The military adoption of AI technologies and avenues toward their potential governance have gained significant salience throughout the early 2020s. Initially, the focus of the global discussion was exclusively on autonomous weapon systems (AWS) that “select and engage to one or more targets without human intervention”. Over time, states and other stakeholders involved in the international debate have increasingly recognised AI in the military domain as a multi-faceted issue that extends beyond weapon systems to encompass how military decision-making, including around the use of force, is structured at a more fundamental level. This renewed attention is, not least, due to a combination of reported advances around generative AI models and the simultaneous growing use of diverse AI capabilities in various conflicts and wars such as those in Syria, Libya, Nagorno-Karabakh, Ukraine, Gaza, and Iran.
The six years since the AutoNorms project started in August 2020 have seen AI in the military domain move from speculation to reality. These developments give a strange vindication to the very premise of AutoNorms which, at the time of writing the application in the summer of 2018, was disputed. I still remember one piece of feedback inviting me to re-think the project’s focus on AWS because “such systems do not exist yet”. Such sentiments are deeply misguided now and were no less misguided in 2018, illustrating a stubborn misconception that has persisted despite years of evidence to the contrary. As those studying this empirical domain closely know, autonomy in weapon systems does not mark a complete departure, nor should we believe the ever-recurring news headlines proclaiming something as the “first AI war”.
Militaries have long been using weapon systems with predecessor technologies such as automation in their targeting functions. We only need to think about guided missiles or air defence systems as examples of this. But we have seen AI and autonomous technologies growing in complexity and spreading over the past decade, making them a firm part of operational military reality and therefore highly salient. There appear to be multiple reasons for this development, including a situation of greater geopolitical uncertainty combined with a growing belief in technological “solutions”.
I designed the AutoNorms project with one clear ambition in mind: to examine how autonomous weapon systems change international norms on the use of force. Realizing this ambition required two essential, mutually constitutive components: first, it required building novel analytical arguments around the project’s core premise that practices make norms. I will speak to the argument in greater detail in the subsequent segment of the essay. Second, addressing AutoNorms’ main research question required building a nuanced, in-depth empirical basis of the degrees to which AI has become integrated into the weapon systems of prominent states, sketching a landscape of diverse practices associated with these states, as well as closely monitoring the evolving global regulatory debates.
To develop this empirical basis, the AutoNorms team combined various qualitative methods including participant observation, informal conversations at the margins of international governance initiatives, participation in closed-door expert meetings, visual analysis, surveys, and open-source analysis of available documentation around technical and military practices. Secrecy affected the available material in uneven ways—generally, there was significantly more material available about particular systems used and developed in the United States (US) as compared to China and Russia. Building knowledge of the empirical landscape involved assessing the veracity of claims around technological capabilities through sustained triangulation and immersion into expert communities. Both the analytical and empirical components of AutoNorms have been essential for the project’s work, which has proceeded throughout in an abductive fashion.
Looking back at the past six years, my main take-away from the AutoNorms project has been the project’s core analytical idea that practices make norms. I will unpack this argument in greater detail and reflect on how it offers new insights both for the particular empirical context of AI in weapon systems and beyond. In other words, the argument that practices make norms travels much further than this specific empirical domain.
1. Practices Make Norms
What does it mean to say that practices norms? I typically unfold this argument by first clarifying that, for the AutoNorms project, norms are not something that is purely legal or that is only enshrined in legal text. Norms are also social. Social norms are understandings of appropriateness that are often implicit, not necessarily written down, and typically not publicly discussed. Practices of developing, of training personnel for, and of using weapon systems integrating AI shape these norms. Thought in this way, social norms are something that emerges from and circulates within these practices that are, quite often, performed outside of the public, deliberative space. Such social norms shape and communicate what states consider “appropriate” behaviour when it comes to AWS and AI in the military domain more broadly.
The vast significance of operational, hidden practices for norm emergence and development had not been recognised prior to the AutoNorms project. In this, “practices make norms” is a departure from how International Relations (IR) scholars have understood where new norms come from. Typically, scholars would look at deliberative forums, such as the United Nations, where states articulate and shape their policy positions. And then potentially, there is something written at the end of this process that then shapes state behavior. This is the expected turn of events around norm emergence that we have seen in historical humanitarian disarmament processes, for example with regard to landmines. There was a landmine convention in the late 1990s which entered into force, and then a majority of states abandoned the use of anti-personnel landmines.
But what this perspective ignores is how in the case of most technological innovations in warfare, states may have been using these weapons for a long time. Militaries around the world have integrated automation and autonomy, which can be considered predecessors of AI, into the targeting functions of weapon systems since at least the late 1960s. The AutoNorms project has tracked this process of slowly developing and spreading historical practices. Of course, this process of tacitly shaping norms does not happen in a normative vacuum. States perform such practices within a normative structure that sets certain boundaries for how they can design and use novel weapons, for example international humanitarian law. But this structure sets general expectations for behaviour rather than specific regulations and there is often a lot of uncertainty around how precisely these general expectations apply to certain AI techniques in the military domain.
Beyond this practice-based process, there is also a public-deliberative process. In the case of AWS, this process only started when the issue entered the international community’s agenda in the mid-2010s and became the subject of a more formal debate at the Group of Governmental Experts on Lethal Autonomous Weapons Systems (GGE on LAWS) since 2017. Since then, the two processes run in parallel to each other, but the practice-based process precedes the public-deliberative process.
The AutoNorms project has also studied how these two processes interact over time. I have differentiated between negative and affirmative ways in which the public-deliberative process of norm emergence may acknowledge the practice-based process. In the case of AI in weapon systems, we have for long seen two dynamics: (1) Wilful ignorance and (2) Affirmative acknowledgement. Wilful ignorance describes a dynamic of interaction where the public-deliberative process of norm emergence does not explicitly acknowledge the practice-based process or perhaps wilfully ignores it, including by engaging in distancing. Existing weapon systems integrating automated and autonomous technologies, for example loitering munitions, have not been the subject of substantive discussion at the GGE on LAWS—although they are used increasingly widely.
Affirmative acknowledgement mentions the practices of developing and using existing autonomous or AI-based weapon systems but argues that these are operated in adherence to human control. Some of these existing systems, such as air defence systems, have been framed as the gold standard of human control because they have a human operator ‘in’ the loop or ‘on’ the loop of the use of force. These arguments shelf a closer discussion of existing weapon systems as they supposedly already fulfil the requirements of the emerging principle of human control and judgment. The prevailing understanding among states parties presenting these cases appears to have been that the only thing to be gained from studying existing systems integrating autonomous and AI capabilities are ‘best practices’ for how human control can be successfully exercised and implemented.
Over time, the AutoNorms project has found that the effect of these two processes is that the practice-based process of norm emergence prevails. So, what happens in the practice-based processes? The AutoNorms project has closely looked at existing weapon systems that incorporate autonomous and AI technologies. Specifically, we have studied air defence systems, loitering munitions, and AI-based decision support systems. I argue that the trajectories of weapon systems across practices of design, training, and use shape what states consider as appropriate when using force.
All three types of systems we studied closely continue to involve humans in their employment, including, for example, in the forms of humans authorizing specific attacks. But the quality of control that humans can exercise is compromised and diminished due to the complexity of the tasks they need to perform in using the systems and the demands they are placed under, for example in terms of speed and overseeing multiple, networked systems. These normative dynamics have not been intentional. Rather, they result from features and logics inherent to AI technologies, such as complexity as well as problematic assumptions about the nature of human-machine interaction.
2. Reflecting on How Practices Make Norms Beyond AI in Weapon Systems
The basic argument that practices makes norms travels far beyond the case of AI in weapon systems. It speaks to a general logic around how we can analyse the normative impact of technologies. The work of AutoNorms has demonstrated that choices about designing and using technologies are essentially deeply political choices with great societal significance. This insight has long been recognized in disciplines such as Science and Technology Studies but has only been slowly brought into the IR mainstream.
It follows that choices about designing and using technologies such as AI need to be made deliberately by decision-makers, they cannot be left to technical practices that then have unintended but highly consequential effects. Emphasizing the agency of decision-makers also serves to counter one of the fundamental hype narratives around AI, where such technologies are presented to be on an inevitable trajectory that decision-makers can only react to. This is not correct, as political decision-makers can and need to make choices about how these technologies are built and what purposes they serve.
I firmly believe, also, that we need to remain attentive to spill-over effects of how AI technologies are designed and used across various societal domains. AI in the military domain often comes across as very remote. But the choices made in this highly consequential, extreme societal domain and how they shape our understandings of appropriateness could have concrete impacts in other societal domains. As AI has become much more of an everyday topic over the past years, I think that many citizens have become more attuned to potential risks and challenges associated with these technologies, including for the exercise of human agency. In other words, what does the use of AI technologies at an increasing scale do to us as humans?
For me, the potential effects that designing and using AI in particular can have on our structures of knowledge, on our understanding of justice, and ultimately on what it means to being human are vital and very much open research questions to come out of AutoNorms.
Recommended AutoNorms outputs for further reading
- Bode, I. (2023). Practice-Based and Public-Deliberative Normativity: Retaining Human Control over the Use of Force. European Journal of International Relations 29(4), pp. 990-1016.
- Bode, I. (2024). AI Technologies and International Relations: Do We Need New Analytical Frameworks? The RUSI Journal 169(5), pp. 66-74.
- Bode, I., & Huelss, H. (2023). Constructing Expertise: The Front- and Back-Door Regulation of AI’s Military Applications in the European Union. Journal of European Public Policy 30(7), pp. 1230-54.
- Bode, I., & Huelss, H. (2024). Artificial Intelligence Technologies and Practical Normativity/Normality: Investigating Practices beyond the Public Space. Open Research Europe.
- Bode, I., Huelss, H., Nadibaidze A., Qiao-Franco, G., & Watts, T.F.A. (2024). Algorithmic Warfare: Taking Stock of a Research Programme. Global Society 38(1), pp. 1–23.
- Bode, I., Nadibaidze, A., & Qiao-Franco, G. (2025). Visuals as Sources of Normative Substance in the Debate about Artificial Intelligence in the Military Domain. Critical Studies on Security, pp. 1-26.
