contradict these norms. The WannaCry attack in 2017 paralyzed dozens of hospitals in the UK's National Health Service, indiscriminately attacking the critical infrastructure of other countries (U.S. White House 2017), and organizations under the Reconnaissance General Bureau have targeted the defense, aerospace, nuclear, and engineering sectors of South Korea and its allies to steal secrets and funds for the regime's military and nuclear plans (CISA·National Intelligence Service 2024). The theft of cryptocurrency amounting to approximately $2 billion in just 2025 violates both cyber norms and non-proliferation norms (Chainalysis 2025), and the use of AI-generated identities and personas for disguised employment and fraud repurposes generative AI as a tool for deception (MSMT 2025.10). Criticizing these acts through channels such as reports from the multinational sanctions monitoring team or the South Korea-U.S. joint cyber advisory is not an insult to the regime but rather a specification of violations of universal norms. Unlike North Korea, which treats criticism of its leadership and regime as a crime, South Korea's restraint is a choice that shifts the target of criticism from the nature of the regime to violations of norms that cross the walls. The clearer this distinction is, the clearer the signal that isolation is not hostility.
When solid defenses and restrained non-aggression are both in place, the isolation between the two Koreas becomes coexistence that builds trust while standing upright rather than hostile disconnection. Only on this foundation of relationship does the preparation of the three branches for future cooperation gain significance.
4.2 Providing Incentives: From Self-Reliance to Co-creation
The first branch is the design of incentives that will influence North Korea's decision-making. Incentives come from two directions. One is to reveal the costs that will be incurred if cooperation does not occur, and the other is to present what North Korea can gain in language that it can accept when cooperation occurs. The combination of power and attraction becomes an incentive only when both operate together.
First, regarding the cost direction. South Korea's government, research institutions, and civil society need to consistently visualize the asymmetric costs that a closed regime will incur in the AI era. As seen in Chapter 3, these costs have already manifested in measurable forms, such as the decline in e-government index or the uncounted incidence of tuberculosis, and due to the self-reinforcing nature of the gap, they will only grow over time. The long-standing insight that North Korea must realize that nuclear line politics will inevitably become a self-defeating strategy gains new relevance in the AI domain. However, this visualization must be a task of recording the facts of the gap, not an attack targeting the regime. Since the target is the nature of the regime rather than the measurable gap, this task does not contradict the restraint mentioned in 4.1.
Next is the direction of benefits. Here, it is necessary to intervene with concepts that can engage with North Korean discourse and concepts. North Korea's self-reliance rhetoric immediately rejects the frame of aid and support, but as confirmed in 2.4, competition, learning from others, and exchanging experiences are already institutionalized methods of movement among internal units. Therefore, the incentive for cooperation should be presented not as benevolent support but as co-creation among equal competitors, that is, as an expansion of the circuit of experience exchange that is closed for internal use to an international level. It may be necessary to newly conceptualize this direction of discourse as co-creation (共力共生) as a modification and expansion of self-reliance. The message is that sovereignty in the AI era is only possible on the basis of multilateral integration, showing that it is in line with the normal path of comparative history, which justified openness while redefining sovereignty. As seen in Chapter 2, the discourse of sovereignty and the reality of dependence on China and Russia have already diverged, and as comparative historical examples show, this tension has been resolved not by the abandonment of sovereignty discourse but by its redefinition. Multilateral distributed cooperation provides a political space that serves as an alternative to single dependency.
On a more concrete level, it is necessary to actively seek out and develop AI application cases verified in low-resource environments and deliver them to North Korea. Examples such as Ghana's math learning assistant Rori (Henkel et al. 2024), AI utilization education experiments in Edo State, Nigeria (World Bank 2025), and the video interpretation AI recommended by the World Health Organization for tuberculosis screening (WHO 2021), as well as agricultural advisory chatbots distributed to hundreds of thousands of farmers in India and Africa, demonstrate that even in environments constrained by power and connectivity, it is possible to narrow the gaps in education, health, and agriculture, and specify the path of adapting globally validated models to the Korean Peninsula. The principles of design should follow the discoveries derived from the comparative history and vaccine cases in 3.3. Small-scale pilot projects that can be adjusted in terms of monitoring levels, are reversible, and are exchanged simultaneously in stages should serve as the starting point. A fourth principle needs to be added. Considering North Korea's history of repurposing AI for surveillance and theft, the targets of cooperation should be limited to civilian sectors, and the design must incorporate criteria for controlling access to models and computational resources while preemptively assessing the potential for surveillance repurposing. This criterion serves as both a response to skepticism about cooperation and a safeguard that maintains consistency with sanctions resolutions.
At the same time, it is necessary to clearly state the limitations of incentives. As long as the supply lines from China and Russia are maintained, North Korea may not feel external incentives as vital variables (Jeon Jae-sung 2025.5). The effectiveness of incentives will only manifest as the path of development, rather than militarized survival, becomes clearer, and as the limitations of the suturing measures discussed in Chapter 2 accumulate. Providing incentives should be viewed not as a path to short-term results but as a way to thicken moderate options within North Korea at the moment of turning points.
4.3 Exploring Multilateral AI Cooperation Spaces
The second branch concerns the multilateral channels that will be practically usable when North Korea decides to engage in AI cooperation. In particular, it is necessary to clearly identify the multilateral cooperation channels that are already open within the constraints of sanctions.
Sanctions regimes are often perceived as being in a state of complete paralysis, but the humanitarian exemption track of the UN Security Council's North Korea sanctions committee (1718 Committee) is still functioning normally, with eight approvals occurring between the 6th and 16th of February 2026 (1718 Committee Exemption List 2026.3). The practice of exemptions has already entered the digital realm. The Food and Agriculture Organization (FAO) has received exemptions from the sanctions committee for projects introducing digital technology into agriculture (2024.10) and for importing information technology devices for projects (2026.2). This means that digital and information technology-related materials can legally enter North Korea through exemption procedures. In South Korea, the Our People Helping Each Other Movement (2019) and Gyeonggi Province (2024) have previously applied for exemptions. This indicates that this pathway was actually open to South Korea's civil society and local governments, and the original text of the sanctions resolution contains a hierarchy of legal pathways.
Medical exchanges are exceptions under paragraph 11 of Resolution 2321, and other scientific and technological cooperation has pathways for party state judgment and prior notification to the committee (the joint research on Mount Baekdu by the UK is a real working precedent), with exemption procedures for material imports and prior approval procedures for permanent joint cooperation bodies. It is important to note that the format of education and training carries the greatest risk of violating the explicit prohibition in paragraph 17 of Resolution 2270, so future cooperation should be designed in the form of joint research and medical exchanges. It is also worth remembering that humanitarian exemptions are the point where the interests of the permanent members of the Security Council converge within the sanctions regime. Given that China has an incentive to counter South Korea's AI approach toward North Korea, it is realistic to start in areas such as health, disaster, and statistics, where China would find it difficult to find counterarguments.
The reason these institutional pathways do not remain mere possibilities is revealed in North Korea's previous attempts to lift these sanctions regimes. At the Hanoi summit in 2019, North Korea demanded the lifting of five Security Council resolutions (2270, 2321, 2371, 2375, 2397) adopted in 2016-2017 in exchange for the dismantling of the Yongbyon nuclear facility (Ri Yong-ho 2019). North Korea presented this as a lifting of clauses that strangle the people's livelihood, but those five resolutions included clauses on the suspension of scientific and technological cooperation (2321), the prohibition of exports of computers and electronic devices (2397 paragraph 7), and the prohibition of computer science education (2270 paragraph 17). This means that the very core of the sanctions North Korea sought to lift was situated in the scientific and technological domain, and now that AI is placed at the top of the regime's agenda, the incentive to gain legitimate benefits in this area has only increased. Therefore, the possibility of cooperation in the scientific and technological and computer sectors through multilateral channels is not merely a hope but can be based on North Korea's revealed preferences. However, lifting sanctions in the scientific and technological domain must not become a unilateral concession. It must be conditioned on corresponding cooperation from North Korea, particularly the cessation of cyber attacks targeting the critical infrastructure of other countries and cryptocurrency theft. When the easing of sanctions in the scientific and technological domain is linked to the cessation of acts that have violated universal norms, it becomes not a reward for a closed regime but an exchange that induces compliance with norms, which also aligns with the restraint of coexistence discussed in Chapter 1.
In addition to finding possibilities for multilateral cooperation within the sanctions regime, it is also necessary to expand the channels of multilateral organizations that are adjacent to North Korea. The meeting between the UN resident coordinator and the North Korean Central Bureau of Statistics in Bangkok in May 2026, after more than five years, suggests that statistical and data cooperation is the first candidate for reopening (UN DPRK 2026.5), and data cooperation is a precursor to AI cooperation. The precedent of the Korea Energy Development Organization (KEDO), where hundreds of South Korean personnel resided in North Korean territory for years under a multilateral flag, shows the upper limits of the space that can be secured through multilateral channels. "AI-based crisis prediction, military conflict prevention, disaster response, healthcare support, agricultural productivity improvement, environmental monitoring, border area management, and humanitarian support systems could become new areas of inter-Korean cooperation in the future. Even if the North Korean regime does not accept this immediately, South Korea must prepare technical, institutional, and diplomatic readiness for future cooperation" (Jeon Jae-sung 2026.5) summarizes the tasks of this branch.
▲ In May 2026, the UN resident coordinator met with a delegation from the North Korean Central Bureau of Statistics for the first time in over five years.
(FAO·AIT Statistical Capacity Strengthening Program). Source: UN DPRK.
4.4 Capacity Enhancement: Preparation for the Moment of Decision
The third branch concerns South Korea's own preparation. The moment North Korea decides on partial cooperation, a division of labor model must be immediately activated, where technology and funds are supplied by South Korea and deployment within North Korea is handled by multilateral organizations. For this to happen, the capacity to handle the assets to be supplied and the procedures for transporting those assets must be in place in advance. South Korea has already accumulated considerable assets in AI governance. The problem is that these assets are not yet aimed at peace coexistence on the Korean Peninsula in the AI era.
The resource base is being established. South Korea ranks among the top five countries in the world in terms of total AI usage (Anthropic 2026.1), and is also rated fifth in terms of usage intensity per capita (Anthropic 2025.9). The APEC AI Initiative adopted at the Gyeongju APEC summit in 2025 (2026-2030) specified the establishment of an Asia-Pacific AI Center operated solely by South Korean funding (APEC 2025), and the AI Basic Act, which stipulates support for international cooperation in Article 22, came into effect in January 2026, with the government's AI budget for 2026 reaching 10.1398 trillion won according to the government proposal and 9.9334 trillion won according to the National Assembly's confirmation, three times the level of the previous year (Ministry of Strategy and Finance 2025.11). The Seoul Declaration of the AI Seoul Summit (2024.5), the REAIM Blueprint on responsible AI in the military sector, and participation in the international network of the AI Safety Research Institute are assets at the normative level, while KOICA's digital official development assistance (ODA) projects and the Korea-ASEAN Digital Innovation Fund are assets at the development cooperation level. In May 2026, the announcement of the Seoul Global AI Hub involving nine UN agencies coincided with the visit of the UNDP Administrator (The Korea Times 2026.5), further broadening the intersection between multilateral organizations and South Korea.
However, a list of assets does not equate to capacity. The legal pathways seen in 4.3 become viable only when there are organizations and people who have actually handled them. The experiences of Our People Helping Each Other Movement and Gyeonggi Province in applying for exemptions are scattered as memories of individual entities and have not been accumulated as national assets. Three things are needed. First, a procedural manual that organizes the requirements, timeframes, and reasons for rejection of exemption applications and project approvals on a case-by-case basis. Second, dedicated practical capacity to handle these procedures, which involves consolidating the functions scattered across the Ministry of Unification, the Ministry of Foreign Affairs, and development cooperation agencies into a single window for North Korea AI and digital cooperation. Third, the prior acquisition of technology stacks that operate in low-resource environments. By having configurations verified in conditions constrained by power and connectivity, such as offline interpretation AI or lightweight learning support tools, development can begin with application rather than starting from scratch when cooperation opens. These three tasks can be initiated regardless of North Korea's attitude, and in that sense, they represent the most certain preparations.
The remaining task is to connect these assets to the possibilities of cooperation with North Korea. There are two branches to this. One is to agenda-set at the level of the South Korea-U.S. alliance, which involves including a proposal for sharing North Korea AI governance in the flow of restructuring the South Korea-U.S. alliance into a comprehensive alliance that combines contributions in advanced industries such as AI, semiconductors, and energy (Jeon Jae-sung 2026.1). The proposal for South Korea to take on the issue of North Korea AI governance as a pillar of governance that the U.S. cannot handle alone could become an agenda that broadens South Korea's voice in discussions on restructuring the alliance. The other is to preemptively design scenarios for North Korea's application in each of the channels listed above. For example, the low-resource environment program of the Asia-Pacific AI Center, the humanitarian AI track of the Global AI Hub, and the Korea-ASEAN Digital Innovation Fund are all efforts to create a place for North Korean cooperation within the channels already held.
Additionally, the assets accumulated by the three branches, such as the manual for exemption procedures, seats within multilateral channels, and the division of labor agenda of the alliance, are bureaucratic and institutional assets with a lighter political color, making them relatively resilient to disruptions caused by regime changes. Considering the history of past North Korea policies being discarded with changes in regime units, this aspect supports the sustainability of the three-branch strategy.
5. Conclusion
In one sentence, the diagnosis of this manuscript can be summarized as follows. North Korea's AI strategy stands on the contradictions of control and capability, and while this contradiction may lead to stronger control and military dependence in the short term, there is a high possibility that it will accumulate into internal pressures that compel a review of partial cooperation in the mid-term. It cannot be determined when or in what form this turning point will come, and the possibility of a shift in direction away from cooperation remains open. However, as comparative history shows, lowering the costs of turning points is dependent on the pathways and capabilities established beforehand.
Therefore, South Korea's tasks are clear. In the current situation where cooperation is impossible, it must send incentives for co-creation in the frame of co-creation rather than aid, conduct preparatory groundwork to establish practical positions within multilateral channels and manuals for sanction exemptions, and prepare for the moment of decision through the restructuring of the South Korea-U.S. alliance and the accumulation of assets for multilateral AI governance. The parallel pursuit of these three branches is anticipatory engagement. And the form that inter-Korean relations should take during this preparation period is coexistence, a relationship that builds trust through intentional isolation rather than hostility, allowing each side to stand upright. This strategy of paving the way and building capabilities for the moment cooperation becomes possible is also a path for South Korea to transition from being a "consumer of order to a designer of order" (Jeon Jae-sung 2026.6). If the core of the new North Korea strategy is "long-term confidence" (Jeon Jae-sung 2026.5), then the basis for that confidence in the AI domain comes from the technologies, institutions, and networks that South Korea already possesses, as well as the strategies that have been prepared more intricately during periods of refusal of cooperation. Preparation is an unavoidable task even when the day of harvest cannot be promised, and now is precisely the time for that preparation.
Finally, I would like to suggest one task that this report has not been able to address in depth. The AI gap of a nuclear-armed closed state is not limited to internal matters on the Korean Peninsula. If the governance base of an actor with nuclear weapons and missiles is shaken, asymmetric risks will be added not only to Northeast Asian security and non-proliferation but also to cross-border governance in health and the environment, and neglecting the AI gap itself can be translated into costs for global governance. However, the entities responsible for managing these costs are not clear. North Korea has been pushed down the priority list of the United States, and China has chosen to keep the current North Korea close to itself (NPR 2026.6). Under conditions where no single country can bear the complex public goods of the 21st century alone, the AI gap of a nuclear-armed closed state is a typical agenda that requires a division of labor among multiple actors (Jeon Jae-sung 2026.6), and South Korea taking on a position in that division of labor could be a positioning that aligns with the changing governance environment beyond merely pursuing status. However, the work of formalizing the North Korean AI issue as an agenda for global governance remains beyond the scope of this report and is left as a separate research task.
References (Selected)
Primary Sources (North Korea)
Rodong Sinmun (2026.2.26). "Report on the 9th Congress of the Workers' Party of Korea."
Korean Central News Agency Commentary (2025.8.27). (On South Korea's designation as a "thoroughly hostile country") Rodong Sinmun (2026.5.7). "Emphasis on the Importance of Numbers is Essential."
Rodong Sinmun (2026.5.11). "Education that Pioneers the Future." Rodong Sinmun (2026.5.13). "Characteristics of the Digital Economy Infrastructure."
Rodong Sinmun (2026.5.14a). "Let Us Actively Engage in Socialist Competition by Sector, Unit, and Region." Rodong Sinmun (2026.5.14b). "Significant Progress Achieved in Improving Education Structure, Content, and Methods."
Foreign Minister Ri Yong-ho's Press Conference in Hanoi (2019.2.28). (North Korea demands the lifting of civilian-related parts of five Security Council resolutions (2270, 2321, 2371, 2375, 2397) adopted in 2016-2017 in exchange for the dismantling of the Yongbyon nuclear facility; these five resolutions include clauses on the suspension of scientific and technological cooperation, prohibition of HS84/85 exports, and prohibition of computer science education. Reported by CNN, VOA Korean, etc.)
Kim Yo-jong's statement (2022.8.19). (Rejection of the bold initiative, "Do not dream a vain dream" — reported by Kyunghyang Shinmun)
The "Pyongyang Language Protection Act" (adopted as Law No. 19 by the Supreme People's Assembly on January 18, 2023). The full text is published in SPN Seoul Pyongyang News (2023.6.19, "[Full Text of the Pyongyang Language Protection Act]").
Comprehensive Strategic Partnership Treaty between the Democratic People's Republic of Korea and the Russian Federation (signed on June 19, 2024, effective December 4, 2024). Full text published by Korean Central News Agency.
"Panmunjom Declaration for Peace, Prosperity, and Unification of the Korean Peninsula" (April 27, 2018). (Article 2: Cessation and abolition of hostile acts and psychological warfare means)
"Military Agreement for the Implementation of the Historic Panmunjom Declaration" (September 19, 2018; June 4, 2024, all effects suspended on the South Korean side).
Korean text.
National Intelligence Service (2025.4.30). Report from the closed meeting of the National Assembly Intelligence Committee. News1 (2025.4.30) "NIS: 'North Korean troops dispatched to Russia, including 600 dead among over 4,700 casualties' (comprehensive report)" and other reports.
Kim Min-jeong (2025.11.25). "Analysis of the Current Status of North Korean Artificial Intelligence and Policy Considerations." INSS Strategic Report No. 350. National Security Strategy Institute.
Kyunghyang Shinmun (2026.4.1). (Report on the restructuring of the Department of Artificial Intelligence at Kim Il Sung University)
KISTI North Korea ICT Research Group (2025). "North Korea ICT Trends Survey 2025." Korea Institute of Science and Technology Information.
NK Economy (2026.5.4). (Report on the Korean language generative AI research at Kim Il Sung University)
Ministry of Economy and Finance (November 7, 2025). "(Press Release) FY2026 AI budget to be 10.1 trillion KRW, following consultations with the National AI Committee, Office of Science and Technology Innovation, and specialized agencies."
Lee, Sang-sook (June 12, 2026). "Key Contents and Significance of North Korea's Constitutional Amendment in March 2026." IFANS Major International Affairs Analysis 2026-18. Institute of Foreign Affairs and National Security, National Diplomatic Academy. (Analysis of the March 2026 amendment, including the addition of territorial clauses [Article 2], the complete deletion of clauses related to unification, peaceful unification, and the nation, and the codification of the Chairman of the State Affairs Commission's nuclear command authority [Article 89].) National Strategy Portal
Ministry of Unification (June 2025). Suspension of loudspeaker broadcasts to North Korea (June 11, 2025) and suspension of civilian leaflet scattering ("Achievements in Promoting Inter-Korean Peace and Coexistence Policy in the First Year of the Lee Jae-myung Administration").
National Security Office (2024). "National Cyber Security Strategy" (February 1, 2024) and "National Cyber Security Master Plan" (September 1, 2024).
National Intelligence Service et al. (2026). "2026 National Information Security White Paper." Jointly published by the National Intelligence Service, Ministry of Science and ICT, Ministry of the Interior and Safety, Personal Information Protection Commission, Financial Services Commission, and Ministry of Foreign Affairs. (Analysis of North Korea's strengthening cyber attack capabilities under its hostile two-nations theory, supply chain attacks targeting defense, energy, and advanced technology companies, theft of classified information, and virtual asset theft [Bybit]. * 'Agentic AI Autonomous Attacks' are discussed in Chapter 2, General Threat Outlook, with an actual case attributed to China-linked organizations [Anthropic, disclosed September 2025], not North Korea.)
Ministry of Foreign Affairs (April 26, 2023). "ROK-U.S. Strategic Cyber Security Cooperation Framework" (Adopted at the ROK-U.S. Summit).
AI Safety Institute (AISI) (Established November 27, 2024). Ministry of Science and ICT, Electronics and Telecommunications Research Institute (ETRI).
Statistics Korea (2022). "Key Statistics Indicators of North Korea 2022." (North Korea's power generation in 2021: 25.5 billion kWh - 16.1 billion kWh from hydropower and 9.4 billion kWh from thermal power, which is 4.4% of South Korea's 576.8 billion kWh.)
Jeon, Jae-seong (May 26, 2025). "[Global NK Commentary] The Art of the Deal vs. Structural Crisis: Trump-Style Negotiations and the Future of North Korean Denuclearization." East Asia Institute.
_______. (September 18, 2025). "[Global NK Commentary] Reshaping the World Order After China's Victory Day and the Rise of North Korea: Choices for South Korean Diplomacy." East Asia Institute.
_______. (January 2, 2026). "[New Year Special Commentary Series] ② Changes in U.S. Foreign Policy and the International Order in 2026." East Asia Institute.
_______. (February 2, 2026). "[Global NK Commentary] The Strategic Ambiguity of North Korea's Discourse on a Multipolar International Order." East Asia Institute.
_______. (May 11, 2026). "[Global NK Commentary] In Search of a New Direction for the South Korean Government's North Korea Policy." East Asia Institute.
_______. (June 8, 2026). "[EAI·JoongAng Sunday Joint Project] The 21st Century Kindleberger Trap… Neither the U.S. nor China Can Solve It Alone." East Asia Institute · JoongAng Sunday.
Ha, Young-sun (December 20, 2010). "[EAI Commentary No. 15] A Comprehensive Strategy for North Korea to Overcome the Yeonpyeong Island Crisis." East Asia Institute.
_______. (January 12, 2021). "[New Year Special Commentary Series - EAI South Korean Diplomacy 2021 Outlook and Strategy] ④ 8th Congress of the Workers' Party of Korea: North Korea's Next Five Years Stand Still." East Asia Institute.
English
Anthropic (2025.8.27). "Detecting and Countering Misuse of AI: August 2025."
Anthropic (2025.9.15). "Anthropic Economic Index Report: Uneven Geographic and Enterprise AI Adoption."
Anthropic (2026.1.15). "Anthropic Economic Index Report: Economic Primitives."
APEC (2025). "APEC Artificial Intelligence (AI) Initiative (2026–2030)." Annex to the Gyeongju Leaders' Declaration.
Bazoobandi, Sara (2024.8.1). "Development of the Knowledge-Based Economy in Iran." Arab Gulf States Institute in Washington (AGSI). (Basis for Khamenei's 2014 proclamation of "scientific jihad" in science and technology policy)
Chainalysis (2025.12). "2025 Crypto Theft Reaches $3.4 Billion." (North Korea stole approximately $2 billion in 2025, accumulating approximately $67.5 billion)
CISA·FBI·NSA·National Intelligence Service, et al. (2024.7.25). Joint Cybersecurity Advisory AA24-207A. (Theft of technology targeting the defense, aerospace, nuclear, and engineering sectors by organizations under North Korea's Reconnaissance General Bureau [Andariel]; jointly announced by the US, South Korea, UK, etc.)
Daily NK (March 12, 2025). (Report on intelligence theft organization affiliated with the Reconnaissance General Bureau using AI) Daily NK English (July 15, 2025). (Analysis of smartphone constant surveillance apps)
Ding, Jeffrey (2024). Technology and the Rise of Great Powers: How Diffusion Shapes Economic Competition. Princeton, NJ: Princeton University Press.
Eloundou, Tyna, Sam Manning, Pamela Mishkin & Daniel Rock (2024). "GPTs are GPTs: Labor Market Impact Potential of LLMs." Science 384(6702): 1306–1308. doi:10.1126/science.adj0998.
Fortune (June 4, 2025). (Report on real-time vocabulary substitution in North Korean smartphones, citing BBC analysis) Genians Security Center (September 15, 2025). "AI-Driven Deepfake-Based Military ID Forgery APT Campaign." (North Korean Kim Soo-ki's use of ChatGPT to forge South Korean military IDs targeting defense agencies)
Guston, D. (2014). "Understanding 'Anticipatory Governance'." Social Studies of Science 44(2): 218–242.
Henkel, Owen, Hannah Horne-Robinson, Nessie Kozhakhmetova & Amanda Lee (2024). "Effective and Scalable Math Support: Evidence on the Impact of an AI-Tutor on Math Achievement in Ghana." arXiv:2402.09809. (Ghana's math support Rori — operated by Rising Academies, joint RCT by Oxford·J-PAL, learning effect approx. 0.36 standard deviation; not a World Bank study)
IISS (March 2026). "Assessing North Korea's AI Ambitions." IISS Online Analysis. (Analysis suggests that while smuggling AI semiconductors (GPUs) via China is possible due to them being export-controlled items, large-scale acquisition is very difficult.)
IMF (2024). AI Preparedness Index (AIPI). IMF Datamapper. (174 countries assessed as of 2023 — North Korea excluded due to lack of data)
Jones, Lee (2014). "Explaining Myanmar's Regime Transition: The Periphery is Central." Democratization 21(5): 780–802. doi:10.1080/13510347.2013.863878.
Kim, Hyuk (August 21, 2024). "North Korea's International Network for Artificial Intelligence Research." 38 North.
Kokabisaghi, F. et al. (2019). "Impact of United States Political Sanctions on International Collaborations and Research in Iran." BMJ Global Health 4(5): e001692. doi:10.1136/bmjgh-2019-001692. Microsoft AI Economy Institute (2026.1). Global AI Adoption in 2025: A Widening Digital Divide. Redmond: Microsoft.
MSMT (2025.10.22). The 2nd Report of the Multinational Sanctions Monitoring Team.
OECD (2025). Governing with Artificial Intelligence: The State of Play and Way Forward in Core Government Functions. Paris: OECD Publishing. doi:10.1787/795de142-en.
Oxford Insights (2026.1). Government AI Readiness Index 2025.
NK Insider (May 27, 2025). "North Korea and China Introduce AI-Based Surveillance at Border, Share Information About Unusual Activity."
NK News (July 7, 2025). "What China's DeepSeek AI Reveals About Beijing's Red Lines on North Korea." NK News (July 22, 2025). "North Korea Dispatching AI Researchers to Russia, DPRK Academic Says."
NK News (September 2021). "North Korea Refuses COVAX Offer of 3 Million Sinovac Vaccine Doses." (Refusal of COVAX vaccines in 2021 - reasons included conditions for international monitoring.)
NPR (June 10, 2026). "China Re-centers North Korea Ties as Nuclear Silence Reshapes Balance." (Xi Jinping's state visit to Pyongyang, June 8-9, 2026 - absence of denuclearization mention in joint statements; cross-referenced with AP, Reuters, DW.)
Pak, U., Y.B. Son, K. Kim, J.H. Kim, M.J. Jang, K.J. Kim & G.R. Pak (2024.11). "Novel Particulate Matter (PM2. 5) Forecasting Method Based on Deep Learning with Suitable Spatiotemporal Correlation Analysis." Journal of Atmospheric and Solar-Terrestrial Physics 264: 106336. doi:10.1016/j.jastp.2024.106336.
Path, K. (2020). "The Origins and Evolution of Vietnam's Doi Moi Foreign Policy of 1986." TRaNS: Trans-Regional and -National Studies of Southeast Asia 8(2): 171–185. doi:10.1017/trn.2020.3.
Pentón-Arias, E. & J.C. Aguilar-Rubido (2021). "Cuban Prophylactic and Therapeutic Vaccines for Controlling Hepatitis B." MEDICC Review 23(1): 21–29. doi:10.37757/MR2021.V23.N1.6. (PAHO/WHO certification of the Hepatitis B vaccine Heberbiova c HB [4 doses in 2001–2003] · supplied to over 30 countries)
Sloterdijk, Peter (2016). Foams: Spheres Volume III — Plural Spherology. Trans. Wieland Hoban. South Pa sadena, CA: Semiotext(e). (Original German Sphä ren III: Schäume, Suhrkamp, 2004)
Stanford HAI (2026.4). The 2026 AI Index Report. Stanford Institute for Human-Centered AI.
The Korea Times (2026.5.28). (UNDP Administrator's visit to Korea · Seoul Global AI Hub)
Thorson, Stuart J. (2012). "Universities and Networks: Scientific Engagement with North Korea." Science & Diplomacy 1(2).
UN DESA (2024). United Nations E-Government Survey 2024. New York: United Nations. (North Korea's EGDI 0.23, ranking 184th out of 193 countries; absolute score declined from 0.3616 in 2012, ranking 131st; South Korea ranks 4th globally — UN E-Government Knowledgebase Country Data)
UN GGE (2015). "Report of the Group of Governmental Experts on Developments in the Field of Informa tion and Telecommunications in the Context of International Security." A/70/174. (11 norms of responsible state behavior — 13(f) not attacking other states' critical infrastructure, 13(c) not allowing ICT illicit acts from one's territory, etc.; South Korea participated, application of international law to cyberspace reaffirmed in 2021 GGE A/76/135)
UNDP (2025). The Next Great Divergence: Why AI May Widen Inequality Between Countries. UNDP Regi onal Bureau for Asia and the Pacific (2025.12).
UN Security Council, 1718 Committee. "Approved Exemption Requests" (List of approved humanitarian exemptions, as of March 2026; basis Paragraph 25 of S/RES/2397(2017) and Implementation Support Notification No.7 [adopted August 6, 2018 · updated June 2, 2023]). Sanctions resolutions cited in the main text: S/RES/2270(2016)·2321(2016)·2375(2017)·2397(2017).
UN DPRK (2026.5.5). (Meeting between UN Resident Coordinator and North Korea's Central Bureau of Statistics on the occasion of the FAO-AIT program in Bangkok)
UNICEF (2024.7.18). "Life-Saving Vaccines for Children and Pregnant Women Delivered to Democratic Pe ople's Republic of Korea." UNICEF East Asia and Pacific Press Release.
U.S. White House (2017.12.19). "Press Briefing on the Attribution of the WannaCry Malware Attack to No rth Korea." (Official attribution of WannaCry malware to North Korea by the U.S. government; joint condemnation by the UK, Australia, Canada, New Zealand, and Japan; paralysis of UK NHS hospitals)
UNICEF·WHO·World Bank Group (2026). Levels and Trends in Child Malnutrition: Joint Child Malnutrition Estimates — Country Estimates (Released March 2026).
WHO (2021). WHO Consolidated Guidelines on Tuberculosis, Module 2: Screening — Systematic Screeni ng for Tuberculosis Disease. Geneva: WHO.
WHO (2025). Global Tuberculosis Report 2025. Geneva: WHO. (Estimated tuberculosis incidence in North Korea 'under review'; approximately 1/3 of global tuberculosis patients undiagnosed)
WHO·UNICEF. Joint Estimates of Vaccination Coverage (WUENIC).
Williams, Martyn (2026.3.18). "The Big Expansion of North Korea's Cellular Sector." 38 North.
World Bank (2025.6.25). "Addressing the Learning Crisis with Generative AI: Lessons from Edo State in Ni geria" (De Simone, Martín E. et al.). World Bank Blogs — Development Talk.
■ Choi, Inho_Senior Researcher, Institute for International Affairs, Seoul National University; Lecturer, Seoul National University.
■ Lee, Sangjun_EAI Researcher
Inquiries: 02 2277 1683 (ext. 211) | leesj@eai.or.kr
*This text is an AI translation of an original written in Korean. Some translations or nuances may be inaccurate.