I don’t need this “AI nonsense.” (Probasco 2023)

Adopting processes that exploit AI-enabled autonomy across the battle network is the path to achieving a higher relative system operating tempo than US competitors. (Work 2020)

What is France’s, Germany’s, or Russia’s approach to defense artificial intelligence (AI)? How does India or Iran think about defense AI? What are the defense AI development priorities of the United States, Israel, South Korea, or Singapore? Is the defense AI approach of authoritarian countries distinctively different from the practice in democratic countries? How do ongoing conflicts shape the defense AI trajectory?

Analysts interested in answering these questions fight a steep uphill battle as information is scattered across a rapidly growing body of literature looking at the way AI is likely to shape military thinking and warfighting practice. To solve this problem and advance the international understanding of how nations approach and implement defense AI, The Very Long Game provides the first collective in-depth analysis of 25 nations that is both comprehensive and comparative, as well as easily accessible. Rather than speculating about how AI could contribute to military power and change warfighting, this book focuses on what nations do now to use defense AI today and in the future. This approach, the current chapter will show, tames hyperbolic imaginaries that range from ascribing superhuman powers to defense AI to portraying the technology as the harbinger of dystopian war scenarios.

By focusing on the current state of play related to defense AI, this volume sits at the intersection of strategic affairs, military innovation, organizational change, emerging technologies, and future force development trajectories. It brings together a diverse set of authors and sheds light on cross-regional commonalities and regional specifics that need to be considered when reflecting upon the interplay between military power and technology. The selected case studies look beyond Canada, Denmark, Estonia, Finland, France, Germany, Greece, Italy, the Netherlands, Spain, Sweden, Turkey, the United Kingdom and the United States as member states of the European Union (EU) and/or the North Atlantic Treaty Organization (NATO) and also cover Australia, China, India, Iran, Israel, Japan, South Korea, Russia, Singapore, Taiwan, and Ukraine to capture the dynamics in different strategic theaters. These 25 country studies look at well-established as well as ambitious new defense exporters to get a better understanding of how defense AI could influence future defense export prospects and provide inroads for newcomers to unlock mature markets. This volume also considers military thought leaders and countries that aspire to become role models for other nations, thus tracking the defense angle of an increasingly “multiplex” world order characterized by the absence of global hegemony through any single nation, broader patterns of interdependence, and intellectual and political diversity sketching out different ways to ensure peace and stability (Acharya et al. 2023: 2341). Furthermore, this volume presents case studies on the use of defense AI in war zones as well as the strategically contested regions of North-East Asia, the Eastern Mediterranean and wider Arab Gulf.

In all the countries selected for this volume, AI and other emerging technologies rank high on the agenda. The assumptions underpinning these agendas are diverse, as the two introductory quotations make clear. In general, there is a growing belief that AI is likely to change future warfare and could also tip the strategic balance. Warfighters, as the Ukrainian voice cited in Probasco’s opening quotation, may beg to disagree. They have a point. Despite the hype around emerging technologies, technology alone is insufficient to modify the way armed forces operate and drive change. Rather, technology needs to be embedded in the broader cultural, conceptual, and organizational context.

This understanding is in line with a growing body of literature on military innovation and the role of defense AI. Focusing on a selection of publications that appeared within the last three to five years, three groups can be identified. First, the number of texts making general assumptions about the military impact of defense AI is growing quickly. This group includes books like Army of None and Four Battlegrounds by Paul Scharre (Scharre 2019; Scharre 2023), Johnson’s analysis of the interplay between defense AI, future wars, and strategic stability (Johnson 2021), and Kenneth Payne’s I, Warbot (Payne 2021). Reflections on the possible impact of AI on international stability and the role of arms control in preventing an AI “arms race” also belong to this group (Cummings 2018; Diehl and Lambach 2022; Horowitz 2018; Horowitz et al. 2018; Horowitz and Scharre 2021; Scharre and Lamberth 2022; Scharre 2021). While primarily interested in the interplay between defense AI and strategic affairs, most of these texts only look at a limited number of countries and combine examples of the current use of defense AI with speculations about its future impact. A combination of AI with uncrewed systems as well as ethical and regulatory considerations is commonplace as well.

In addition to these broad treatises, a second group of authors tries to understand how AI might affect military power with a focus on specific military domains and what factors enable or block the diffusion of AI and information-driven ideas of change. This body of literature includes, for example, Sam Tangredi’s and George V. Galdorisi’s book on AI and naval warfare (Tangredi and Galdorisi 2021) and the analysis of defense innovation in the information age by Jensen et al. (2022). These texts focus on understanding AI’s added value for operators and/or theoretical advancements. Along these lines, Lin-Greenberg (Lin-Greenberg 2020) scrutinize the impact of AI on coalition decision-making, while Lindsay (Lindsay 2023/2024) sheds light on the institutional context for AI-enhanced military innovation. Analyses of the risks posed by defense AI form a prominent sub-group inside this specialized impact literature, for example, with papers focusing on countering intelligence algorithms (Phillips and Pohl 2021) or the use of AI in wargames (Barzhaskha 2023). This literature is also linked to the prominent discourse on the ethics of defense AI (CIGI Undated; Hofstetter and Verbovzsky 2023; Galliott and Scholz 2020; Stanley-Lockman 2021; Rowe 2022).

The third group of publications takes a more comparative approach by focusing on the specific approaches undertaken in different countries. Artificial Intelligence, China, Russia, and the Global Order, edited by Nicholas D. Wright (Wright 2019), for example, covers different aspects of defense AI in the two countries mentioned but does not adopt a comprehensive analytical approach to provide cross country assessments. By contrast, The AI Wave in Defense Innovation. Assessing Military Artificial Intelligence, Strategic Capabilities and Trajectories, edited by Raska and Bitzinger (Raska and Bitzinger 2023), presents case studies on defense AI in the US, China, Russia, Japan, and South Korea as well as Australia. Moreover, one chapter looks at the use of military AI in Europe with an EU/NATO focus.

As this short overview illustrates, most of the literature is focusing on broad questions while somewhat pushing aside the more complex and protracted issues of how countries specifically think about defense AI, how they prepare for its adoption, and how they develop existing concepts, structures, processes and develop the respective capabilities (Goldfarb and Lindsay 2021/2022: 11). This is the gap that this volume is seeking to address. Therefore, the editors started in 2022 to commission the 25 case studies brought together in updated and abridged chapters for this volume.Footnote 1

Together with other technologies like robotics and computing, AI is considered an emerging technology likely to prompt military innovation. As argued elsewhere (Borchert et al. 2021: 13–17), military innovation is a popular but controversial concept as it is challenging to gauge what constitutes innovation. For military innovation to occur, armed forces need to master the advent of new technology in tandem with conceptual, cultural, and organizational transformation. That’s why the case studies understand defense AI as a socio-technological phenomenon that requires a broad analytical framework. Each chapter follows the same structure—inspired by the DOTLMPFIFootnote 2 lines of effort—to discuss how nations:

  • think about defense AI to illustrate how existing military concepts define and shape the use of defense AI;

  • develop defense AI to outline existing defense research and development (R&D) priorities related to AI, highlight important research projects and priorities, and portray the national defense AI ecosystem;

  • organize defense AI with a focus on specific structures and processes that have been put in place to advance organizational adaption of defense AI;

  • fund defense AI to highlight spending priorities;

  • field and operate defense AI to give an overview to what extent AI is already used to support existing military missions;

  • train for defense AI to analyze how armed forces prepare for a future in which they cooperate with cognitive machine intelligence.

In so doing, each chapter looks at the interplay of four critical elements relevant for capability development and force adaption: First, there are ideas, concepts, and policies that underpin military power. As the case studies highlight, most nations understand defense AI as a means to improve existing practice, rather than to embark on exploring new ways of applying military power. This serves as a cautionary tale and deflates much of the rhetoric about the “disruptive” power of defense AI.

Second, novel ideas need champions to diffuse them within and across hierarchies. In addition to intra- and interservice rivalries that the military innovation literature has identified as stimuli for change (Mansoor and Murray 2019: 11–14), or the idea that software-driven modernization can enable military powers to leapfrog on capability growth (Soare 2023), the following case studies identify incumbent and emerging new defense companies as important technology bridges to induce change. This, however, raises thorny questions related to controlling the exports of companies that serve military and commercial clients, the relationship between sovereign and private algorithm development, and the acceptance of “black boxes” when defense AI solutions are embedded in foreign weapons supplies.

The rise of the “new kids on the defense block” is closely related to each country’s institutional setup as the third important aspect to consider. Institutional maturity and complexity can create market barriers for new defense suppliers and cause organizational pathologies that render the successful application of defense AI almost impossible, as the case study on defense AI in Canada makes clear. Most nations play around with new organizational elements and enhance cross-institutional coordination for defense AI. As the case studies show, military organizational change in the US receives a lot of attention as many nations seek to emulate what the US Department of Defense is doing.

Finally, change will never materialize without resources. As argued below, adapting the business models of digital platform providers that are awash in data that consumers share freely in return for products and services, is of limited use for defense AI. The consumer data reality has nothing in common with the defense data reality. Adhering to data-centricity, although popular among all armed forces discussed in this volume, is likely to slow down, rather than boost the military’s use of AI.

Against this background, the remainder of this chapter will first provide a quick introduction into the technicalities of AI. A summary of the main findings of all case studies follows second. Rather than providing country-specific abstracts, I use the six lines of effort discussed above to advance a cross-country understanding of who is doing what, why and how concerning defense AI. Third, the chapter will interpret these findings and suggest that some of them confirm existing expectations, whereas others are surprising and tell uncomfortable truths. Concluding remarks related to contemporary imaginaries underpinning the defense AI discourse round off the chapter.

1 Defense AI: What’s in a Word

AI is popular, but challenging as there is no unifying definition. In one way or another most definitions used by the armed forces analyzed in this volume coalesce around the notion—put forward in the AI strategy of the US Department of Defense (Department of Defense 2018: 4)—that AI “refers to the ability of machines to perform tasks that normally require human intelligence.” This definition is straightforward but can cause controversy as it challenges the notion—underpinning the thinking on defense AI prevalent in almost all nations analyzed in this volume—that humans must always control machine intelligence and action. According to Brandlhuber (Brandlhuber 2021: 6), a less contentious understanding of AI would thus emphasize the role of hardware and software systems in implementing reasoning mechanisms without the need to pre-program the solutions that these systems are expected to accomplish.

In so doing, AI uses and depends on different methods with much of the literature focusing on machine learning (ML). ML posits that “knowledge is learned from data” and that the respective algorithms run “on a training dataset and produce an AI model” (Allen 2020: 3). This notion has become very popular due to the prevalence of digital commercial business models that seek to leverage (consumer) data as well as the rapid improvement of computing resources to handle large volumes of data. The ML focus, however, has created a narrow perspective on AI.

The problem stems from overemphasizing the role of data. Digital business models for consumer markets constitute the implicit reference point. These are data-centric because users provide their data in exchange for digital products and services. In addition, digital companies want to work “model-free” as AI systems should replace and scale the steps that were previously developed using specialist expertise and engineering skills. To this end, ML systems should process large amounts of data to reproduce a desired behavioral pattern.

This data-centric logic has also captured military thinking. All the countries surveyed describe data as their most important strategic asset and are gearing data strategies towards exploiting this “data treasure,” also by investing in powerful hardware. But armed forces do not operate in a consumer environment with abundant data; rather they struggle with data scarcity.Footnote 3 Data-centric approaches are resource-intensive and require personnel, computing power, energy, infrastructure, bandwidth and recording time, which become even scarcer in the event of war (Chahal et al. 2020: 10–13; Michel 2023: 16–21). Paradoxically, they are also past-oriented, i.e., armed forces can only evaluate what has been collected. While collected data can describe the dynamics of the past, it cannot describe dynamic operating principles of the physical environment in which armed forces operate (Borchert et al. 2023a: 47).

In contrast to this prominent view, military users should be cognizant of the fact that AI may be a general-purpose technology (Horowitz 2018: 39–41; Scharre 2023: 3), but it also is a “bag full of methods” (Hofstetter 2014: 136–142). Some AI methods are better suited than others to address different military tasks. Being precise about which method best suits what task is important to set the goals for defense AI to achieve. Four general approaches are important to distinguish (Brandlhuber 2021: 14; Allen 2020: 4):

  • Unsupervised learning, which does not require labels for data used, is helpful for data analytics, anomaly detection or auto-coding.

  • Supervised learning, which uses data labelled by human operators, is used for speech or image recognition, video analysis, auto translation, and to classify signals.

  • Reinforcement learning, which allows for AI agents to generate their own data based on interacting with the relevant environment, in which they operate, is instrumental for optimal sequencing of actions over several iterations, hedging strategies to support risk management or strategic decision-making.

  • Finally, cooperation and emergence result from AI agents that interact with their environment and are useful for dynamic resource management (e.g., to improve sensor capabilities), optimal resource sharing (e.g., optimal allocation of sensors and effectors to engage targets), or efficient routing.

Against this background, the US Defense Advanced Research Projects Agency (DARPA) has proposed differentiating three waves of AI (DARPA Undated; Borchert et al. 2023b: 27):

  • First wave AI leverages handcrafted knowledge. Human experts construct expert systems that capture the specialized knowledge of human experts in rules that systems can apply.

  • Second wave AI focuses on correlational learning. Statistical and probabilistic methods are used to train neural networks to perform classification and prediction tasks.

  • Third wave AI emphasizes contextual reasoning. Here, computing systems have full situational awareness, which means that these systems reason in context and understand the consequences of their action and actions undertaken by third parties like adversaries, for example.

The transition from second to third wave AI is essential to understand given the latter’s ability to self-learn. First and second wave AI focus on extracting patterns from (big) data by using classification and regression. Third wave AI, in contrast, strives to develop solutions that learn how best to learn by using context-aware, complex, and multi-stage decision-making commensurate with the relevant operational environment, mission tasks, and overall rules of engagement. This is pivotal for defense AI, because reinforcement learning, which is important for ML, has yielded impressive performance results in ideal-type games with perfect information (Silver et al. 2017; Vinyals et al. 2019). Imperfect information and uncertainty, however, are typical for military operations. That’s why third wave AI needs to address, for example, the fact that tactical military decisions occur along a decision sequence in which prior decisions are contingent for later decisions, non-decisions may significantly limit a commander’s future freedom of action, and adversarial operations in the electromagnetic spectrum may jam or even neutralize sensors and thus hamper situational awareness and situational understanding. Third wave AI thus needs to be able to interpret this mission-relevant context to avoid prejudicing decisions to hamper mission success at later stages.

Markov-Decision-Processes constitute one mathematical approach relevant to tackle these challenges as they model decision-making in an environment, “which changes state randomly in response to action choices made by the decision maker” (Littman 2001). Additional methods include adversarial learning to advance the robustness of ML (Bai et al. 2021), transfer learning (Zhuang et al. 2019) to develop defense AI solutions that can be transposed from the defense metaverse—that digitally mimics key parameters of the battlefield (Borchert et al. 2023b)—into physical reality, as well as meta learning (Vettoruzzo et al. 2023) to enable decision-making policies to evolve commensurate with a non-stationary mission environment that changes over time. As a result, third wave AI addresses uncertainty by emphasizing emergence, not linearity or regularity, as the key principle and posits that formation, rather than formulation, matters for successful adaption (Mintzberg et al. 2005: 177; Popescu 2018). Essentially, emergence also emphasizes the “capacity to experiment” (Mintzberg et al. 2005: 189) to explore new avenues and exploit existing avenues at the same time (Reeves et al. 2013).

Distinguishing between three waves of AI is useful to consider what defense AI can accomplish. As Table 1 illustrates, the contribution of defense AI to implement the principles of war that describe how to use military power, becomes more significant the more military users envision employing third wave AI. So far, however, almost all nations portrayed in this volume focus on second wave AI that emphasizes the central role of data exchange via digital platforms. This “hub and spoke” logic works for benign environments, in which centralizing is efficient and provides scale. The military combat environment, however, is non-benign and requires AI solutions that operate under uncertainty. That’s why successfully using defense AI on the battlefield will require armed forces to transition to decentralized, self-learning solutions that third wave AI empowers (Bousquet 2022: 210–211).

Table 1 Contribution of three waves of defense AI to the principles of war
Full size table

2 Who Does What, Why, and How?

This section provides a comparative summary of the 25 case studies along the six lines of effort discussed above. Before moving on, a word of caution is needed. Although detailed, the case studies, which are based on open-source information and expert interviews, only provide a snapshot of the national defense AI endeavors in late 2023 and early 2024. Based on the case studies, the summary is selective. In addition, the summary also simplifies, for example, by equating a single project with a national focus on developing or deploying defense AI. This, in turn, might not always adequately describe the respective levels of maturity and could inflate the importance of individual projects or initiatives.

2.1 Thinking About Defense AI

The countries analyzed are not very specific about the concrete goals defense AI is meant to accomplish. Most descriptions remain generic referring to potential gains such as being able to analyze ever-growing datasets, sneak into the enemy’s OODA loop (Orient, Observe, Decide, Act), or accelerate decision-making. Overall, most armed forces want to improve and optimize what they are doing today, rather than exploring how using defense AI might empower them to accomplish new tasks and missions. What may surprise at first sight, becomes less puzzling when looking at the underlying drivers shaping national defense AI perspectives.

2.1.1 Strategic Motives

Three strategic motives are at play (Table 2). The first group of nations has adopted a threat-based approach. In this case, one strategic challenger or a complex set of regional challenges has been identified, and defense AI is seen as a key solution. The textbook example are China and the United States, which see each other as prime challengers and focus on defense AI to contain the other. The same logic applies to Russia and to neighboring pairs of countries like Greece and Turkey, South Korea and Japan (vs. China and North Korea), as well as Ukraine and Russia. Israel and India also operate under this perspective, but with additional regional drivers shaping their view. India, for example, is also concerned that defense AI might empower other nations to dominate the country, which prompts New Delhi to invest in indigenous solutions. A similar logic is at play in Iran, which considers Israel, the United States, and their allies in the region as an existential threat shaping its approach to defense AI. Finally, Taiwan is a special case as all three strategic motives are at play with the threat from China as the key driver.

Table 2 Three strategic drivers of defense AI
Full size table

The second group of countries fears missing out (FOMO) or falling behind. This is a different threat-based perspective that emphasizes the competitive disadvantage that could result from the inability to embrace defense AI. This includes well-established defense exporters like France and Italy, transatlanticist Denmark, as well as Greece, whose armed forces went through several years of underfunding after the international financial crisis in 2008/2009. In Athens’ case, the fear of missing out is also directly linked to the threats the country sees originating from Turkey.

Most of the countries belong to the third group, which takes a less pronounced position and interprets AI primarily as a capability multiplier. Members of the other two groups also share this perspective but drive defense AI development via threat or FOMO-based foci. Adherents of the capability multiplier perspective, by contrast, tend to have less well-defined development and deployment priorities as they put more emphasis on exploring various defense AI use cases.

2.1.2 Who Shapes Whom?

These strategic motives also play a role when asking if and to what extent others are shaping national defense AI perspectives. Partners, Table 3 illustrates, play a foundational role. Here, the United States remains the pivotal player, particularly for its allies in the Asia-Pacific region. European partners of the United States also look to other members of NATO and the EU for inspiration. In this context, inspiration implies that partners adjust their defense capstone documents within the context of US thinking, for example by emphasizing the role of AI in Multi-Domain Operations, mimic US organizational reforms, and emulate US defense systems concepts and architectures, like the Joint All-Domain Command and Control (JADC2) concept, to remain interoperable with the US. Emulation, however, also comes with risks. Taiwan’s AI-pilot project—modelled after DARPA’s Alpha Dogfight—illustrates these risks by proving to be “far-reaching but conceptually ill-informed,” as the chapter notes.

Table 3 Who shapes the thinking on defense AI?
Full size table

Strategic challengers are as powerful as partners and allies in shaping defense AI perspectives. Again, the case studies on China and the United States highlight how closely both nations monitor each other’s defense AI moves. In addition, the threat-based perspective discussed above also shapes the perspectives of those countries that feel existentially threatened by neighbors or a cocktail of different risks. Interestingly, the match is not always perfect. While Greece’s thinking on defense AI is motivated by the military and defense industrial challenges posed by Turkey, the latter is more and more replacing its traditional focus on the US by emphasizing its own foreign policy and indigenous defense industrial ambitions.

Between these two poles, only few countries remain. France traditionally emphasizes its role as a self-determined nuclear power and tries to chart its own course regarding defense AI. Sweden and Spain, by contrast, are a bit more difficult to locate. Spain’s national security strategy talks about the China challenge, but it is not much of a strategic driver for its thinking on defense AI, which seems more influenced by initiatives driven from inside NATO and the EU. Sweden, traditionally very close the US, takes an agnostic perspective as its thinking is shaped by internal drivers like total defense and NATO.

2.1.3 Human or Tech-Centric Understanding

Whether countries adopt a human or a technology-centric approach to defense AI is important to understand the relationship between human operators and machines. The human-centric approach builds on the idea that AI is to complement, not replace human beings. By contrast, a technology-centric approach posits that AI shall facilitate and accelerate full technical autonomy and machine-machine interaction, as illustrated in the quotation by Bob Work, former US Deputy Secretary of Defense, at the beginning of this chapter.

Table 4 illustrates that most of the 25 nations adopt a human-centric approach, but with notable nuances. Estonia, for example, supports a human-centric approach but is not a “normative hawk” on regulation, as discussed below. South Korea belongs to this group as well but demographic aging and a dramatically shrinking personnel basis of the armed forces could become drivers for a more technology-centric view in the future. The US, for the time being, is also clearly human-centric but the most recent Replicator initiative announced by the Department of Defense envisions a future with swarms of uncrewed assets to overwhelm adversaries that is firmly anchored in the technology-centric perspective (Hicks 2023; Tucker 2024).

Table 4 Human or tech-centric understanding of defense AI
Full size table

Already today, two countries edge towards the technology-centric approach. Turkey wants to leverage AI in tandem with the country’s expertise on uncrewed assets to advance autonomous operations across domains. Turkish defense company Havelsan has been championing the “Digital Troops” concept that reflects this idea and is building a product portfolio around it. In addition, the country’s defense engineers posit that it is more difficult to arrange for man-machine integration than to enable machine-machine interaction, which adds to the country’s edge towards a technology-centric approach. Ukraine follows a similar idea borne out of the current war. Achieving machine autonomy with defense AI is one of the country’s declared development priorities for the near future, also because the war shows that connectivity that provides one option for human operators to continue piloting uncrewed assets, is brittle in combat.

Finally, there are several nations that take an agnostic view, but across this group there are strong drivers that point towards the likelihood of a more technology-centric stance in the future. Denmark’s need for wide area surveillance with uncrewed assets and defense AI could entice a more technology-centric approach. Estonia, Iran, Singapore, and Taiwan see the option to free up scarce manpower with the help of AI and uncrewed systems. Greece is still agnostic, but strategic rivalry with Turkey could tilt the balance depending on where Turkey is heading. Russia’s position is moving between both poles, in particular regarding the interplay between defense AI and lethal autonomous weapon systems (LAWS). Finally, Sweden, like Denmark, sees a need for wide area surveillance, for which defense AI could augment uncrewed assets. In addition, the need to counter hypersonic weapons could prompt a more technology-centric stance with defense AI playing a prominent role in analyzing data related to this specific threat.

2.2 Developing Defense AI

Before addressing current defense AI development priorities in detail, it is worth asking which of the three waves of defense AI shapes national mindsets. As argued above, this differentiation serves as an indicator for defense AI applications that are aligned with the status quo (second wave AI) or illustrate potential “breakout” solutions (third wave AI).

The overwhelming majority of the countries (Table 5) follows a data-focused understanding of defense AI, whereas the United States is the only nation that has so far officially discussed and explored the military benefits of third wave AI thanks to dedicated programs managed by DARPA. Few countries hold an agnostic position. Estonia, Iran, and Spain are in the process of establishing defense AI practices and might still need to mature their respective thoughts and concepts. Turkey’s development priorities indicate a weak tilt towards emergence in combination with uncrewed systems. Canada’s approach to defense AI suffers from current organizational stovepipes.

Table 5 Data or emergence-based defense AI development
Full size table

Combining the data vs emergence perspective with technology and human-centric approaches to defense AI produces a most interesting 2x2 matrix depicted in Fig. 1. This chart clearly underlines that for the time being all 25 nations operate within the same data and human-centric paradigm. When challengers and incumbents operate along the same mindset, surprise will hardly manifest. But the matrix also clearly suggests that there are different breakout trajectories that could materialize in the near future:

  • The United States is the most prominent candidate to embrace emergence for future operations, as the recent Replicator initiative illustrates. US concepts like JADC2 would also greatly benefit from a move towards decentralized and horizontal, rather than vertical, command and control (C2) approaches, but current service preference seem to stand in the way of fully leveraging this idea.

  • Once the United States move towards emergence, China is likely to follow suit as it wants to play on par. Whether China would move towards emergence via emphasizing machine autonomy before embracing context-aware defense AI solutions is hard to say right now. The country’s most recent experiments with large language models (LLM) to support swarm C2 could be seen as an early indicator towards this direction. However, if an LLM-based swarm of autonomous systems would withstand battlefield realities is doubtful as its ability to generate forward-looking tactics that are difficult for adversaries to mimic is likely limited.

  • Ukraine, which recently decided to set up a new unmanned systems branch (President of Ukraine 2024), and Israel also constitute two obvious candidates for change as both see value in advancing machine autonomy with AI given current battlefield realities. The same is true for Turkey, which is leveraging defense AI in tandem with the country’s broad portfolio of uncrewed systems across all domains.

  • Finally, Iran and South Korea both qualify as potential adopters of a more pronounced machine autonomy paradigm. Both see value in compensating demographic aging and shrinking armed forces with more technical autonomy. And Iran, given the erroneous downing of Ukrainian International Airline flight 572 in 2020 by a ground-based air defense battery, considers more autonomy to avoid human errors.

Against this background, Table 6 summarizes today’s key defense AI development use cases across all case studies. The table shows interesting priority clusters. First, a regional perspective suggests that defense AI priorities play a prominent role in EU and NATO countries. This may be related to the technological maturity of the armed forces as well as current and most recent wars in Europe and adjacent regions. Defense AI development priorities are less pronounced across the Greater Middle East. China, Russia, and Turkey have ambitious defense export plans for the region and could thus set the pace regarding diffusing defense AI across the respective countries. Israel’s opportunities to shape regional defense AI capabilities, however, seem much more muted as a fallout of the Gaza war that questions the efforts to normalize relations with the country, particularly among the Gulf states. Apart from Australia and China, defense AI development priorities remain nascent across the Asia-Pacific countries analyzed. The fact that Japan, South Korea, Taiwan, and Singapore maintain close relations with the United States is important as these nations will want to develop their defense AI capabilities in tandem with the US and reach out to US defense AI suppliers for help, as developments in Australia illustrate.

Fig. 1
A 4 paradigm 2 by 2 matrix to develop defense A I. The rows are for a human centric perspective and technology centric perspective, from bottom to top. The columns are for data-driven defense A I and context aware defense A I from left to right. The cells are labeled field 1, 2, 3, and 4.

Four paradigms to develop defense AI

Full size image
Table 6 International defense AI development use cases
Full size table

Second, current defense AI development priorities underpin several capability categories:

  • The combination of AI with uncrewed systems is the most prevalent area of application. Here, quality differences stem from the portfolio of uncrewed assets maintained today (domain specific or multi-domain preferences) and the tasks these systems are expected to perform. Most often, these tasks focus on intelligence, surveillance, and reconnaissance (ISR) either to improve common operational pictures (situational awareness and situational understanding) or support targeting and strike. In this regard, manned-unmanned teaming is relevant as well.

  • Predictive maintenance, logistics, as well as cyber operations constitute a second group that is directly related to the prevailing preference for data-driven AI solutions, which is also reflected in the focus on data analytics and data management.

  • C2 in combination with data analytics and data management forms a third priority capability cluster as many nations look at AI for help in assessing growing data volumes. Few countries, by contrast, want to use AI to develop new C2 approaches.

  • While AI for electronic warfare (EW) and to counter adversarial EW plays a role for around a third of the countries, AI for use in missiles, torpedoes, fire support and air defense rank lower on the development agendas. It will be interesting to see to what extent the Russia-Ukraine war and current missile attacks against Israel will shift these priorities.

  • While around a third of the countries consider defense AI for red teaming and wargaming, far fewer countries develop defense AI for (mission) planning and tactics development. Interestingly, Germany has paved the ground for exactly these types of future applications, which may be a harbinger of significant change if the German Ministry of Defense (MoD) were to succeed in transitioning development results into mature defense systems and capabilities. Here, lessons from the war in Ukraine could serve as an accelerator, in particular, for future German air/missile defense solutions.

  • Finally, the significant congruence between Russian and Chinese defense AI development priorities is noteworthy. Among other aspects, both nations share an interest in defense AI for target identification, EW, and swarming, which could, in combination with the focus on uncrewed systems, point towards fully autonomous reconnaissance-strike complexes able to operate under adversarial electromagnetic spectrum dominance in the future.

Reference to congruent defense AI priorities highlights the fact that more and more countries look at opportunities to co-develop defense AI capabilities with partners.Footnote 4 Pillar two cooperation among Australia, the United Kingdom, and the United States (AUKUS) is one example illustrating a multinational capability development strand. Cooperation between the United Kingdom, Italy, and Japan in the Global Combat Air Program (GCAP) could also address defense AI in the future. And multinational development projects that receive co-funding via the European Defense Fund (EDF) play a key role for defense AI development in Estonia, Finland, Italy, Greece, and Spain.

As attractive as multinational defense cooperation may be it will raise thorny questions regarding the future (national) ownership of (multinational) defense AI efforts and the interoperability between AI building blocks originating from different countries operating under heterogeneous data protection/sharing regimes. Even more importantly, multinational defense AI efforts are likely to run into national interests, as, for example, Australia, Germany, France, the Netherlands, and the United Kingdom consider defense AI a sovereign defense industrial capability—which in turn might limit multinational cooperation. The Netherlands, for example, have only tasked TNO, the government-owned research entity, not industry, to develop defense AI algorithms that are intimately related to the way its armed forces plan and execute certain military operations.

A final note is due on the question of who adopts from whom, which matters for national and multinational development efforts. Developing AI against the background of an already existing system or platform will always shape AI in line with the prime tasks of the respective system. By contrast, enabling new behavior on the battlefield by developing AI tactics, would reshape the design and performance parameters of these systems and platforms in line with tactics—but might be less palatable because of this. There is no right or wrong on this question but given the congruence in defense AI development priorities to improve existing solutions discussed above, the latter will likely become more important in the future to achieve and sustain a competitive edge.

2.3 Organizing Defense AI

According to Jensen et al. (2022: 3), structure is one of two key factors determining if and to what extent the use of information technology will lead to military innovation, because the “structure of institutions channels the flow and interchange of information across the formation.”Footnote 5 Whether nations undergo organizational reform to prepare for the use of defense AI is thus an important question. The case studies send mixed signals on how reform could or should look like (Table 7).

Table 7 How to organize defense AI
Full size table

The United States and France are among the few countries that have set up new project-specific organizations to advance defense AI. Project Maven in the US explored the use of defense AI for ISR and made progress thanks to the new, small, dedicated project structure—but later faced organizational challenges and resistance upon transitioning into the prevailing functional organization. France set up ARTEMIS.IA as a public-private partnership to provide the armed forces with a set of different applications to analyze big defense data volumes but was not able to deliver the key goal of identifying and selecting best-in-class proposals because loopholes in the design provided options for circumventing the setup’s competitive nature.

Several countries decided to set up new defense AI units mostly at the level of the MoD, while others also established new research and development (R&D) entities to drive defense AI forward. The list of newly created defense AI units includes.

  • France’s Defense AI Coordination Unit (CCIAD);

  • Plans for a new unit focusing on developing future technologies in the R&D directorate of Israel’s MoD;

  • India’s Defense AI Council (DAIC) and Defense AI Projects Agency (DAIPA);

  • South Korean plans for a National Defense AI Center, explicitly modeled on the basis of new units established in the United States and the United Kingdom;

  • The Netherlands’ appointment of a Chief Information Officer and the establishment of the Data Science Center of Excellence;

  • Russia’s new special department on developing AI technologies in the MoD;

  • UK’s Defense AI Center (DAIC);

  • The Pentagon’s new Chief Digital and AI Office (CDAO);

  • Ukraine’s Brave1 and the Innovation Development Accelerator.

Canada, which is also listed in this group of countries, holds a special place as its current organizational setup is so dysfunctional that it almost prevents the armed forces from using AI properly. The Defense AI Center of Excellence (DAICoE) has thus been proposed as an institutional solution to overcome an excessively siloed organization. Like the other new organizations mentioned, this unit would be tasked with horizontal coordination of different stakeholders, which can create conflicts given different functional tasks and organizational interests. Thus, the jury is still out whether these new units will facilitate defense AI advancement and synchronization.

This uncertainty is also the reason why several other countries entrusted existing organizations with the task of bringing defense AI into the armed forces. Turkey, for example, has added the task of coordinating defense AI stakeholders to the portfolio of the State Secretariat for the Defense Industry (SSB). Finland rejected the idea of setting up a new unit to deal with defense AI and instead chose a cross-organizational matrix approach. Estonia created a new position inside the MoD and uses the Cyber Command as a transmission mechanism to reach out to industry and involve the military services. Australia does not foresee big organizational implications of defense AI but has rather changed its overall innovation pathway to expedite the transfer of new ideas to in-service solutions.

In addition to organizational change at ministerial levels, several countries also introduced novel elements at service and command levels. The United States launched the AI and Data Accelerator (AIDA) initiative to “embed teams of data experts within combatant commands” (Barnett 2021). France uses AI coordinators at service levels in a similar way. The Netherlands involves experts of the Data Science Center of Excellence as defense AI advisors to cooperate with specific warfare centers to formulate doctrine and set up projects. Australia has established a central organization within the Joint Capabilities Group to coordinate efforts among the services. India has adopted a similar approach with setting up a new AI Subcommittee and a Joint Working Group on AI for the services. Singapore has set up the Digital Ops-Tech Center at the new Digital and Intelligence Service for the same purpose. South Korea goes furthest with designating dedicated experimental units for each service to advance defense AI adoption.

2.4 Funding Defense AI

Sustained funding is essential for defense capability development. Defense AI is no exception to this rule. But funding has presented the biggest analytical challenges across all case studies. The significant spread in funding, that ranges from a few hundred thousand euros or dollars per annum in some countries to close to USD5bn in the United States, makes a volume-based comparison useless. In addition, national budgetary laws and diverging definitions of funding categories move cross-country analyses close to comparing apples with peaches. Therefore, the findings depicted in Table 8 need to be interpreted with care.

Table 8 How to fund defense AI
Full size table

Starting from left to right, the category “somehow financed” suggests that it is not clear, what the respective countries spend on defense AI because budgets are not disclosed. While money is available, different defense budget lines are tapped to fund the respective projects. If funding is sustainable, is hard to say in these cases. The second category “dedicated AI budget line” suggests that defense AI officially appears in the defense budget, most often in the form of R&D projects that are either listed at aggregate levels or with respect to different R&D priorities, use cases, or technology fields. Italy’s spending, for example, also includes funds to advance the networking of ecosystem partners to include specialized civilian AI research institutes. When defense AI is mentioned, like in Australia, France, and Iran, as a program of record, it signifies that procurement activities are ongoing, which also suggests multi-year funding. The final two categories in Table 8 show that the listed countries budget defense AI as part of ongoing defense R&D efforts as well as procurement projects. While this sounds like good news at first, integrating defense AI into procurement projects also implies that it is—absent access to more classified budgetary material—close to impossible to gauge the true level of defense AI spending related to the respective procurement project.

Ensuring funding, however, is only one element. Even more important is the question how funds are made available. This question poses challenges as information is sparse. However, the case studies on France and Germany, for example, clearly illustrate that traditional funding mechanisms set up for hardware development might reach their limit when considering software-enabled AI. In France, traditional hardware-focused procurement projects have separated R&D from equipment budgets. And Germany, apart from basic defense research funding, is integrating defense R&D into defense procurement programs without properly disclosing specific R&D amounts. South Korea, by contrast, is the only country in this volume that has engaged civilian ministries, primarily the Ministry of Science and Information and Communications Technology, in co-funding the development of defense AI.

The prevailing opaqueness of defense AI funding is a problem as it hampers informed debate on what is being spent and how effectively money has been invested. Overcoming the situation is challenging, as an internationally accepted taxonomy on defense AI spending is missing. In view of developing such a taxonomy, it will be important to combine input and output/outcome perspectives to assess what has been made available and what has been achieved. In this regards, three aspects are key. First, a spending taxonomy could create transparency regarding the input factors needed to produce defense AI solutions such as hardware (e.g., edge computing, high-performance computing), data management and data curation, data analytics, digital models of physical assets (e.g., missiles, radars, vehicles), mathematical models, advanced simulation environments, and software development. Second, the taxonomy needs to consider the performance of AI models and approaches commensurate with the tasks to be fulfilled. The metrics for this category should vary as second wave AI, for example, focusing on classification or pattern recognition, would need benchmarks like accuracy or false positive rates, whereas third wave AI, that can be used to develop AI tactics, will need benchmarks like exploitability, defined as the ability of one player to take advantage of a suboptimal adversarial move. Finally, the taxonomy needs to look at effectiveness and focus on outcome and impact generated by defense AI. Benchmarks could ask if an AI-enhanced ground-based air defense system engaged more adversarial targets with less munition and in shorter time or what the minimum size of an AI-empowered swarm that successfully neutralizes or destroy an adversarial tank formation would need to be.

2.5 Fielding and Operating Defense AI

Armed forces are using defense AI applications, but it is challenging to gauge the true level of contemporary defense AI diffusion. The picture presented in Table 9 is markedly different from the defense AI development priorities (Table 6) discussed above.

Table 9 International defense AI fielding use cases
Full size table

Form a regional perspective, current use cases for fielded defense AI solutions seem more prevalent in EU and NATO countries, but the differences to other regions are less stark. The most illuminating regional cluster emerges from the current war between Russia and Ukraine, which reflects an almost identical tit-for-tat use of defense AI among the warring parties. Absent conceptual novelties and tactical surprises, it is difficult to see how defense AI could benefit one of the two sides when both focus on fielding defense AI to advance situational understanding, data analytics, decision and planning support, fire support, ISR, influence operations, precision effects, and uncrewed systems.

Combining defense AI with uncrewed systems and ISR is also the most prominent use case across all 25 nations followed by target detection and data analytics. Tellingly, more countries have already deployed defense AI solutions for precision effects than identified this use case as a development priority. Among other countries, France, Germany, Iran, Russia, and the United Kingdom consider developing and fielding defense AI for precision effects among their priorities, which is in line with their well-established missile portfolios.

Almost every second country currently uses defense AI for predictive maintenance and logistics as well as simulation-based training. Around one third of the countries use it to conduct cyber network operations and improve air defense. Interestingly, fewer countries use, rather than develop, defense AI to augment situational awareness and situational understanding and for C2. Battle or Combat Management Systems are closely related to both functions. This application is interesting as Denmark has placed itself in a prominent position due to the SitaWare software suite developed by Systematic, which is used by Australia, Denmark, Finland, Germany, Sweden, the United Kingdom, the US, and other countries (Systematic 2023).

Fielding defense AI solutions for border security is relevant for Greece and Turkey, Iran, and India, as well as the United Kingdom, but no country counts border security as a development priority. This could suggest that using AI for surveillance, video and image analysis or in combination with biometrics is already rather mature. By contrast, tactics development and safety drop from the list of current deployment use cases. Defense AI to counter uncrewed assets, in support of swarming, loitering munition, and mission planning are also not yet well-established fielding priorities, suggesting that defense AI for these use cases is still at a lower level of technological readiness.

While Table 9 illustrates current use cases for defense AI, it does not yet indicate how nations roll out defense AI. In this regard Table 10 offers interesting insights. Right now, most of the countries use experiments and/or single projects to field specific defense AI solutions, but very few of these projects make it into official programs of record. This suggest that there are two “valleys of death.” The first, well-known, describes the challenge of transforming ideas into market ready products; the second is more “internal” and captures the transition from R&D to procurement. The US experience is particularly enlightening as both valleys of death are very pronounced here despite launching different AI-focused programs like Maven, which uses AI for video and image analysis, and Scarlet Dragon, which builds on results from Maven for AI-enhanced targeting assistance and setting up JADC2 as a proper program of record involving defense AI applications.

Table 10 Modes of fielding defense AI solutions
Full size table

Finally, “learning by procuring” is a prominent inroad for defense AI to enter a foreign market via the defense solutions procured from a partner. As the overview illustrates, this avenue is not only relevant for countries with less advanced national defense industrial bases but also for defense industrial heavyweights like Germany, Italy, the Netherlands, and the United Kingdom. In all four countries defense AI solutions enter the national defense ecosystem via purchases from the US and/or Israel. In fact, a modern fifth generation fighter jet like Lockheed-Martin’s F-35, advanced uncrewed systems like the Reaper or air defense solutions like Arrow need to be considered as pivotal defense AI transmission mechanism.

While using these systems can solidify the advancement of defense AI, this option also comes with challenges. A first challenge stems from the inevitable crowding-out effect that defense AI embedded with these systems might produce vis-à-vis indigenous defense AI applications. The second stems from properly understanding what kind of defense AI you get when importing foreign systems. In this regard, Moshe Patel, Director of the Israel Missile Defense Organization, made a telling statement at a May 2023 event at the Center for Strategic and International Studies in Washington, DC, when he argued that Israel would be integrating its air defense algorithms “inside the Finnish command and control” (CSIS 2023). How deep, one may ask, will nations be willing to integrate “foreign algorithms” into national sovereign systems and to what extent will the buyer have a say in calibrating and adjusting these “foreign algorithms?”

2.6 Training for Defense AI

AI needs talents but competition for talents is fierce as armed forces, defense companies, and commercial industries vie for key experts.Footnote 6 Given the financial power of big tech, for example, Marie Louise Cummings already speculated years ago, that the defense community’s relative “AI illiteracy” could tilt the balance towards big private interests (Cummings 2018: 17). This, and the conviction, explicitly expressed in countries like Denmark, Estonia, and Finland, that armed forces should only operate AI systems they truly understand, has prompted many nations to step up training efforts.

Given the fact that most nations acknowledge the general-purpose character of AI and its broad impact on all facets of private and corporate life, the limitedness of existing training efforts is puzzling. Most of the countries concentrate on making sure that soldiers understand and handle AI properly, Table 11 reveals. Updating the curricula of defense academies, setting up new courses, and advancing wargaming with AI are some of the initiatives undertaken by these countries.

Table 11 Who is trained for defense AI?
Full size table

A second group looks at the workforce more broadly and includes civilian defense personnel as well. In this regard, Greece and South Korea are of particular interest. Greece has been setting up comprehensive training programs at the military academies, while in South Korea the Ministry of Defense is cooperating with the Ministry of Science and Information and Communications Technology to educate AI literate soldiers and officers. Both countries emphasize that significant training efforts also of the civilian defense personnel are needed to make sure that defense talents are well prepared should they transit over to the civilian labor market. Similar drivers are at work in Israel, where the armed forces also use AI to early identify potentially outstanding future commanders, personalize training programs, and identify soldiers likely to extend their service. Similar initiatives are also being launched in the United Kingdom.

The third group of nations includes countries that look at training the military service and defense industrial workforces. This is the case in Spain and France, where industry plays an active role in advancing corporate AI training programs also in view of better integrating small and medium-sized enterprises into the AI-relevant supply chains of leading defense players. Turkey also belongs to this group, with the YETEN project playing a special role in AI-enhanced corporate matchmaking to identify the defense company best suited to develop specific technologies indigenously, which also includes a skills aspect.

Finally, Russia is for the time being the only country reviewed in this volume that sets its focus on expanding the AI expertise of its civilian defense, defense industrial, and military services workforce. While also using AI for training purposes, Russia uses “military scientific units.” The conscripts, with whom these units are staffed, are expected to follow a military-scientific carrier that could either lead them to work for military institutes or as experts of the armed forces. Like this idea, but with a different focus, France and Israel pool local expertise via “digital reserve elements,” that can be activated in times of need.

An interesting new angle comes into play when considering the growing importance of mini- or multilateral frameworks to develop defense AI, as discussed above. In view of co-developing defense AI, the “ability to move talent between partner countries to improve the speed of technology innovation is vital” (Cohen and Nott 2023: 9) to advance the human skillset required for defense AI. The problem, however, is that defense AI experts are on short supply everywhere. This prompts the need rethink how talents can be attracted and retained. Like jointly funding sovereign technology development via the NATO Innovation Fund, a new initiative could champion the idea of establishing a defense-focused sovereign talent management regime. Most importantly, the regime would clear experts for sensitive work on defense AI and thus facilitate, like the Schengen Agreement in Europe, free movement among its members. This idea might be particularly appealing to retain and bring back reservists that have embarked on a civil career track relevant for defense AI. In addition, national defense academies could be tasked to offer dedicated programs that ensure knowledge transmission among national and international experts. Furthermore, innovation and startup hubs could be brought in to make sure that defense industrial newcomers also benefit from knowledge transfers.

3 Interpreting the Findings

Operational requirements, technological readiness, and conceptual maturity need to come together for armed forces to innovate. The above discussion underlined that the adoption of defense AI along these three trajectories is very uneven. Consequently, some of the findings produced by this collection of case studies are confirming elements of the defense innovation literature and practices, whereas others are surprising and some even tell uncomfortable truths.

3.1 Confirming Expectations

New military ideas, concepts, and technologies diffuse in a multi-stage process. As Raska (Raska 2016: 168–169) has argued, this process ranges from emulating what others have been doing, via adopting existing practices to developing novel concepts and tactics. Right now, defense AI emulation is prevalent. First, challengers and partners look at how the United States is handling the integration of AI into military concepts and capabilities. China, Russia, and Iran scrutinize the US practice in view of better understanding how to prepare against the US use of defense AI and identifying possible weak spots that might be exploited for their own use of AI on the battlefield. The same is true for US allies and partners in Europe and the Asia-Pacific region. In this case emulating US practice signals “closeness” meant to facilitate interoperability and thus also cooperation.

Second, nations well versed in handling new technologies like AI also look at others in a process of “reverse emulation.” On the one hand, these nations look at challengers to understand if their own processes are fit enough to avoid being surprised by challengers. This becomes most obvious regarding the need to create a holistic defense AI ecosystem bringing together defense end users, research institutes, established defense and new non-defense companies, as well as investors. While most nations have adopted this idea and struggle with its proper implementation given high market entry barriers for non-defense companies and startups, China’s seemingly perfect mastering of military-civil fusion creates an implicit international benchmark. This perspective, however, overlooks the difficulties even the Chinese leadership has in implementing its top-down approach to driving technology-induced defense modernization.

On the other hand, established players and ambitious newcomers look at current wars to assess how defense AI might affect warfighting and what needs to be adapted to incorporate initial lessons identified. Volunteer-led software driven novelties that originate from Ukraine are catching the eye of many observes in EU and NATO countries. This view, however, tends to ignore that Ukraine’s all-source data and information fusion would be impossible to implement in most EU and NATO countries given current privacy protection and data sharing regulations. The prominent role of Western defense AI solution providers in Ukraine should also serve as cautionary warning not to become victim of false mirror imaging that emerges from the fact that the use of Western technology by Ukrainian forces is considered a true indigenous innovation that would, in turn, be impossible at home given the broader regulatory leeway these companies enjoy in Ukraine. In addition, war zone analyses tend to overemphasize the contribution of single assets or applications thus neglecting the need for a more systemic view (Borchert et al. 2021: 37–52).

3.2 Surprises

One of the most counter-intuitive findings of this volume is that there is no real disruptor when it comes to defense AI. One reason is conceptual and stems from the fact that disruption in military terms is hard to define. Change in the use of military power, that could lead to disruptive outcomes, can result from conceptual, organizational, technological, or operational modifications and will most often only become visible in retrospect. A second reason stems from the fact that despite the rhetoric about thinking and acting out of the box, no nation seems willing to break out as it is unclear if the “first mover advantage” will incur strategic benefits that outweigh the risks. Yet the case studies also make it amply clear that there are two thresholds that—if passed in the long-run—could constitute game-changing impacts: The move from second to third-wave defense AI that emphasizes context and consequence awareness and a relaxation of the “human in the loop” principle to advance machine autonomy. But as long as no nation is crossing these thresholds, data-driven AI in combination with the “human in the loop” principle continues to be the prevailing paradigm (Fig. 1).

A second surprise stems from the fact that digitalization is a misleading indicator for a nation’s defense AI prowess. All countries considered leaders in digitally modernizing the public sector, for example, have a hard time pulling through this edge into the military domain. Estonia, considered a thought and practice leader on e-government and cybersecurity, faces the challenge of a conservative military and a clear focus on meeting current capability needs; both forces prevent the country from quickly adopting defense AI. The same is true for Israel (Adamsky 2010: 132–133) whose “organized mess” in defense AI produces results but is far from providing optimal inroads for the country to leverage the technology base it has. A similar industrial-military dysfunction is at play in Taiwan, where commercial and defense industrial players operate in their own silos thus depriving the country’s armed forces of access to commercial talent and technology. The very same problem also hampers defense AI in South Korea, where the defense industry is considered unattractive to work with. And while India might be considered an “AI talent hotbed, it is not an AI innovation one,” the case study argues. All these countries thus serve as a cautionary tale that public (and private) sector digitalization does not easily transfer into the armed forces to create added value. This finding is even more relevant as most of the countries analyzed in this volume consider defense digitalization a prerequisite for the successful use of defense AI.

Combining this finding with the above discussion about the struggle to build adequate ecosystems leads to a third surprising insight: Irrespective of the industrial level of maturity, defense industrial innovation is the often-neglected sibling of defense and military innovation. Novel modes of defense industrial cooperation including big tech companies or smaller startups from the commercial world do not happen overnight. Rather, there is a growing need for dedicated defense industrial transformation management that very few governments have on their agenda.

South Korea, for example, is a textbook example of synchronizing the activities of many different ministries to advance defense capabilities. But interagency jointness meets a completely bifurcated industrial ecosystem. In Germany and Spain, industrial bifurcation is in full swing as well, but without a common understanding of several ministries to join forces in augmenting defense. And while France invests a lot of effort in designing a sovereign ecosystem for defense AI, the country’s leading defense companies have embarked on very different digital modernization journeys thus rendering synchronization with non-defense companies challenging. This is also the case in the United States, despite the efforts of the Department of Defense to set up defense innovation units with outreach offices in the country’s technology hotspots. In slight contrast, the case studies on Russia and Turkey suggest that both countries explore using AI to advance indigenous defense industrial capacities. Although it is unclear, yet, if the respective initiatives hold up to the announced promises, Moscow’s approach deserves attention given the country’s “continually improving adaptation cycle that links battlefield lessons to Russia’s industry and strategies” (Ryan 2024).

While this volume focuses on defense AI, the ecosystem challenge is also relevant for successfully bringing other emerging technologies into the defense sector. Defense industrial innovation and defense ecosystem management should thus be considered strategic tasks that require more government attention, because governments set the overall regulatory framework and incentives. Industrial innovation also needs more corporate efforts, because ecosystem design is all about appropriately readjusting corporate supply chains considering the need to sustain defense industrial capacities for periods of long and protracted warfighting.

3.3 Uncomfortable Truths

Some of this volume’s findings do not sit easily with prevailing assumptions underpinning the current discourse on defense AI ethics. First, defense AI ethics matters, but is more pronounced in some countries than in others and shaped by very different motives. France and the United Kingdom have established special MoD committees to oversee development of reliable AI for defense. Russia and China do see the need for defense AI regulation at the global level, but primarily as an instrument to contain the strategic leeway of the United States and to avoid US leapfrogging that would prevent both nations from overcoming existing gaps. The United States, in turn, might have a similar interest in using international regulation to prescribe a certain use of defense AI that does not create surprises on the battlefield. Spain, in contrast, has mainly focused its efforts during the EU Presidency in the second half of 2023 on advancing AI regulation irrespective of the country’s defense needs, which prompted the defense community to develop its thinking via the military channels in the EU and NATO. Thus, analysts need to put more light on the motives that prompt countries to engage on defense AI ethics at all. India, for example, sends very mixed signals when arguing in favor of the responsible use of defense AI but abstaining from signing, for example, the 2023 REAIM Summit’s Call for Action on exactly this principle. The Netherlands have established a national ELSA Lab Defense to assess the ethical, legal, and societal consequences of defense AI also in view of leveraging this approach to co-shape international norm-building. Singapore follows a similar understanding and considers military AI governance an important element of its defense diplomacy outreach, which led the city-state to publish its own guiding principles for defense AI in 2019. While consenting with the need for responsible defense AI, Estonia and Finland, for example, are concerned that an overemphasis on regulation might hamper ethically justified technology development and thus also business interests. And in Greece decades of underinvestment and strategic rivalry with Turkey have pushed the defense AI ethics discourse to the backburner.

Second, the case study on Ukraine—as well as current developments in Israel—illustrate that war readjusts normative preferences. Under threat, both nations recalibrate the rules of engagement for AI on the battlefield. When “the emphasis is on damage and not on accuracy,” as a spokesperson of the Israel Defense Forces said on 9 October 2023 (Abraham 2023), the threshold for the use of AI in military operations is lowered—by the “human in the loop,” and not by technology. Adversarial electromagnetic spectrum dominance, that makes it more difficult to provide connectivity to remotely operate assets using AI, can render fail-safe provisions to keep AI under human control more challenging, as the Ukrainian case study shows. Therefore, both case studies serve as powerful reminders that norm preferences are context-driven since war can entice governments to “embrace once-controversial technologies with gusto” (O’Brien 2024). An in-depth discussion of this topic goes beyond the scope of this volume, but future research on defense AI ethics could provide added value by analyzing, if and how norm adjustments, that occurred under war, “survive” the transition to peace, how war-torn societies use their own experience in shaping international norm discussions, and what role other factors like Ukraine’s new status as an EU admission candidate play in this regard.

Finally, the findings of this volume cautiously warn against making non-democratic nations nine feet tall when it comes to implementing defense AI. Rather, the China, Iran, and Russia chapters show, that these countries suffer from the same pathologies that hinder defense innovation in democratic nations. The idea that individuals or leading party-affiliated groups have a completely free hand in autopiloting their defense establishment towards AI-enhanced military superiority is a caricature of reality. Yes, civil-military relations differ from a democratic and rules-based approach prevalent in democratic nations. This, however, does not yet suggest that military-civil fusion is easier to achieve as it depends if and to what extent novel ideas and technologies of non-military origin can penetrate the military industrial complex, which is an important power player in these countries, to yield military advantages (Evron and Bitzinger 2023; Scharre 2023: 21). Second, non-democratic governments may have a different risk calculus, but most of them do not gamble regime survival for novel, but immature ideas, concepts and, technologies. This illustrates that strategic competition can produce lookalikes, if challengers believe that “the perceived risks of failing to imitate another state outweigh the perceived benefits of pursuing a novel but risky new technology” (Liou et al. 2015: 159).

At the same time, however, there is reason to stay vigilant about how defense AI will evolve in these three countries. Two aspects deserve special attention. First, more research is needed on emulation among non-democracies that have learned to cope with sanctions targeting their economies, strategic industries, and critical technologies. China, Iran, and Russia are convinced that the global influence of the United States is waning. And this might prompt them to test the resolve of Washington and its allies also by using defense AI. In this regard the use of defense AI in the context of the respective nuclear arsenals and to exert domestic control are two important topics to monitor closely.

Second, Russia is a well-established defense exporter, China is ramping up defense exports into the Greater Middle East, Africa, and Latin America, and Iran is maintaining a pan-regional network of proxy forces. While it is too early to tell, to what extent all three nations are willing to export defense AI and engage in knowledge and technology transfer with recipients, these development vectors also need more attention. On the one hand, AI could shift the intra-power balance between the three since Iranian military capabilities empowered with AI could directly threaten Russia and endanger Chinese interests in the Middle East. On the other hand—and building on the often-overlooked aspect of innovation driven by violent non-state actors (Veilleux-Lepage and Archambault 2022)—countries like Iran could use proxy forces like Hezbollah or Hamas as “battle lab assistants” that evaluate and test the benefits of new AI-related concepts and technologies in different theaters of operation before fielding them on their own.

4 Conclusion

Defense AI is advancing across the 25 countries analyzed in this volume, but motives, drivers, pace, and priorities differ. The paradigm that is underpinning defense AI, however, is surprisingly stable irrespective of a country’s overall strategic ambition, its technological and industrial maturity, or the character of its domestic political system: data and human-centric defense AI describe today’s dominating focus and understanding of defense AI. This prompts two final remarks related to the prevailing imaginaries about defense AI.

First, what do we (believe to) see when we talk about defense AI? This question addresses the boundaries of the prevalent socio-technical defense AI imaginaries (Jasanoff and Kim 2015). Currently, the dominant body of literature, briefly discussed at the beginning of this chapter, is fixated on data-centricity and thus second wave AI. Current analyses predominantly ascribe limitations and opportunities to defense AI that are true for traditional machine learning, but do not apply in the same way for third wave AI. Understanding the limits of this analytical perspective is important as it can lead to inadequate assumptions about the overall impact of defense AI. The often-discussed idea of so-called “flash wars” (Scharre 2018), analogous to stock market flash crashes caused by technology glitches, builds on the erroneous idea that AI will unstoppably pursue its path in a certain direction creating an escalatory dynamic. Third wave AI, by contrast, knows when “going out of action” due to confusing signals from the battlefield would be needed or when de-escalation would be more beneficial than aggressively staying the course (Hofstetter 2014).

Therefore, the rise of third wave AI has significant consequences for the regulation of defense AI, as the respective technologies are subject to fierce geoeconomic competition among nations aiming at containing raising challengers to maintain their own edge. In view of containment, third wave AI is ambivalent. The fact that it is less data hungry will make it less prone to data-induced restrictions. But third wave AI still requires significant computing power opening the door for regulation to target this angle to limit adversarial capacities. This, however, is easier said than done as sophisticated mathematical effort goes into applying methods that reduce third wave AI’s need for computing power. In sum, as mathematical modelling and theoretical as well as conceptual sophistication are more important for third wave defense AI, the diffusion of these solutions is very likely harder to control and contain but will likely also occur at slower pace given more demanding requirements.

This is, perhaps, also the reason why third wave AI remains still difficult to understand and requires a more nuanced vocabulary to describe its outcomes. Eric Lipton’s August 2023 story about Valkyrie, an experimental pilotless aircraft of the US Air Force using AI, serves as a telling example. He wrote:

One of the things Major Elder watches for is any discrepancies between simulations run by computer before the flight and the actions by the drone when it is actually in the air (…) or even more worrisome, any sign of ‘emergent behavior,’ where the robot drone is acting in a potentially harmful way. (Lipton 2023, emphasis added)

Emergence, in the context of this quotation, meant that the AI-enhanced Valkyrie went into a series of rolls to make optimal use of the infrared sensors on board. Human operators were not expecting AI to perform these moves, but these moves yielded better performance. So, the pure fact that emergence might imply tactical behavior that human operators did not expect does not make this kind of behavior “potentially harmful.” Or, to put it differently: Equating emergence with wrongful behavior will deprive armed forces of the added value third wave AI can deliver, for example, by creating surprise, a core principle of war, in a way “differently than a human” would act (Demarest 2024). The more important question is if the system, that is producing AI-driven emergent behavior, is producing it in an explainable and verifiable way as this will be needed for third wave AI solutions to be certified for military use.

Second, when do we know that AI has been used on the battlefield? This question is anything but easy to answer because AI, as software, mostly eludes our sensory perception. In fact, to be effective, AI needs to be embedded in sensors, missiles, platforms, decision-making procedures or other assets and technologies. Thus, the true impact of AI can only be evaluated in tandem with them—and it is this dependence that also shapes the performance of AI.

This, in turn, reinforces the need to be much more precise about the ultimate goals defense AI is expected to accomplish, the roadmap needed to deliver AI-induced capability growth, and the metric needed for performance measurement. Going back to the principles of war, defense planners need to consider what benchmarks they use, for example, to assess the contribution of AI in creating surprise, advancing flexibility, or enabling the initiative. In so doing, it becomes obvious that asking which nation is leading is popular, but quite misleading. There is no aggregate benchmark that would capture the state of transformation regarding conceptual, organizational, technological, and operational maturity in adopting defense AI. Rather each country’s progress needs to be analyzed in view of its own level of ambition, the contextual challenges it faces, the sophistication of it defense-industrial technology base, and the savviness of its armed forces. Consequently, countries that look for inspiration when exploring defense AI can choose from a continually growing number of role models. Whether this accelerates instability or reinforces stability very much depends, to paraphrase Alexander Wendt (Wendt 1992), on what military users of AI make of it.