Abstract
This article analyses military AI governance using the Advocacy Coalition Framework (ACF), focusing on two key advocacy coalitions: the ‘tight regulation’ coalition which emerged in the context of the UN’s Group of Governmental Experts on Lethal Autonomous Weapons Systems, and ‘responsible use’ coalition which emerged in the context of the REAIM Summits and the US-led Political Declaration on Responsible Military AI Use. The ‘tight regulation’ coalition advocates for a legally binding treaty to restrict certain AI technologies, while the ‘responsible use’ coalition emphasises voluntary codes of responsible conduct. Through the ACF lens, the study examines policy learning, coalition coordination, and evolving dynamics within this emerging policy domain.
Introduction
The military applications of artificial intelligence (AI) are no longer speculative glimpses of future warfare but an integral component of modern conflict. While AI-enabled systems enhance battlefield capabilities, they also intensify ethical, legal, and strategic dilemmas. Efforts to establish global norms remain fragmented (UNIDIR, 2024). Addressing these governance challenges requires an analytical lens capable of capturing the diverse actors and evolving normative divides that shape this emerging domain.
Traditional approaches to international security (realism, liberalism, and constructivism) offer insights into state behaviour, institutional design, and norm development (e.g. Choi, 2015; Finnemore and Wendt, 2024). Yet they tend to privilege structural conditions or broad ideational shifts over the dynamic processes through which states, technology firms, and civil-society actors mobilise resources, advance competing claims, and recalibrate strategies. Realism focuses on power politics and state interests but lacks a mechanism for explaining how domestic and transnational actors engage in long-term regulatory struggles (Balfour, 2023). Liberal institutionalism emphasises cooperation through formal institutions but struggles to capture fragmented and iterative governance when technological change outpaces legal frameworks (Johnson and Heiss, 2018). Constructivism highlights the importance of ideas and norms but tends to offer retrospective explanations rather than specifying how contested security policies evolve. Absent a framework that systematically captures policy coalitions, belief-driven contestation, and long-term adaptation (Finnemore and Wendt, 2024; Jung, 2019), conventional perspectives provide only a partial understanding of how global security rules emerge and evolve amid accelerating technological change.
We therefore turn to the Advocacy Coalition Framework (ACF) as a theoretical lens attuned to the complexities of global military AI governance and applicable to other transnational security issues. Originally developed to explain policy change in domestic settings, ACF has increasingly informed foreign policy analysis (Brummer, 2024; Haar and Pierce, 2021), yet its application to international security remains limited. ACF examines how otherwise disparate actors coalesce into durable coalitions anchored in shared belief systems, encompassing core values, causal assumptions, and long-term policy goals, and how these coalitions mobilise expertise, select or abandon negotiating venues, and adapt to shocks and shifts in normative traction. It provides a tool to understand dynamic processes of coalitions, policy learning and adaptation (see also introduction to this Symposium, Lo et al., 2026).
Applying ACF to military AI governance offers a way to analyse how regulatory preferences evolve through changes in coalition coordination, external shocks, and long-term policy-oriented learning. The framework clarifies why governance efforts have followed distinct trajectories and identifies the conditions under which rival coalitions align around shared regulatory principles despite ideological divides. The governance of military AI, we argue, arises from not solely top-down strategic calculation but an ongoing contest between advocacy coalitions with competing visions of security, technological development, and legal oversight.
Extending ACF beyond its conventional applications strengthens both international security studies and policy process research in three ways. First, it offers a theoretical perspective for analysing governance challenges that transcend state-centric models, highlighting the contested nature of security policymaking. Second, it maps the key actors, ideological commitments, and strategic interactions that shape military AI regulation, underscoring the influence of middle powers and non-state actors. Third, it explains how policy-oriented learning and adaptation occur in security governance, illuminating both incremental and transformative shifts in regulatory practice. As debates over military AI continue to unfold, ACF provides a theoretically and empirically grounded framework for understanding the forces shaping its trajectory.
The remainder of the article is structured into two main parts. The first applies ACF to international security, outlining its core concepts and explanatory value for understanding belief-driven coalition dynamics. The second employs this framework to examine global military AI governance through a qualitative analysis of two competing coalitions; we draw on a set of primary and secondary sources. We conducted online anonymised interviews with diplomats, policymakers, and experts, selected for their affiliation with principal actors or recognised expertise in the field, that provide insights into coalition beliefs, strategies, and interactions. The primary evidence is complemented by secondary materials, including academic literature and expert commentary. Using process tracing, we examine key developments over time, identifying shifts in coalition dynamics, policy learning, and institutional change.
Applying ACF to international security
ACF provides a multilevel lens for understanding international security policymaking in an era of contested governance, technological disruption, and evolving threats. At its core, ACF explains policy dynamics through the interaction of belief systems, coalitions, and learning (see also Lo et al., 2026). Its central unit of analysis is the advocacy coalition: networks of public and private actors that share core beliefs about policy problems, coordinate strategies within a policy subsystem, and pursue influence. The subsystem represents the functional arena in which actors concerned with a specific policy issue interact over extended periods, competing and learning as they seek to shape collective outcomes.
ACF distinguishes among deep core (fundamental normative values), policy core (strategic preferences within a policy domain), and secondary (instrumental implementation choices) beliefs. Coalitions exert influence across domestic, regional, and international arenas through coordination, brokerage, and strategic venue shopping. Policy change arises from external shocks that disrupt the balance of power between coalitions and/or policy-oriented learning through which actors revise their beliefs and strategies (Nohrstedt et al., 2023; Nowlin, 2024). Because dominant coalitions tend to resist change, major shifts are infrequent, emerging when sustained external pressures or accumulated policy failures erode existing alignments. This architecture clarifies how belief-driven contention, resource mobilisation, and institutional adaptation interact across levels, linking micro-level learning to system-level change and illustrating how actors, institutions, and ideas jointly shape policy change.
ACF shares the broadest conceptual overlap with alliance theory and practice theory, two approaches rooted in the major traditions of international security. Alliance theory explains why states form, sustain, or dissolve security partnerships, focusing on power dynamics, threat perceptions, and alliance management (Topal, 2024; Walt, 1987). Traditionally viewed as an instrument for balancing adversaries or maintaining bloc cohesion (Masala, 2010), it remains largely state centric, overlooking how non-state actors and transnational coalitions shape policy evolution. Practice theory, in contrast, examines how norms are enacted and stabilised through everyday interactions and professional routines (Bode, 2023). It illuminates micro-processes of meaning reproduction but offers limited analytical leverage for explaining how such practices scale up, gain institutional authority, or generate policy change over time.
What remains absent is a framework that explains how belief-driven coalitions operate across levels of governance and over time, linking micro-level learning to macro-level transformation. ACF fills this gap through an integrative causal model that connects belief systems, resources, venues, and learning, thereby linking actor preferences to institutional change across arenas and time horizons and offering a set of interrelated analytical features that clarify its explanatory value.
First, it places coalitions at the centre of analysis, tracing how they mobilise expertise, authority, and public narratives, forge alliances across sectors, navigate institutional constraints, and adapt strategies in response to shocks and feedback. Second, it captures both incremental learning and punctuated breakthroughs, reconciling continuity and transformation within a single theoretical logic and clarifying why policy change usually unfolds gradually yet occasionally accelerates into major reform. Third, its multilevel perspective integrates domestic, regional, and international arenas, revealing how belief-driven coordination, resource asymmetries, and institutional opportunity structures shape coalition influence across governance levels.
Fourth, it explains why regulatory outcomes vary even within the same institutional forum, showing how shifting coalition configurations, evolving resource distributions, and cumulative learning generate differentiated results under otherwise stable conditions. ACF also recognises informal venues – policy declarations, industry standards, and multistakeholder dialogues – as legitimate arenas for normative development and strategic recalibration. Finally, coalition behaviour rests on three interrelated elements: policy-oriented learning (ACF’s core process), resource mobilisation, and venue shopping, each operating within institutional opportunity structures, such as decision rules and mandates (Gupta et al., 2025). Together, these elements link belief-driven contestation to observable patterns of continuity and change, clarifying how advocacy struggles evolve and why regulatory outcomes differ across governance settings.
ACF provides a synthetic, unifying framework that mitigates the fragmentation of existing approaches and integrates material, institutional, and ideational dimensions into a coherent account of how complex security policies evolve in an era of contested governance. Military AI governance exemplifies such complexity: belief-driven coalitions operate across specific venues – the terrain the next section maps.
Applying ACF to governance of military AI
We understand military applications of artificial intelligence (‘military AI’) as a policy issue, and its governance functions as the corresponding policy subsystem within international security. Its boundaries are inherently fuzzy (Nohrstedt et al., 2023), yet we delineate them around global efforts to govern military AI, which together define its geographical and thematic scope (Henry et al., 2022). As a field of global security governance, this subsystem remains ‘nascent’ (Nohrstedt et al., 2023), with norms, venues, and coalitions only beginning to crystallise.
Debates on military AI governance have centred primarily on lethal autonomous weapon systems (LAWS). Although AI now permeates every level of military activity, regulatory attention remains concentrated on these, leaving broader governance challenges underexamined. The governance of LAWS, a pivotal component of military AI, has unfolded mainly within the UN framework. Since 2014, the Group of Governmental Experts on Lethal Autonomous Weapons Systems under the Convention on Certain Conventional Weapons (CCW) has met regularly, serving as a forum for examining the ethical, legal, operational, and technical dimensions of autonomy in weapon systems. Within it, a ‘tight-regulation’ advocacy coalition has coalesced around humanitarian-minded middle powers, such as Austria, Mexico, and Brazil, working closely with civil-society networks, including the Campaign to Stop Killer Robots and International Committee of the Red Cross (ICRC; see Bode and Qiao-Franco, 2024).
By the late 2010s, several states, notably the United States, the United Kingdom, France, and Japan, had begun releasing national strategies outlining the development, deployment, and use of military AI. Yet, given the rapid proliferation of AI technologies, national regulatory efforts alone appeared insufficient to address the imperative of global governance, which is essential for averting arms races and mitigating the attendant risks to international stability.
In pursuit of these objectives, a second advocacy coalition (‘responsible-use’) emerged, led by the Netherlands, South Korea, and the United States, coalescing around two complementary initiatives. The first was the Summit on the Responsible Artificial Intelligence in the Military Domain (REAIM), launched jointly by the Netherlands and South Korea in February 2023. It seeks to initiate and facilitate dialogue on the responsible development, deployment, and use of military AI (Interview with diplomat, 2024). The final declarations of the first (February 2023) and second (September 2024) summits were endorsed by 57 and 63 countries, respectively (Government of the Netherlands, 2023; RoK Ministry of Foreign Affairs, 2024).
REAIM summits adopt a broad, multistakeholder format, engaging industry representatives, civil-society actors, and experts. Because industry is central to developing military AI, REAIM organisers view its participation as essential to achieving meaningful norms of responsible use (Interview with diplomat, 2024). To further institutionalise these efforts, and consistent with the ACF’s expectation that advocacy coalitions advance their goals through collaboration with non-state actors, the Netherlands supported creating the Global Commission on Responsible AI in the Military Domain. It comprises individual commissioners (former diplomats and experts) and an advisory group facilitated by the Hague Centre for Strategic Studies, a non-state actor.
The United States launched its own initiative in February 2023 during the REAIM Summit in The Hague, aimed at promoting the responsible use of military AI. The Political Declaration on the Responsible Military Use of Artificial Intelligence and Autonomy (Political Declaration) is intended to foster developing robust international norms of responsible behaviour. Unlike REAIM, it is open exclusively to states. Fifty-eight countries have endorsed its 10 foundational measures outlining principles of responsible military AI governance.
The tight-regulation advocacy coalition within CCW and responsible-use advocacy coalition of the Political Declaration constitute the two principals shaping the governance of military AI (Table 1). For both, global stability and the reinforcement of international law are fundamental deep core beliefs. However, their policy core and secondary beliefs diverge significantly.
Summary of advocacy coalitions within the military AI governance subsystem.
With respect to policy core beliefs, the tight-regulation advocacy coalition centres its discussions on prohibiting LAWS, extending this stance to certain military AI applications. The New Agenda for Peace and December 2023 UN General Assembly resolution (A/RES/78/241) have tempered these calls for a total ban, instead advocating restrictions limited to weapon systems lacking meaningful human control. In contrast, the responsible-use advocacy coalition promotes an alternative vision, emphasising responsible development and deployment without seeking to prohibit specific categories of military technology.
Another differentiation between the two coalitions is their secondary beliefs. The CCW coalition advocates creating a legally binding instrument. The responsible-use coalition promotes adopting voluntary codes of conduct (Interview with diplomat, 2023) or, in the words of US officials, an ‘international framework of responsibility’ (US Department of State, 2024).
Both coalitions consist of several sovereign states as their common members, as states are members of CCW. Both include several states as overlapping members, given that states participate in both the CCW and REAIM processes. Non-state actors, including private individuals and non-governmental organisations (NGOs), also engage with both coalitions. The ICRC, while closely aligned with the CCW coalition, participates in REAIM discussions and other multistakeholder forums alongside broader civil-society networks. Industry actors, by contrast, are represented primarily within the responsible-use advocacy coalition.
Policy brokers, actors who operate between multiple advocacy coalitions, may facilitate communication, negotiation, and compromise (Nohrstedt et al., 2023). For military AI governance, Austria might cautiously fulfil this role. Vienna has advocated for a legally binding instrument for LAWS at the UN level (Interview with diplomat, 2023). It also not only endorsed the Political Declaration but also spearheaded one of its specialised working groups focusing on oversight (Freedberg Jr., 2024). However, Austria’s firm stance, as highlighted during its own global conference ‘Humanity at the Crossroads: Autonomous Weapons Systems and the Challenge of Regulation’ in Vienna in April 2024 (Austrian Federal Ministry for European and International Affairs, 2024), indicates that its role in policy brokerage may clash with the evolving priorities of states within the ‘responsible-use’ coalition.
The success of a coalition, defined by which one emerges as dominant, depends on the coordinated behaviour of policy actors. The decade-long tight-regulation deliberations within CCW, which have yielded virtually no tangible results, indicate a pattern of weak coordination (Nohrstedt et al., 2023). CCW’s consensus rule, coupled with rivalries and tensions among great powers, has hindered agreement and the implementation of concrete measures. By contrast, the responsible-use coalition appears to represent a more like-minded grouping, where states willingly endorse a common set of principles. Though relatively young and operating in a politically sensitive domain, it benefits from strong coordination (Nohrstedt et al., 2023) that enhances its collective influence.
Within the ACF, a coalition’s capacity for learning, and its resulting contributions to policy change or stability, is a key determinant of its dominance. The framework also emphasises the need for long-term time frames (e.g. a decade or more) to capture the gradual evolution of policy processes (Nohrstedt et al., 2023). Although the governance of military AI remains in its early stages, the protracted effort to regulate LAWS through CCW offers valuable insights for both coalitions. The CCW coalition has moderated its position on the scope of governance, shifting from a total to a selective ban – an adjustment in its secondary beliefs. The Political Declaration follows a more pragmatic path, promoting codes of conduct among like-minded states to avoid repeating CCW’s deadlock. This change among humanitarian-oriented states and NGOs within CCW, to supporting a more targeted approach, constitutes a major policy shift (Henry et al., 2022). The resulting policy evolution created favourable conditions for the Political Declaration process to consolidate an agenda-setting position, evidenced by growing state endorsements and working-group activity, while formal rule change remains nascent.
Nonetheless, under the ACF, an internal or external shock may displace a dominant coalition and elevate a minority coalition to prominence. These include a breakthrough in deploying military AI during a large-scale conflict among great powers or US withdrawal from the Political Declaration under a future administration reprioritising unilateralism. The latter scenario could arise from a combination of domestic and international factors, such as efforts to repudiate the previous administration’s legacy and reassert US military primacy (Associated Press, 2025).
Absent such a disruption, the responsible-use coalition, supported by Western states, industry, and key expert platforms, is likely to retain its agenda-setting advantage. However, if either form of disruption were to occur, current progress could stall, attention might shift back to CCW’s agenda, or alternative pathways, such as the 2024 UN General Assembly resolution on military AI (adopted by 159 votes), could gain renewed momentum. In such moments, advocacy coalitions typically revise their secondary beliefs while preserving their deep core commitments – a hallmark of the ACF’s policy-oriented learning dynamic.
Conclusion
The two advocacy coalitions highlight three dynamics that established IR frameworks struggle to capture. First, issue-specific resource alignments, rather than institutional rules alone, account for within-forum divergence, as illustrated by CCW’s ability to update Protocol V on explosive remnants while remaining gridlocked on autonomous weapons. Second, ACF’s notion of venue shopping clarifies how the responsible-use coalition behind the Political Declaration circumvented great-power tensions through an opt-in endorsement mechanism – which neither realism’s power-centric logic nor institutionalism’s rule typologies can predict. Third, ACF shows how compound shocks generate incremental adjustments in regulatory language without necessarily displacing the dominant coalition, a pattern that constructivist norm-cycle models blur and practice theory leaves empirically underdeveloped.
Beyond military AI, ACF provides a coherent framework for analysing transnational security domains where public authority is fragmented, private capabilities are pivotal, and technological change outpaces treaty-making. Whether the issue concerns offensive cyber operations, counter-space weapons, drone swarms, or dual-use biotechnology, policy contests increasingly involve heterogeneous coalitions of states, platform firms, standards bodies, humanitarian NGOs, and epistemic communities. ACF’s focus on coalition dynamics, multidimensional resource mobilisation, and cross-venue manoeuvring fits these arenas particularly well. It helps explain, for example, how a coalition of cloud providers and middle powers might leverage ISO standards to bypass a Security Council stalemate on cyber norms or successive ransomware crises operate as compound shocks that gradually recalibrate policy core beliefs about offensive cyber deterrence. Research could extend this approach to the emerging counter-space regime, mapping how commercial launch firms and small-satellite operators are reshaping great-power bargaining.
ACF’s emphasis on policy-oriented learning over extended time horizons also provides a practical diagnostic tool. When coalitions converge on secondary beliefs yet retain the same venue strategy, a shift in rule-making arenas may be imminent. If deep core disagreements persist, incremental drift rather than major breakthroughs is more likely. Applying ACF beyond military AI will require the calibrations identified in this study – attention to strategic imperatives, compound shocks, and boundary-spanning coalitions – but the framework’s integrative logic remains intact. In short, ACF complements rather than replaces established IR theories, while offering the causal precision increasingly needed to understand and govern fast-moving, multi-actor security domains – military AI governance chief among them.
Footnotes
Acknowledgements
The authors thank Nicholas Thomas and Hengyi Yang for their helpful comments on the first draft. Tasha Bigelow proofread the text with much care. The authors are also grateful to the reviewers for their constructive comments and to the journal editors for their continued support.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the REMIT research project, funded from the European Union’s Horizon Europe research and innovation programme under grant agreement no. 101094228.
