Bridging Understandings of the Military AI Lifecycle: A Transdisciplinary Socio-Technical Approach to Governance

[Jessica Dorsey is an Assistant Professor of International Law at Utrecht University School of Law; Zena Assaad is an Associate Professor at the School of Engineering, Australian National University; Elke Schwarz is a Professor of Political Theory at Queen Mary University London; Ingvild Bode is a Professor of International Relations, University of Southern Denmark. The authors are all members of the Independent Advisory Board on Legal Reviews of the Responsible by Design Institute.]

As governments, military officials, technical experts, legal advisers, scholars and civil society representatives gather in Geneva next week for the United Nations Office for Disarmament Affairs (UNODA) informal consultations on emerging technologies and international security, discussions of military artificial intelligence (AI) will repeatedly engage a familiar concept: the lifecycle.

The term has proliferated in recent years and across contexts: It is found in technical frameworks, governance frameworks, procurement guidance, safety engineering practices, best practices toolkits, and increasingly in international discussions concerning military AI and autonomous systems. Yet despite its prevalence, there are unresolved divergences in how the lifecycle for military AI systems is understood and represented.

Some may chalk this up to a semantic disagreement, but, we posit, it goes beyond this. Different understandings of the military AI lifecycle shape fundamentally different approaches to governance, responsibility, accountability, and ultimately the role of human judgment in military decision-making. This proposition sits at the center of a side event we are holding at next week’s UNODA meetings, Assurance, Accountability, and Autonomy: Responsible Military AI Lifecycle Governance. In preparing for this event, we encountered these interpretive divergences ourselves and found them important to surface via this post, with the aim to ground the discussion in a more nuanced approach to governance. Bringing together perspectives from systems safety engineering, international humanitarian law, ethics, political science, and political theory, our approach is structured around a shared dilemma: what does it mean to govern military AI across the entire lifecycle? The answer depends, in part, on which lifecycle we are talking about.

The Duality of Lifecycle Language

Within engineering communities, the lifecycle has a clear and specific meaning. It represents a structured sequence of stages through which a system is developed and brought into operation. Frameworks such as those developed within the IEEE ecosystem by one of the authors of this piece (and about which two others of us expanded upon here and here) frequently identify phases such as planning, design, development, deployment, operation, maintenance, and decommissioning. The engineering lifecycle consists of a linear set of phases which bring a system into operation. While there are cyclical processes which are conducted in the various lifecycle phases, the overall engineering lifecycle of a system involves a linear progression of activities. 

This understanding helps identify where safety practices, testing requirements, validation procedures, and assurance mechanisms should be implemented across the lifecycle for systems, including AI systems. It allows organizations to establish processes for assurances, legal compliance and evaluating system performance, among many other activities. But when lawyers, ethicists, political theorists, political scientists and international relations experts speak about lifecycles, they are often describing something quite different. As our own interdisciplinary exchanges (bringing together insights from multiple academic disciplines) have demonstrated, the “lifecycle” of military AI systems functions as a “travelling concept” in Mieke Bal’s sense. It is a shared term that moves across disciplines and conversations while carrying different assumptions, meanings and governance implications, a dynamic one of us has previously explored in the context of military AI governance. 

From an interdisciplinary perspective, the lifecycle of an AI system is not confined to a linear progression of engineering stages. It is a socio-technical approach to governance through which technologies, institutions, organizations, norms, legal and moral obligations, professional practices, and political incentives continuously interact. Responsibility and accountability are not confined to individual moments of development or deployment. Instead, they emerge through ongoing relationships between engineers, commanders, operators, procurement officials, legal advisers, intelligence personnel, policymakers, and affected populations. It is not a series of steps, but rather a non-linear recursive process of reflexivity, iteration and modulation at each nexus to make sure that emergent properties and practices remain responsible. 

The nuance here is important because many of the most significant risks associated with military AI arise out of their socio-technical character, through interactions between technological systems, humans and human institutions rather than solely within technical design decisions or operational use. Questions concerning accountability, legal compliance, explainability, agency, human judgment, biases within systems, technical complexity, organizational learning, and civilian protection do not fit squarely within a single phase of the engineering lifecycle. They cut across the entirety of it. Therefore, in developing any kind of responsible approach to military AI governance, stakeholders must wrestle with the challenges that emerge in trying to combine the duality of the lifecycle concept; the engineering lifecycle and the socio-technical lifecycle of military AI systems.

The Responsible-By-Design Approach: Moving Governance Upstream

One implication of a responsible approach to military AI governance is that governance cannot begin once a system is already being deployed. Too often, we note, discussions of military AI assume that legal review, ethical assessment, or accountability mechanisms are primarily questions for the end user. Yet many of the most consequential decisions from the perspective of human control, agency and moral and legal responsibility and accountability occur much earlier.

Long before engineers begin coding, organizations determine the goals and objectives of a system (the planning phase of the engineering lifecycle). In the proceeding engineering lifecycle stages, they define operational objectives, establish acceptable risk thresholds, allocate responsibilities between humans and machines, and identify intended use cases. It begins not with the question: how should we use it responsibly, but should we use this system in the first place. These decisions shape not only the systems that are then deployed into operations but the operations themselves. If human control, agency and moral and legal responsibility and accountability are considered only after development has begun, many governance choices have already been made implicitly. It should be noted, the engineering lifecycle is not devoid of governance mechanisms and in fact is structurally emblematic of through-life governance of a system. What we argue here is that the intersection of the engineering lifecycle and the socio-technical lifecycle of military AI systems has introduced additional governance considerations – such as human control, agency and moral and legal responsibility and accountability – which require through-life consideration which is currently not being afforded.

A socio-technical approach to the governance of military AI systems therefore requires adding socio-technical governance specific to military AI systems from the initial planning phase of the engineering lifecycle. This means asking targeted questions around the socio-technical implications of military AI systems before development starts. What legal and ethical assumptions underpin its design? What forms of human judgment must remain central? Who has agency and at what points? These questions span engineering, legal, ethical, political and organizational boundaries and therefore require a parallel approach to answering them. We posit that this approach helps to shape system requirements, safety mechanisms, accountability approaches and procedures, and decision-making architectures from the outset, rather than attempting to retrofit safeguards after key decisions have already been made.

Beyond Compliance Checklists: Documentation Requirements to Ensure Accountability

The growing popularity of lifecycle governance frameworks specific to military AI systems (see, e.g., GGE LAWS, GC REAIM, UN General Assembly) is encouraging and a natural extension of existing engineering governance frameworks. Yet there is also a risk that lifecycle language becomes little more than another procedural checklist. A framework that identifies development stages without specifying meaningful oversight mechanisms may create the appearance of responsible governance without substantively constraining harmful uses. For this reason, we argue that lifecycle governance must be translated into concrete institutional practices. One particularly important mechanism is documentation, or maintaining a record of relevant activities.

Documentation may be perceived as an administrative burden. In reality, it may be one of the most important governance tools available, linking procurement processes to system requirements to accountability mechanisms. Maintaining documentation throughout the lifecycle is a standard practice which creates traceability between decisions made during planning, design, development, testing, deployment, and operational use. It supports the establishment of accountability structures, legal review processes, enables post-incident assessment, and facilitates organizational learning. It also better enables legal compliance with future operations by understanding pre- and post-strike data more comprehensively.

Documentation can also operationalize otherwise abstract concepts such as explainability, traceability and accountability. For example, organizations can document intended use cases, operational assumptions, testing methodologies, performance limitations, accountability structures, records of human review, and lessons learned from operational deployment. Such documentation helps create a transparent record of how systems are expected to function, how they actually perform, and how human decision-makers engage with them. Many of these processes already exist; however, not necessarily tailored specifically to military AI systems. As one of us has pointed out before, transparency afforded in part by this documentation has a number of benefits for states, one of which is that it can help lead to various mechanisms of accountability and ultimately lead to a higher perception of the legitimacy of military operations. In line with standard engineering practices, documentation should begin during the planning phase and continue throughout the entirety of the engineering system lifecycle. 

The Importance of Transdisciplinary Approaches to Governance 

One persistent tendency in the growing field of military AI governance we also note is to divide actors into separate categories. Developers develop. Operators operate. Lawyers conduct legal or compliance reviews. Commanders make operational decisions. This is a natural tendency in many fields as the requirements of expertise can make it difficult to communicate across disciplinary or professional boundaries. However, we are also seeing the risk that socio-technical governance failures may occur because different communities do not fully understand one another’s assumptions, constraints, objectives, and concerns.

A socio-technical lifecycle approach therefore requires a transdisciplinary approach, which integrates interdisciplinarity with the expertise and experiences of non-academic stakeholders to address complex real-world challenges. In the context of military AI, this means continuous engagement among technical developers, operational experts, commanders, legal advisers, acquisition officials, ethicists, and policymakers and can help in establishing understandings across disciplinary boundaries. Operational personnel should participate in planning and development processes. Their experience can inform realistic use cases, testing scenarios, and operational constraints. Likewise, the expertise of technical developers remains essential during deployment, operational evaluation, incident analysis, and post-deployment learning. We also encourage legal advisors and ethics experts to be at the table at every step of the way. 

Lifecycle Governance as Continuous Learning

Finally, governance of military AI systems should be considered across both the engineering lifecycle and the socio-technical lifecycle and should be understood as an ongoing learning process. Military organizations routinely conduct after-action reviews, operational assessments, lessons-learned exercises and, in some cases, exchange best practices. Similar feedback mechanisms should be embedded into military AI governance. Deployment should generate new information that informs future design decisions, testing procedures, procurement requirements, and operational practices in an iterative feedback loop.

This means creating mechanisms for incident reporting, operational monitoring, reassessment following significant modifications, and continuous evaluation of how systems affect human decision-making in practice, echoing the documentation suggestions outlined above. Such feedback loops are standard in other safety-critical domains, including aviation and safety engineering. Military AI should be no different. 

The development and implementation of responsible military AI may ultimately depend on our collective capacity to reconcile and operationalize both technical and socio-technical understandings and frameworks within a transdisciplinary approach to governance. Such a model would aim to hold the distinction between engineering and socio-technical perspectives in productive tension, ensuring that technical rigor and institutional responsibility evolve together across time, stakeholders, operational contexts and decision points.