Pathways Towards Ensuring Human Control and Judgement Across the Military AI Lifecycle

An Australian Army Perspective

Authors: Ingvild Bode and Anna Nadibaidze

Professor, Center for War Studies, University of Southern Denmark, bode@sam.sdu.dk

Anna Nadibaidze

Postdoctoral Researcher, Center for War Studies, University of Southern Denmark, anadi@sam.sdu.dk

Introduction

Militaries around the world are increasingly integrating various types of AI technologies into weapon systems and forms of military decision-making, including in relation to targeting. AI techniques that may be used for such applications include computer vision systems, natural language processing, large language models (LLMs) and big data analytics.[1] In parallel to this accelerating trajectory, the past decade has seen a diversified governance discussion across several international forums. This discussion is characterised by both formally institutionalised processes at the United Nations (UN), especially the Group of Governmental Experts on Lethal Autonomous Weapon Systems (GGE on LAWS) and more ad hoc, multi-stakeholder initiatives such as the Responsible AI in the Military Domain (REAIM) summits.[2] The discussion has matured into lists of potential governance principles in the form of voluntary commitments.[3] This article focuses on the principle that pertains to maintaining the central role of humans in the use of force.

From the beginning of the international debate, stakeholders have been invested in examining how the use of various AI technologies affects the role of humans in military decision-making. Consequently, the importance of safeguarding the human role is mentioned as a central concern in practically all outcome documents pertaining to autonomous weapon systems and AI in the military domain.[4] Various terms have been used to describe this human role, with the most prominent being human control—sometimes with the qualifier ‘meaningful’—and human judgement. The June 2026 version of the GGE on LAWS rolling text, for example, refers to ‘human judgement and control’ in combination.[5] At the same time, ‘human control and judgement’ remains contested as a term and lacks a universal definition. Many actors, including from Australia, have expressed reservations about employing the specific term ‘human control’.[6] In this article, we do not take this concept to imply either constant or homogenous types of human supervision in decision-making about the use of force. We understand human control as the capacity to exercise agency: to reasonably understand and foresee an AI system’s functions and effects in a relevant context, to deliberate on context-appropriate actions, and to act on these deliberations.[7] Focusing on human agency as an integral component of human control and judgement allows us to move away from the binary assumption of a hierarchical relationship of humans and AI technologies.[8] Broadly understood as ‘the ability to take actions or affect outcomes’,[9] the term ‘human agency’ captures a nuanced account of the relationship between humans and AI technologies, including how types and degrees of human–AI interaction influence the capacity for any exercise of human agency.[10]

Given that safeguarding human control and judgement is such a longstanding concern, there have been some efforts directed at turning this principle into practice.[11] Many of these efforts recognise that securing human control and judgement cannot take place at the stage of employing particular applications of AI alone but needs to happen across the whole lifecycle of AI systems, and especially at the early lifecycle stages of pre-development, design and procurement, as well as testing, evaluation, verification and validation (TEVV).[12] Securing the exercise of human control and judgement will therefore involve various groups of humans who exercise their agency in a way that is appropriate to the context.[13]

This paper sketches concrete pathways for implementing the widely recognised principle of human control and judgement in the use of force through highlighting sets of best practices across the AI lifecycle for policymakers, military personnel and developers.[14] These pathways are based on work conducted by the AutoPractices project, which brought together 49 diverse stakeholders to co-create a best practices toolkit to sustain the human role in the context of using AI-based systems in the military domain.[15] To make the potential application of these practices more concrete, we connect their exercise to the case of AI-based decision support systems (AI DSS). We define AI DSS as ‘tools that use AI techniques to collect and analyse data and provide information about the operational environment, as well as actionable recommendations, with the aim of aiding military decision-makers’.[16]

The remainder of the paper is structured as follows. First, we highlight why human control and judgement continues to occupy such a central place within debates about AI in the military domain, foregrounding safety, legal and ethical reasoning. Second, we situate the Australian Government’s position regarding human control and judgement in the context of integrating AI in military decision-making, including Australia’s participation in the international governance debate. Third, we review how sets of best practices along the AI lifecycle can contribute to operationalising human control and judgement through unpacking its safety, legal and ethical components. We focus, in individual sub-sections, on best practices that can be performed by Australian policymakers, military personnel and developers within the context of developing and using AI DSS, as well as potential points of tension and implications for the Australian Department of Defence (Australian Defence). Finally, we conclude by sketching open questions and areas of inquiry.

The Importance of Exercising Human Control and Judgement: Safety, Legal and Ethical Purposes

Securing the retention of human control and judgement in military decision-making pertaining to the use of force continues to receive broad support.[17] For example, at the March 2026 meeting of the GGE on LAWS, state delegations characterised the human dimension as central to the group’s mandate of proposing ‘a set of elements of an instrument … to address emerging technologies in the area of LAWS’.[18] It is important to underline this enduring importance, because the governance debate around AI in the military domain is under stress at the time of writing (June 2026). The United States, for example, moved away from its previous commitment to soft governance and explicitly suggested removing references to human control and judgement from the GGE on LAWS rolling text at the March 2026 meeting. Simultaneously, the 2026 REAIM Summit outcome document was only endorsed by 39 states, compared to the 61 states who had endorsed the 2024 REAIM Blueprint for Action. The AI governance debate is therefore also affected by a deteriorating global security situation and an ongoing rupture in the international rules-based order, visible in the strains put on the regulatory capacity of international law.

The importance of the human control and judgement principle has become an almost intuitive reference point in the international debate, but it is important to spell out clearly why this is the case.[19] As Davidovic highlights, policies implementing this principle need to be precise—both about what types of human roles there are across the AI system lifecycle and about what purpose the human role serves.[20] We unfold safety, legal and ethical purposes of ensuring the exercise of human control and judgement. As described in the introduction, we consider human agency as an integral component of enabling human control and judgement. It is particularly important to ensure that AI DSS are employed in a way that strengthens humans’ ability to act and affect outcomes, instead of undermining their potential to exercise their agency.[21]

Safety purposes: Retaining human control and judgement is often characterised as an important safety guardrail in deploying AI technologies in the military domain. This is essentially a functional argument that is tied to the continued concerns around both the reliability and predictability of AI technologies.[22] For example, LLMs inevitably produce factually incorrect information as part of their core statistical design.[23] The purpose of retaining human control and judgement aligns with that of other safety-critical domains where the human being adds value in compensating for LLM unreliability and spotting incorrect information.[24]

Legal purposes: Across the long-running legal debate, two (somewhat) distinct purposes of human control and judgement have been spelled out. First, some contend that such a principle is already part of the spirit of international law, including international humanitarian law (IHL), albeit in implicit form. As Heyns argues, ‘since the use of force throughout history has been personal, there has never been a need to make this assumption explicit’.[25] As the very institution of (international) law and justice is therefore human-centric,[26] the role of the human needs to be maintained for the purpose of overall institutional stability.[27] Second, others argue that the exercise of human control and judgement is the basis for and a means to ensure compliance with international law,[28] as ‘only natural persons are capable of administering the performance of IHL obligations binding on States’.[29]

Ethical purposes: Centrally retaining the exercise of human control and judgement ensures the continued presence of ethical and moral reasoning. This serves the purpose of countering the risk of a technological systematisation of violence associated with AI technologies in the military domain.[30] This line of reasoning underlines that warfare is a human activity that is characterised by vast moral dilemmas and therefore requires contextual human decision-making. Having unfolded the safety, legal, and ethical purposes that human agency serves, we now highlight Australia’s approach to AI in defence and its positioning in the global governance debate.

Responsible AI and Human Control in Australian Defence

Australia is both a key developer of AI in defence and a prominent participant in governance debates on AI in the military domain.[31] Its 2026 and 2024 National Defence Strategies highlight that ‘as a middle power, Australia must seek military advantage in innovative ways, including through developing asymmetric advantage’,[32] which includes developing technologies grouped under the term ‘AI’. This objective is also stated in the 2024 Defence Innovation, Science and Technology Strategy, which describes innovation as key for Australia’s strategic position as a middle power in the Indo-Pacific great power competition.[33] As part of this strategic positioning in the regional and global orders, Australia’s approach seeks to particularly leverage AI for human–machine teaming, as well as to innovate and generate asymmetric advantage.

Many of the AI capabilities that Australian Defence has been developing, testing and demonstrating focus on AI technologies being used as a tool to support military personnel. In January 2026, the Australian Government announced an investment of almost AU$40 million in collaborations with research institutions for ‘decision advantage’.[34] Official documents and discourses highlight that the main objective is not to replace humans but to assist Australian Army personnel in decision-making, as this is what is perceived to give strategic advantage.[35] As highlighted by Wing Commander Kate Conroy during Exercise Talisman Sabre 2025, which involved an AI-based system tested on a flight of the P-8A Poseidon aircraft, ‘The AI adds value when the crew need it, and helps them make faster, smarter decisions … AI is a helper, not a decider’.[36] The Royal Australian Air Force has outlined its vision of becoming ‘AI-ready’ and developing a ‘concept of human-machine teaming’ where AI DSS play a key role.[37] One of the systems disclosed by the Air Force is the cognitive assistant avatar Sergeant AIMEE, a chatbot that is portrayed as being useful in the education and training of personnel.[38]

At the same time, Australia has been actively involved in global debates on potential governance of AI in defence, such as discussions on LAWS held at the UN under the framework of the Convention on Certain Conventional Weapons. Throughout its participation in UN expert meetings on LAWS, Australia has highlighted its commitment to the principle of safeguarding the human role in the use of force, although without supporting new legally binding measures on human control and judgement in the development and use of autonomous weapons. It ‘considers “human control” or “human involvement” as one means—but not the only means—of ensuring compliance with’ international humanitarian law, adding that the ‘level of human control or human involvement will vary depending on the context of the use of a LAWS’.[39]

Australia has also endorsed all the outcome declarations from the three REAIM summits: the Call to Action (2023), Blueprint for Action (2024) and Pathways to Action (2026). It joined the 2023 US Political Declaration on Responsible Military Use of AI and Autonomy.[40] Moreover, it voted in favour of UN General Assembly resolutions on LAWS (2023 and 2024) and on ‘Artificial intelligence in the military domain and its implications for international peace and security’ (2025). As noted in the introduction, most of these initiatives include a reference to safeguarding the human role, although these references might be formulated using different terms.[41] For example, the 2026 REAIM Pathways to Action document mentions that ‘AI-enabled decision support systems should support, not replace, the exercise of human judgement’.[42]

Building on Australia’s commitments, Australian documents have proposed practical ways of implementing the principle of human control and judgement in the use of force, including in the context of autonomous weapon systems. In its 2019 non-paper submission to the GGE on LAWS, Australia outlined the concept of ‘system of control’ applying to ‘all aspects of a weapon system from design through to engagement’, which it also implements in its use-of-force procedures.[43] This system is meant to enable ‘controls to be tailored to the specific AWS and its unique operating environment’.[44] This non-paper notes that Australia finds the specific term ‘human control’ to insufficiently ‘cover the plethora of practical controls or systems utilised by states and their militaries’.[45] However, the proposal for the system of control emphasises the need to implement the human role[46] and to ensure that the use of weapon systems is ‘driven by human direction’.[47] It outlines control mechanisms as part of various stages of the lifecycle of autonomous weapons, ranging from initial political decision-making (‘legal and policy framework’) to post-weapon use review (‘after-action evaluation’). While this framework of control concerns the case of weapon systems, it signals Australian Defence’s commitment to engage in practical efforts to operationalise responsible AI in the military domain, as well as to implement a lifecycle-based approach that stretches beyond the stage of employing weapon systems.[48]

The Australian Government also highlights its objective to adopt a comprehensive framework spanning different domains, noting that all technologies, ‘including AI, are subject to multi-layered control measures, comprised of technical, legal and doctrinal measures including test and evaluation, and capability-specific training and policies’.[49] From this perspective, there are various control measures that may be implemented at different stages of the lifecycle, depending on the context of development and intended use.[50] Australia’s Policy Settings for Responsible Use of Artificial Intelligence in Defence, which ‘implement Australia’s commitments’ under the REAIM output documents and the US Political Declaration, state that the responsible use of AI ‘must be considered at all stages of the technology life cycle’.[51] This includes, according to these policy settings, risk-based control measures to ‘identify, assess and mitigate’ risks related to AI, ‘at every stage of the technology lifecycle’.[52] As its delegation noted during the 2026 UN Informal Exchanges on Artificial Intelligence in the Military Domain, Australia supports an approach where responsible use of AI ‘must be integrated from the early lifecycle stages’[53] and where ‘effective oversight throughout the lifecycle is critical’.[54]

In sum, Australia’s strategic and diplomatic documents underline that the development and acquisition of AI systems is meant to support military personnel’s decision-making. At the same time, its practices in relation to responsible AI and governance debates demonstrate its official commitment to a context-appropriate exercise of human control and judgement across the lifecycle of AI systems. The Australian Government has indicated its interest in moving towards ‘practical efforts’ to implement responsible development and use of AI in defence. [55] Getting the right balance between innovation and robust governance throughout the AI lifecycle is therefore a ‘key priority for Australia’.[56] In the next section, this article explores sets of best practices across the AI lifecycle that may be relevant to ongoing efforts within Australian Defence, considering its commitment to a lifecycle-based approach.

Practices Contributing to the Exercise of Human Control and Judgement Across the AI Lifecycle: the Case of AI DSS

Various efforts to operationalise the human control and judgement principle draw on AI lifecycle frameworks applied to the military domain—including in Australia, as demonstrated in section 2. Lifecycle-based frameworks combine understandings across the engineering lifecycle and the military targeting cycle, while also being attentive to the inherent socio-technical dynamics.[57] Taken as a whole, a lifecycle framework depicts multiple avenues for exercising human agency and thereby ensuring human control and judgement across the recurring stages of AI development for military contexts, encompassing design, training, testing, procurement, deployment and post-use review. We follow a simplified version of the lifecycle framework developed by the IEEE Standards Association Research Group on Issues of AI and Autonomy in Defense Systems.[58] This model distinguishes between four broad stages:

(1) Plan: this includes defining the scope and determining the objectives of a system

(2) Design: this involves outlining specific requirements, architectures, functions, interfaces etc. of the system

(3) Develop: this involves the actual building of the system, which includes TEVV

(4) Deploy and post-use: this involves integrating the system into operations, which includes maintenance and post-use assessment throughout the lifecycle of the system.[59]

We review different practices applicable to three core sets of stakeholders—policymakers, military personnel, and developers—with a focus on the context of developing and employing AI DSS. In the fourth sub-section, we reflect on the potential tension between some of the practices and the importance that Australian Defence attributes to speed and optimisation.

Policymakers

Policymakers are central focal points for the exercise of human agency as a key dimension of human control and judgement across both ‘stage 1: plan’ and ‘stage 4: deploy’. We highlight three selected practices for policymakers to perform at these stages. These practices not only mark important types of human agency in their own right but also have knock-on effects, like all practices at early AI lifecycle stages, for subsequent exercises of human control and judgement along the lifecycle.[60] Therefore, such policymaker practices play a central role in ensuring that any employment of AI DSS across a military organisation reflects deliberate, explicit choices made early on, rather than having those determinations pushed to later stages of the lifecycle or even left to happenstance.

Set clear objectives and boundaries for the AI DSS to be developed or acquired (stage 1: plan). Before the development of any AI DSS or other capability, political and strategic decision-makers set as clear a plan as possible around foundational system components. This includes considering at least three aspects: (a) the clear and concrete goals that the AI DSS is meant to serve, (b) planned contexts and environments for the use of the AI DSS, and (c) whether the development and use patterns of the AI DSS align with broader political, strategic, ethical and societal considerations. This step recognises that such foundational parameters of system design amount to socio-political questions that require making explicit, recorded choices.

Setting clear objectives and boundaries for the AI DSS before development and acquisition serves safety, legal and ethical objectives. In terms of safety, it aligns with engineering practice in safety-critical settings, which includes documenting a concept of operations (CONOPS) ‘describing the characteristics and intended usage of a proposed or existing system from the viewpoint of its users’.[61] This is intended to align expectations across different stakeholder groups involved throughout the AI DSS lifecycle and therefore creates an important throughline of communication and expectation.

Going through this practice as a process should also serve to create human-set guardrails around legally and ethically defined undesirable ways of employing AI DSS. In legal terms, any AI DSS that ‘contribute[s] to the conduct of hostilities (e.g., target identification, collateral damage estimation, selection of weapons) fall[s] within the scope of Additional Protocol I (AP I) Article 36 legal reviews’.[62] This means that the central question of whether the use of AI DSS may, even inadvertently, lead to indiscriminate targeting, which is clearly prohibited under IHL, needs to be raised and addressed at this early stage.[63] In ethical terms, this practice could serve to set limits on the use of AI DSS that exceed positive legal obligations from the get-go, such as clear commitments to higher levels of civilian protection or exercising a higher degree of precautions as a result of using particular AI DSS. Highlighting these vital legal and ethical considerations should therefore also entail identifying use cases and applications where AI DSS are unsuitable. This has the potential to counterbalance the detrimental kneejerk response that AI DSS are the solution for every possible identified capability need.

Design a comprehensive risk assessment framework and make (political) choices regarding risk acceptance and risk thresholds (stage 1: plan). This practice involves constructing a framework to evaluate various risks associated with employing a particular type of AI DSS and using such a framework as a guide throughout the AI DSS’s lifecycle. Assessing risks associated with AI DSS is not only an important exercise of human agency by policymakers before such a system is even developed. It also ties policymakers to setting clear benchmarks to ensure that AI DSS are used for intended objectives and enable retaining human control and judgement.

There is currently no risk assessment framework that is specific to AI DSS in the military context, at least not in the public domain. Such a risk assessment framework could be designed to incorporate quantitative and qualitative evaluations of risk,[64] organised into tiers that classify AI DSS according to their risk profile—for example high-risk, medium-risk and low-risk categories.[65] Particular types of AI DSS would be classified into these risk categories depending on factors such as intended purpose of use (e.g., targeting, command-and-control, logistics), intended context of use (e.g., populated/unpopulated areas) or data quality. These factors are key in delineating high-risk, medium-risk and low-risk categories because they matter in assessing the risk of physical harm as well as the risk posed to civilian populations that can be associated with using particular types of AI DSS.

As risk assessment frameworks inform the practices of actors across the entire lifecycle, we consider them to be a particularly vital way of exercising human control and judgement. Thinking about AI DSS in terms of comprehensive risk assessment frameworks and the human roles therein serves legal, ethical and safety purposes. It follows that classifying a system as low, medium or high risk triggers different measures for the design, deployment and use of such systems. To illustrate, we focus on a high-risk AI DSS where the stakes are highest.

At the point of planning the design of high-risk AI DSS, policymakers need to consider how to minimise civilian harm. Such planning should include design provisions for the AI DSS—for example, the capacity to recognise and track civilian populations and alert forces to their presence.[66] Such provisions could also elevate ethical responsibility in using AI DSS,[67] thereby explicitly integrating the potential humanitarian benefits of such systems in proactive ways. This could entail, for example, prioritising civilian protection as a positive duty and featuring procedures for human control along the lifecycle in ways that better address unintended but foreseeable harms. At a more fundamental level, such measures should also include working towards actively addressing the higher likelihood of enemy dehumanisation and countering the risk of moral numbness in the operator that has been analytically associated[68] and empirically reported in the use of AI DSS.[69]

Systems with high risk profiles would necessitate greater human intervention.[70] This could involve, for example, rigorous scrutiny and careful review processes where human operators of AI DSS cross-verify the system’s outputs before acting on them. This serves the legal purpose of ensuring that the use of a particular AI DSS in a particular context complies with core IHL principles, particularly the principle of precaution. A high risk profile should also involve slowing down the overall decision-making speed. AI technologies can allow military personnel to access a much wider array of sources in making crucial decisions—for example, legal judgements—around key principles such as precautions and proportionality. This also echoes longstanding military practices around tactical patience, clearly acknowledging that speeding up decision-making across the board cannot be an end in itself.[71]

Ensure that capability needs and objectives are met by the products offered (stage 4: deploy). This practice is at the heart of procurement processes, serving to ensure that critical choices made by policymakers at stage 1 are implemented at the moment of acquiring a specific AI DSS, rather than being displaced by shifting vendor priorities.[72] While maintaining such a throughline of policymaker decisions from planning to procurement will generally be central, the risk of the widespread hype narratives around AI technologies shaping procurement decisions warrants particular attention.[73] An important element of this practice would be to engage in open conversations with diverse potential vendors rather than becoming locked into working with a limited range of suppliers circulating (potentially) hyped products. Unpacking the hype narratives should also translate into careful communication of TEVV results to military commanders and operators who will use the procured AI DSS.

At its heart, this practice therefore concerns safety requirements by ensuring that relevant military personnel have access to essential information needed to comprehend how specific AI DSS operate within their intended operational contexts. Yet the ability to pierce hype narratives around AI technologies associated with this practice also has important ethical implications because it can contribute to maintaining a sense of moral agency among military personnel using AI DSS. It can allow military personnel to retain a sense of the moral trade-offs that are inherent in warfare and avoid falling into the techno-solutionist trap that ‘armed conflict can be remade into manageable problems and “solved” once the right algorithm is developed’.[74]

Military Personnel

Military personnel are the central actors across ‘stage 4: deploy and post-use’ and their practices have long been a focal point for sustaining human agency.

Develop comprehensive training programs and protocols for all military personnel that either employ or recommend employing AI DSS. Educating humans, including but not limited to operators and commanders, about relevant AI systems and how they are integrated into a network (which is often the case for AI DSS) plays a key role in retaining adequate levels of human control. Training should be comprehensively structured. It should include technical training around, for example, AI literacy—in other words, a basic technical and conceptual understanding of AI systems.[75] Such basic training should be seen as a premise for ensuring that military personnel can fulfil both legal and safety functions in exercising their role. Yet training should also cover human–machine interaction across socio-technical components and how humans engage with AI systems. This type of training holds particular significance as it helps cultivate and deepen awareness and understanding among military personnel of decisive phenomena including automation bias, cognitive biases and sycophantic behaviour.[76]

To counteract inherent biases and assumptions embedded in human–machine interaction, a sustained training program addressing both the dynamics of such interactions and their effects on human control is essential.[77] This type of training should therefore provide broader, more contextualised knowledge about how AI DSS work in practice and how this shapes the decision-making space for military personnel. This knowledge is likely to be particularly crucial for legal and ethical purposes as it centres on the retention and sustenance of human agency. This requires training that takes a deeper dive into the underlying, socio-technical logics of AI DSS.

Adequate comprehension of AI DSS and how humans interact with them should strike a balance that guards against users becoming overly confident or reliant on such technologies. Continuous training initiatives also aim to inform involved military personnel about the potential ramifications of deploying AI DSS in different contexts, spanning strategic, legal and ethical implications. Such knowledge enhances humans’ capacity to intervene effectively and reflect on, dispute or reject AI DSS recommendations as necessary and appropriate to the context—all qualities that underlie legal and ethical purposes that human control and judgement need to fulfil. A comprehensive training regime, moreover, should include recognition that its delivery and curriculum will have to be adaptive and evolve alongside any technological and operational changes.

Establish ways for military personnel to understand when AI DSS do not work according to planned objectives and requirements. The history of integrating AI predecessor technologies (such as automation and autonomy) exhibits various examples where known failure modes were not (adequately) communicated to the human operators of such systems.[78] Yet this is an essential component that allows humans to decide how best to act in cases of system malfunctions, through training protocols around steps to follow in unexpected situations. This practice also involves alerting military operators to blind spots in the AI DSS training data, as well as edge cases. This practice can be primarily associated with the safety purposes of human control, as it can enable military personnel to prevent mistakes from happening.[79]

It can therefore be an important step towards strengthening human operators’ ability to understand and assess the outputs of AI DSS. It also provides an important counterbalance against potential risks of human de-skilling and cognitive dependence that have been found in research around human–AI interaction in the civilian domain.[80] Such a practice also encompasses generating awareness around potential effects of machine–machine interaction—especially when various AI DSS may be used as part of a complex system of various intersecting types.[81]

Institute feedback loops between military personnel and developers across the AI DSS lifecycle. This serves to sustain the exercise of human control and judgement in providing a consistent throughline of information from end user to developer. It ensures that key design choices fulfil their intended objectives and enable rather than delimit human agency. This practice is based on the ability to discuss particular uses of AI DSS in context and on military personnel sharing experiences of interacting with AI DSS. Creating such a throughline of communication serves primarily both safety and legality purposes. In safety terms, it allows reporting of potential malfunctions and making assessments about their relative significance in an operational context. In legal terms, it provides important information about whether design choices actually had the intended effect in terms of sustaining human control and judgement. The impact of this practice ultimately rests on militaries maintaining an organisational culture that is somehow open to feedback and exchange.

Developers

Practices contributing to the exercise of human control and judgement are applicable to developers, including those from the private sector and research institutes collaborating with Australian Defence.[82] As the Global Commission on Responsible AI in the Military Domain notes, industry actors’ ‘decisions significantly influence whether downstream users can integrate [AI] technologies in ways that align with ethical, legal, and operational standards’.[83] The role of developers of AI systems such as AI DSS is particularly central to stages 2 (design) and 3 (develop).

Incorporate feedback from users via communication mechanisms. As mentioned in the last practice listed in section 3.2, appropriate communication mechanisms between end users and developers are crucial to ensure that AI DSS are developed to match the operational and strategic objectives, as well as meeting legal obligations and ethical commitments. Feedback from military personnel is useful in making design choices about AI DSS and developing the systems, as ultimately the systems are meant to enhance the users’ capability to exercise their agency according to the appropriate context. This includes involving users in the design of the interface, particularly to ensure that these design choices match the intended role of AI DSS and their operational value within the context. This practice fulfils primarily a safety purpose, as its main objective is securing alignment between the development of AI DSS and their intended uses and addressing risks associated with those contexts. At the same time, institutionalising such feedback mechanisms also contributes to enhancing military personnel’s ability to engage in legal decision-making in their use of AI DSS, particularly in relation to IHL principles which require the exercise of contextualised human control and judgement.[84]

At stage 2, design AI DSS interfaces that present information in a manner appropriate to the context, while considering the limitations and risk assessment frameworks set previously (at stage 1). The way in which interfaces present information to human decision-makers deserves particular attention, as interfaces are meant to strengthen users’ capacity to consider various options, deliberate on courses of action, and decide on the one(s) most appropriate to the context. Developers should pay attention to interface design because it has potential to enable and constrain the exercise of human agency and contribute to safety, legal and ethical considerations.[85]

For example, interfaces that constantly change the way they present information or that prime users towards a particular course of action would be less likely to enable users’ agency at later stages, including at stage 4. Interface designs that feature a balance between necessary information and cognitive overload contribute to military personnel’s ability to exercise control and judgement from safety, legal and ethical standpoints. A balanced approach to interface design that includes some level of preselection but does not lean into oversimplification is more likely to empower users of the AI DSS. Design choices at stage 2 should incorporate decisions made at stage 1, including the clear objectives and boundaries of AI DSS and any risk assessment frameworks identified by relevant policymakers. Interfaces of AI DSS should also incorporate ways for users to be aware of when systems are not performing as expected.

Conduct TEVV procedures that incorporate concrete metrics on various dimensions (stage 3). TEVV procedures should incorporate concrete metrics on a variety of aspects, especially on whether intended users are empowered to engage in decision-making that is compliant with IHL principles and appropriate ethical frameworks, as well as operational and security objectives. This practice therefore fulfils safety, legal and ethical purposes associated with the exercise of human agency, which contributes to human control and judgement. Testing of AI DSS and their components should ultimately ensure that intended users possess the appropriate mechanisms to deliberate on actions in relation to not only expected conditions of use but also edge cases or unexpected scenarios. Comprehensive procedures that match the limitations and risk assessment frameworks set previously (at stage 1) are particularly crucial in TEVV. Australia’s system of control framework from 2019 also highlights the importance of comprehensive testing and evaluation in autonomous weapon systems.[86]

At stage 3, test the AI DSS interfaces to identify and document limitations. Developers can enable intended users’ capacity to exercise agency by involving these users in the testing of the interface and monitoring what the use of these AI DSS means for military personnel’s decision-making in a particular context. In other words, it is beneficial for testing practices to include aspects of human–machine interaction, including dynamics such as automation biases or sycophancy, rather than exclusively focusing on the technical performance of AI DSS. This practice builds on other interface design related practices mentioned earlier, underlining the importance of consistency and alignment across the lifecycle stages.[87]

Ensure that appropriate datasets are used and document potential limitations or biases in data. Across both stages 2 and 3, developers can engage in many practices that ensure transparency surrounding the data used to train the AI DSS. These practices are important as they strengthen users’ ability to understand how AI DSS may (mal)function and reasonably foresee the impact of using these AI DSS within particular contexts. Developers should take measures to improve data quality as much as possible and record limitations, as part of both design and testing. Documentation regarding training data is a way to provide transparency surrounding potential biases in the data that the AI DSS were trained on. Such documentation informs other humans across the lifecycle, such as military personnel and policymakers, about limitations to the data and strengthens their ability to judge whether AI DSS are appropriate to a certain context. There are several key questions regarding the training data, including concerning the selection, labelling and assessment of the data involved in the development of the AI DSS. This practice is particularly relevant for the safety purposes of ensuring the exercise of human control and judgement. It seeks to mitigate safety risks involved with incomplete or wrong data in a military context, which may exacerbate security risks such as the brittleness of AI DSS.[88]

Implications for Australian Defence

The practices outlined throughout section 3 contribute to operationalising the principle of human control and judgement across the lifecycle of AI DSS used in the military domain. They are therefore valuable to Australian Defence as it seeks to both promote innovation in its armed forces and implement its commitment to ensuring the human role in the use of force. Many of these practices also match the approach adopted in Australia’s ‘system of control’ 2019 non-paper, such as comprehensive TEVV, training of personnel, setting specific parameters of use, and conducing post-use reviews.[89] Although Australia’s 2019 proposal is specific to the case of weapon systems, and the practices listed here focus on the development and use of AI DSS, there are many points of intersection that are relevant to Australian Defence.

At the same time, the application of some of these practices may entail points of discussion, tension and even resistance. In certain contexts, a comprehensive risk assessment framework which would involve human intervention, review processes and cross-checking mechanisms can be considered an obstacle to reaping the benefits of employing AI DSS. The development and the use of AI DSS are associated with speed, scale and optimisation.[90] Mechanisms that would slow down the tempo of decision-making are associated with reduced efficiency, which would be seen as countering the priority attributed to developing AI technologies in defence. As Australia’s 2026 National Defence Strategy states, ‘Defence must keep pace with these changes [in emerging technologies including AI] to leverage opportunities, respond to evolving threats and remain interoperable with allies and partners’.[91] There is therefore a potential point of tension between operationalising the role of the human on one hand, and the strategic importance given to increasing the armed forces’ efficiency and the speed of decision-making on the other.

To meet Australian Defence’s strategic objectives of developing asymmetric advantage and bolstering ‘the ADF [Australian Defence Force]’s ability to deter and respond’,[92] the use of AI DSS should ‘extend users’ will’ rather than undermine it.[93] This extension of the will—and the use of AI to truly support and help operators—requires humans across the lifecycle to exercise their control and judgement in a way that matches Australia’s priorities as well as increasing its interoperability with allies.[94] Ultimately, the exercise of human agency which contributes to human control and judgement in the development and use of AI DSS involves a set of common practices but not a ‘one-size-fits-all model’.[95] Each set of actors, whether Australian policymakers, military personnel or developers/industry, will therefore have to navigate respective trade-offs to ensure the exercise of human control and judgement.

Conclusion

Global governance debates on military applications of AI are considering the operationalisation of high-level principles agreed across various forums. As a key developer of AI in defence and an active participant of governance debates, Australia has demonstrated its commitment to the principle of the human role in the use of military force and to pursuing the debate on how to implement it in practice. This article has presented security, ethical and legal reasons why this work should continue and has offered sets of practices applicable to various groups of humans—policymakers, military personnel, and developers—within the context of development and use of AI DSS. These practices are based on the findings of the AutoPractices project, which co-created a toolkit of best practices to strengthen the exercise of human control and judgement across the lifecycle of military AI systems.

The practices included in this overview match the Australian approach to the human role presented in its 2019 non-paper, even though this proposal states that the term ‘human control’ is not the most useful. As Trabucco notes, Australia’s approach broadens the system’s timeline and ‘lengthen[s] the perspective horizontally to broaden the range of actors and decision-making capacities and fold[s] them into a network of human control’.[96] The practices explored in this article can also be integrated as part of a networked approach to the exercise of control and judgement across human actors involved at different stages of the lifecycle, with a particular focus on the early stages (1–3). While there are potential points of tension or even resistance to some practices, this article argues that strengthening this network of practices would assist Australia in achieving its objective of employing AI to gain a decision advantage while remaining committed to legal obligations and safety considerations, especially as achieving this balance is a ‘key priority’ for Australia.[97]

Acknowledgements: We are grateful for the constructive feedback we received on earlier drafts of the article from Nicholas Carroll and Neil Renic at the Special Issue workshop, as well as from Vincent Boulanin as part of the review process. We also thank all the stakeholders who participated in the AutoPractices project process.

Funding acknowledgements: This research was supported through funding by the European Research Council via the European Union’s Horizon 2020 research and innovation program (grant number 852123, the AutoNorms project; and grant number 101156237, the AutoPractices project) and the Independent Research Fund Denmark (grant ID 3119-00023B, the HuMach project).

Endnotes

[1] Nathan G Wood et al., ‘Stop Saying “AI”’, arXiv:2602.17729, arXiv (2026): 28, at: https://doi.org/10.48550/arXiv.2602.17729.

[2] While the GGE on LAWS debate involves the 128 High Contracting Parties of the UN Convention on Certain Conventional Weapons, the three REAIM summits that have been held to date have been more limited in geographical scope with, respectively, 60, 61 and 39 states endorsing the REAIM 2023, REAIM 2024 and REAIM 2026 summit outcomes.

[3] Government of the Netherlands, REAIM 2023 Call to Action’, 16 February 2023, at: www.government.nl/documents/2023/02/16/reaim-2023-call-to-action; Ministry of Foreign Affairs, Republic of Korea, Blueprint for Action, 11 September 2024, at: www.mofa.go.kr/www/brd/m_4080/down.do?brd_id=235&seq=375378&data_tp=A&file_seq=9; Ministry of Foreign Affairs, Spain, REAIM Pathways to Action, 2026, at: www.exteriores.gob.es/en/REAIM2026/Documents/REAIM%202026%20Pathways%20to%20Action.pdf; UNODA, GGE on LAWS Rolling Text, Status Date: 05 June 2026 (2026), at: https://docs-library.unoda.org/Convention_on_Certain_Conventional_Weapons_-Group_of_Governmental_Experts_on_Lethal_Autonomous_Weapons_Systems_(2026)/GGE_LAWS_-_Rolling_Text_-_5_June_2026.pdf.

[4] Ingvild Bode, Emerging Norms Around Military Applications of AI: The Case of Human Control, GC REAIM Expert Policy Note Series (The Hague Centre for Strategic Studies, 2025), at: https://hcss.nl/wp-content/uploads/2025/05/Bode-1.pdf.

[5] UNODA, GGE on LAWS Rolling Text, Status Date: 05 June 2026, para. 12.

[6] Australia, Australia’s System of Control and Applications for Autonomous Weapon Systems, submission to United Nations Convention on Certain Conventional Weapons, Geneva, 2019 (UNODA, 2019), para. 42, at: https://docs-library.unoda.org/Convention_on_Certain_Conventional_Weapons_-_Group_of_Governmental_Experts_(2019)/CCWGGE.12019WP.2Rev.1.pdf.

[7] The AutoPractices Project, Strengthening Human Agency in the Military Domain: Best Practices Toolkit for Policymakers, Developers, and Users of AI Systems (Center for War Studies, University of Southern Denmark, 2026), p 6, at: https://doi.org/10.5281/zenodo.17671715.

[8] Bérénice Boutin and Taylor Woodcock, ‘Aspects of Realizing (Meaningful) Human Control: A Legal Perspective’, in Robin Geiß and Henning Lahmann (eds), Research Handbook on Warfare and Artificial Intelligence (Edward Elgar Publishing, 2024); Klaudia Klonowska and Taylor Kate Woodcock, ‘Rhetoric and Regulation: The (Limits of) Human/AI Comparison in Legal Debates on Military AI’, in Bérénice Boutin et al. (eds), Legal, Ethical, and Technical Dilemmas in Military Artificial Intelligence (T.M.C. Asser Press, 2026); Anna Nadibaidze, ‘Investigating the Human Role in Warfare: Beyond “Control” and Towards ‘Agency’?’, The HuMach Blog, 17 February 2025, at: https://humach.eu/investigating-the-human-role-in-warfare-beyond-contro….

[9] Alan Chan et al., ‘Harms from Increasingly Agentic Algorithmic Systems’, 2023 ACM Conference on Fairness, Accountability, and Transparency, 12 June 2023, 651–66, at: https://doi.org/10.1145/3593013.3594033. We use this brief definition as a pragmatic starting point for agency in the context of this practically oriented paper. The notion of agency has a rich conceptual heritage across various social sciences disciplines that we cannot capture in this contribution. See Werner Rammert, Distributed Agency & Digital Technology: A Social Pragmatist View on Human-Technology Interaction (Springer Fachmedien Wiesbaden, 2025).

[10] Ingvild Bode et al., ‘Ensuring the Exercise of Human Agency in AI-Based Military Systems: Concerns Across the Lifecycle’, Ethics and Information Technology 27, no. 50 (2025), at: https://doi.org/10.1007/s10676-025-09861-2.

[11] Vincent Boulanin et al., Limits on Autonomy in Weapon Systems: Identifying Practical Elements of Human Control (Stockholm International Peace Research Institute and International Committee of the Red Cross, 2020), at www.sipri.org/sites/default/files/2020-06/2006_limits_of_autonomy.pdf; Jovana Davidovic, ‘Rethinking Human Roles in AI Warfare’, Nature Machine Intelligence 7, no. 10 (2025): 1593–95, at: https://doi.org/10.1038/s42256-025-01123-6; Merel Ekelhof and Giacomo Persi Paoli, The Human Element in Decisions About the Use of Force (United Nations Institute for Disarmament Research, 2020), at: https://unidir.org/wp-content/uploads/2023/05/UNIDIR_Iceberg_SinglePages_web.pdf.

[12] IEEE SA Research Group on Issues of Autonomy and AI in Defense Systems, A Framework for Human Decision-Making through the Lifecycle of Autonomous and Intelligent Systems in Defense Applications (IEEE SA, 2024), at: https://ieeexplore.ieee.org/document/10707139.

[13] Vincent Boulanin and Dustin A Lewis, ‘Responsible Reliance Concerning Development and Use of AI in the Military Domain’, Ethics and Information Technology 25, no. 8 (2023): 1–5, at: https://doi.org/10.1007/s10676-023-09691-0; Boutin and Woodcock, ‘Aspects of Realizing (Meaningful) Human Control’; Lena Trabucco, ‘AI-Enabled Autonomous Weapons and Human Control. Part I: Human Control and Machine Learning Design and Development’, International Law Studies 106 (2025): 534–78.

[14] IEEE SA Research Group on Issues of Autonomy and AI in Defense Systems, A Framework for Human Decision-Making Through the Lifecycle of Autonomous and Intelligent Systems.

[15] The AutoPractices Project, Strengthening Human Agency in the Military Domain; The AutoPractices Project, Map of Practices (Center for War Studies, 2025), at: www.autonorms.eu/governing-ai-technologies-in-the-military-domain-from-the-bottom-up-map-of-practices.

[16] Marta Bo et al., ‘Views of Members of the Scientific Community and Civil Society Pursuant to Resolution 79/239 “Artificial Intelligence in the Military Domain and Its Implications for International Peace and Security” Adopted by the General Assembly on 24 December 2024, in Accordance with the Request of the UN Secretary-General Contained in Note Verbale ODA/2025-00029/AIMD’, 11 April 2025, at: https://docs-library.unoda.org/General_Assembly_First_Committee_-Eightieth_session_(2025)/79-239-Dorsey-Bo-Bode-Schwarz-EN.pdf.

[17] Laura Bruun, Towards a Two-Tiered Approach to Regulation of Autonomous Weapon Systems: Identifying Pathways and Possible Elements (Stockholm International Peace Research Institute, 2024), at: https://doi.org/10.55163/LPED7967.

[18] UN Convention on Certain Conventional Weapons, Final Report. Meeting of the High Contracting Parties to the Convention in Certain Conventional Weapons. CCW/MSP/2025/8, 27 November 2025, at: https://docs-library.unoda.org/Convention_on_Certain_Conventional_Weapons_-Meeting_of_High_Contracting_Parties_(2025)/CCW-MSP-2025-8_English.pdf.

[19] Rebecca Crootof et al., ‘Humans in the Loop’, Vanderbilt Law Review 76, no. 2 (2023): 429–510.

[20] Davidovic, ‘Rethinking Human Roles in AI Warfare’; Jovana Davidovic, ‘On the Purpose of Meaningful Human Control of AI’, Frontiers in Big Data 5 (2023): 1017677, at: https://doi.org/10.3389/fdata.2022.1017677.

[21] Anna Nadibaidze, ‘Do AI Decision Support Systems “Support” Humans in Military Decision-Making on the Use of Force?’, Opinio Juris, 29 November 2024, at: https://opiniojuris.org/2024/11/29/do-ai-decision-support-systems-support-humans-in-military-decision-making-on-the-use-of-force; Alexander Blanchard and Laura Bruun, Autonomous Weapon Systems and AI-Enabled Decision Support Systems in Military Targeting: A Comparison and Recommended Policy Responses (Stockholm International Peace Research Institute, 2025), at: https://doi.org/10.55163/YQBY3151; Nathan Wood, ‘Autonomous and AI-Enabled Systems: Extensions or Replacements of Human Will and Control?’, Ethics and Information Technology 28 (2026): 1–16, at: https://doi.org/10.1007/s10676-025-09876-9.

[22] Brendan Walker-Munro and Zena Assaad, ‘The Guilty (Silicon) Mind: Blameworthiness and Liability in Human-Machine Teaming’, Cambridge Law Review 8, no. 1 (2023); Mary Cummings, ‘Prohibiting Generative AI in Any Form of Weapon Control’, paper presented at 39th Conference on Neural Information Processing Systems, 2025, at: https://openreview.net/pdf?id=uEY7kQsiZz.

[23] Arthur Holland Michel, Decisions, Decisions, Decisions: Computation and Artificial Intelligence in Military Decision-Making (International Committee of the Red Cross, 2024), at: https://shop.icrc.org/decisions-decisions-decisions-computation-and-artificial-intelligence-in-military-decision-making-pdf-en.html; Jonathan Kwik, ‘Digital Yes‐Men: How to Deal with Sycophantic Military AI?’, Global Policy 16, no. 3 (2025): 467–473, at: https://doi.org/10.1111/1758-5899.70042.

[24] Qiaochu Zhang, ‘Navigating the Complexities of Exercising Human Agency in Human-Machine Interaction Across the AI Lifecycle’, The AutoNorms Blog, 21 October 2024, www.autonorms.eu/navigating-the-complexities-of-exercising-human-agency….

[25] Christof Heyns, ‘Autonomous Weapons Systems: Living a Dignified Life and Dying a Dignified Death’, in Autonomous Weapons Systems: Law, Ethics, Policy, ed. Nehal Bhuta et al. (Cambridge University Press, 2016), 8.

[26] International Committee of the Red Cross, Submission to the United Nations Secretary-General on Artificial Intelligence in the Military Domain (April 2025), www.icrc.org/sites/default/files/2025-04/ICRC_Report_Submission_to_UNSG…; Joanna J. Bryson et al., ‘Of, for, and by the People: The Legal Lacuna of Synthetic Persons’, Artificial Intelligence and Law 25, no. 3 (2017): 273–91; Anna Nadibaidze and Ingvild Bode, ‘Five Questions We Often Get Asked About AI in Weapon Systems and Our Answers’, The AutoNorms Blog, 25 September 2023, at: www.autonorms.eu/five-questions-we-often-get-asked-about-ai-in-weapon-systems.

[27] Davidovic, ‘Rethinking Human Roles in AI Warfare’.

[28] Arthur Holland Michel, Autonomous Refusal in Lethal Weapons (Center for War Studies, University of Southern Denmark, 2026), p. 5, at: https://doi.org/10.5281/zenodo.17811999.

[29] Dustin A Lewis and Hannah Sweeney, Exercising Cognitive Agency: A Legal Framework Concerning Natural and Artificial Intelligence in Armed Conflict (Harvard Law School, 2025), p. 10, at: https://pilac.law.harvard.edu/exercising-cognitive-agency.

[30] Neil Renic and Elke Schwarz, ‘Crimes of Dispassion: Autonomous Weapons and the Moral Challenge of Systematic Killing’, Ethics & International Affairs 37, no. 3 (2023): 321–343; Atay Kozlovski, ‘Meaningful Human Control over AI Military Decision Support Systems: Exploring Key Challenges’, in Bernhard Koch and David Winkler (eds), Artificial Intelligence Ethics in Military Medicine and Humanitarian Healthcare (Springer Nature Switzerland, 2026).

[31] Kate S Devitt and Damian Copeland, ‘Australia’s Approach to AI Governance in Security and Defence’, in Michael Raska and Richard A Bitzinger (eds), The AI Wave in Defence Innovation (Routledge, 2023); Peter Layton, Evolution Not Revolution: Australia’s Defence AI Pathway, DAIO Study 22/02 (Defense AI Observatory, 2022), at: https://defenseai.eu/wp-content/uploads/2023/01/DAIO_Study2202.pdf.

[32] Department of Defence, 2026 National Defence Strategy (Australian Government, 2026), at: www.defence.gov.au/about/strategic-planning/2026-national-defence-strategy-2026-integrated-investment-program; Department of Defence, 2026 Integrated Investment Program (Australian Government, 2026), at: www.defence.gov.au/about/strategic-planning/2026-national-defence-strategy-2026-integrated-investment-program, p. 80. See also Department of Defence, 2024 National Defence Strategy (Australian Government, 2024), at: www.defence.gov.au/about/strategic-planning/2024-national-defence-strategy-2024-integrated-investment-program; Department of Defence, 2024 Integrated Investment Program (Australian Government, 2024), p. 64, at: www.defence.gov.au/about/strategic-planning/2024-national-defence-strategy-2024-integrated-investment-program.

[33] Department of Defence, Accelerating Asymmetric Advantage: Delivering More, Together (Australian Government, 2024), at: www.defence.gov.au/about/strategic-planning/accelerating-asymmetric-advantage-delivering-more-together.

[34] Department of Defence, ‘Harnessing Tech to Drive Faster Decisions Across Defence’, Defence (website), 5 January 2026, at: www.defence.gov.au/news-events/releases/2026-01-05/harnessing-tech-drive-faster-decisions-across-defence.

[35] ‘Artificial Intelligence to Improve Decision-Making’, Department of Defence (website), 18 November 2021, at: www.defence.gov.au/news-events/news/2021-11-18/artificial-intelligence-improve-decision-making; ‘Enhancing AI to support the warfighter’, Department of Defence (website), 9 August 2024, at: www.defence.gov.au/news-events/news/2024-08-09/enhancing-ai-support-warfighter.

[36] ‘AI Takes Flight on Talisman Sabre’, Department of Defence (website), 11 August 2025, at: www.defence.gov.au/news-events/news/2025-08-11/ai-takes-flight-talisman-sabre.

[37] Alexander Manning, ‘Getting Ready for Artificial Intelligence’, Department of Defence (website), 23 November 2021, at: www.defence.gov.au/news-events/news/2021-11-23/getting-ready-artificial-intelligence; Royal Australian Air Force, ‘Getting Ready for Artificial Intelligence’ (video), YouTube, 24 November 2021, at: www.youtube.com/watch?v=J4K71RPcv5U.

[38] Royal Australian Air Force, ‘Getting Ready for Artificial Intelligence’.

[39] Australian Government, Australia’s Submission to the United Nations Secretary-General’s Report on Lethal Autonomous Weapons Systems (UNODA, May 2024), para. 26, at: https://docs-library.unoda.org/General_Assembly_First_Committee_-Seventy-Ninth_session_(2024)/78-241-Australia-EN.pdf.

[40] Defence Ministers, ‘Australia Joins Declaration on Safe and Responsible Artificial Intelligence in the Military’, media release, 3 November 2023, at: www.minister.defence.gov.au/media-releases/2023-11-03/australia-joins-declaration-safe-responsible-artificial-intelligence-military.

[41] Bode, Emerging Norms around Military Applications of AI.

[42] Ministry of Foreign Affairs, Spain, REAIM Pathways to Action, para. 12.

[43] Australia, Australia’s System of Control and Applications for Autonomous Weapon Systems, para. 3.

[44] Ibid., para. 1.

[45] Ibid., para. 42.

[46] Trabucco, ‘AI-Enabled Autonomous Weapons and Human Control. Part I’, p. 546.

[47] Australia, Australia’s System of Control and Applications for Autonomous Weapon Systems, para. 8.

[48] Department of Foreign Affairs and Trade, Australia’s Submission to the UN Secretary-General’s Report on “Artificial Intelligence in the Military Domain and Its Implications for International Peace and Security”’ (Australian Government, 7 August 2025), para. 8, at: https://docs-library.unoda.org/General_Assembly_First_Committee_-Eightieth_session_(2025)/79-239-Australia-en.pdf.

[49] Ibid., para. 8.

[50] Lauren Sanders, ‘Symposium on Military AI and the Law of Armed Conflict: Bridging the Legal Gap Between Principles and Standards in Military AI—Assessing Australia’s “System of Control” Approach’, Opinio Juris, 3 April 2024, at: https://opiniojuris.org/2024/04/03/symposium-on-military-ai-and-the-law-of-armed-conflict-bridging-the-legal-gap-between-principles-and-standards-in-military-ai-assessing-australias-system-of-control-appro.

[51] Department of Defence, Policy Settings for Responsible Use of Artificial Intelligence in Defence (Australian Government, 2026), pp. 5–6, at: www.defence.gov.au/sites/default/files/2026-03/Policy-Settings-for-Responsible-Use-of-Artificial-Intelligence-in-Defence-%5BOFFICIAL%5D.pdf.

[52] Ibid., p. 9.

[53] Australian Delegation at the Informal Exchanges on AI in the Military Domain and Its Implications for International Peace and Security, United Nations in Geneva, 15 June 2026. Quote recorded and transcribed by the authors.

[54] Laura Varella, ‘Report on the Informal Exchange on Artificial Intelligence in the Military Domain’, in AWS Diplomacy Report (Reaching Critical Will, 2026), p. 14, https://reachingcriticalwill.org/images/documents/Disarmament-fora/aws/2026/reports/AWSR3.1.pdf.

[55] Department of Foreign Affairs and Trade, Australia’s Submission to the UN Secretary-General’s Report on “Artificial Intelligence in the Military Domain and Its Implications for International Peace and Security”’, para. 8.

[56] Australian Delegation at the Informal Exchanges on AI in the Military Domain and Its Implications for International Peace and Security, United Nations in Geneva, 15 June 2026. Quote recorded and transcribed by the authors.

[57] Jessica Dorsey et al., ‘Bridging Understandings of the Military AI Lifecycle: A Transdisciplinary Socio-Technical Approach to Governance’, Opinio Juris, 12 June 2026, at: https://opiniojuris.org/2026/06/12/bridging-understandings-of-the-lifecycle-interdisciplinary-approaches-to-military-ai-governance.

[58] IEEE SA Research Group on Issues of Autonomy and AI in Defense Systems, A Framework for Human Decision-Making through the Lifecycle of Autonomous and Intelligent Systems in Defense Applications.

[59] Jessica Dorsey et al., ‘Designing for Responsibility: Outlining an Interdisciplinary Lifecycle Framework Research Agenda for AI-Enabled Weapons and Decision-Support Systems’, conference paper under review, 2026; Jonathan Kwik, ‘Iterative Assessment for Military Artificial Intelligence (AI) Systems’, in Bérénice Boutin et al. (eds), Legal, Ethical, and Technical Dilemmas in Military Artificial Intelligence (T.M.C. Asser Press, 2026).

[60] The AutoPractices Project, Strengthening Human Agency in the Military Domain; Ariel Conn and Ingvild Bode, ‘Establishing Human Responsibility and Accountability at Early Stages of the Lifecycle for AI-Based Defence Systems’, Ethics and Information Technology 27, no. 51 (2025), at: https://doi.org/10.1007/s10676-025-09862-1.

[61] Ali Mostashari et al., ‘Developing a Stakeholder‐Assisted Agile CONOPS Development Process’, Systems Engineering 15, no. 1 (2012): 1, at: https://doi.org/10.1002/sys.20190.

[62] Klaudia Klonowska, ‘AI-Enabled Decision Support Systems as Means and Methods of Warfare’, Legal Review of Weapons Information Portal, 1 October 2025, at: https://legalreviewportal.org/technologies/decision-support-systems.

[63] Ibid.; Tobias Vestner and Altea Rossi, ‘Legal Reviews of War Algorithms’, International Law Studies 97 (2021): 509–555.

[64] See, for example, Department of National Defence and Canadian Armed Forces, Artificial Intelligence Strategy (Government of Canada, 2024), at: www.canada.ca/content/dam/dnd-mdn/documents/reports/ai-ia/dndcaf-ai-strategy.pdf.

[65] Zena Assaad, ‘A Proposed Risk Categorisation Model for Human-Machine Teaming’, paper presented at EICS ’22: Engineering Interactive Computing Systems conference, 21 June 2022, CEUR Workshop Proceedings (2022), at: https://ceur-ws.org/Vol-3404/paper11.pdf; Alexander Blanchard et al., ‘A Risk-Based Regulatory Approach to Autonomous Weapon Systems’, Digital Society 4, no. 23 (2025), at: https://doi.org/10.1007/s44206-025-00181-y; Jack Shanahan, ‘Symposium on Military AI and the Law of Armed Conflict: A Risk Framework for AI-Enabled Military Systems’, Opinio Juris, 1 April 2024, at: https://opiniojuris.org/2024/04/01/symposium-on-military-ai-and-the-law-of-armed-conflict-a-risk-framework-for-ai-enabled-military-systems.

[66] International Committee of the Red Cross, Submission to the United Nations Secretary-General on Artificial Intelligence in the Military Domain, p. 6.

[67] We are grateful to Neil Renic for drawing our closer attention to these points.

[68] Renic and Schwarz, ‘Crimes of Dispassion’; Neil Renic, ‘Tragic Reflection, Political Wisdom, and the Future of Algorithmic War’, Australian Journal of International Affairs 78, no. 2 (2024): 247–256, at: https://doi.org/10.1080/10357718.2024.2328299.

[69] Yuval Abraham, ‘“Lavender”: The AI Machine Directing Israel’s Bombing Spree in Gaza’, +972 Magazine, 3 April 2024, at: www.972mag.com/lavender-ai-israeli-army-gaza.

[70] Kozlovski, ‘Meaningful Human Control over AI Military Decision Support Systems’.

[71] International Committee of the Red Cross, Submission to the United Nations Secretary-General on Artificial Intelligence in the Military Domain.

[72] Netta Goussac and Vincent Boulanin, Responsible Procurement of Military Artificial Intelligence (Stockholm International Peace Research Institute, 2026), at: https://doi.org/10.55163/YOLG1827.

[73] Kevin LaGrandeur, ‘The Consequences of AI Hype’, AI and Ethics 4, no. 3 (2024): 653–656, at: https://doi.org/10.1007/s43681-023-00352-y.

[74] Neil Renic, ‘AI-Optimized Violence and the Suffocation of Moral and Political Wisdom’, Cambridge Forum on AI: Law and Governance 1 (2025): 8, at: https://doi.org/10.1017/cfl.2025.10025.

[75] Stephanie M Tully et al., ‘Lower Artificial Intelligence Literacy Predicts Greater AI Receptivity’, Journal of Marketing 89, no. 5 (2025): 1–20, at: https://doi.org/10.1177/00222429251314491.

[76] Alexander Blanchard and Laura Bruun, Bias in Military Artificial Intelligence (Stockholm International Peace Research Institute, December 2024), at: https://doi.org/10.55163/CJFT9557; Laura Bruun and Marta Bo, Bias in Military Artificial Intelligence and Compliance with International Humanitarian Law (Stockholm International Peace Research Institute, 2025); Kwik, ‘Digital Yes‐Men’.

[77] Holland Michel, Decisions, Decisions, Decisions.

[78] Jeffrey M Bradshaw et al., ‘The Seven Deadly Myths of “Autonomous Systems”’, IEEE Intelligent Systems 28, no. 3 (2013): 54–61; John K Hawley, Patriot Wars: Automation and the Patriot Air and Missile Defense System (Center for New American Security, 2017), at: www.cnas.org/publications/reports/patriot-wars; Ingvild Bode and Tom FA Watts, Meaning-Less Human Control: Lessons from Air Defence Systems on Meaningful Human Control for the Debate on AWS (Drone Wars UK and Center for War Studies, 2021), at: https://dronewars.net/2021/02/19/meaning-less-human-control-lessons-from-air-defence-systems-for-lethal-autonomous-weapons.

[79] Davidovic, ‘On the Purpose of Meaningful Human Control of AI’, p. 2.

[80] Nataliya Kosmyna et al., ‘Your Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing Task’, arXiv:2506.08872, arXiv, 10 June 2025, at: https://doi.org/10.48550/arXiv.2506.08872; Michael Gerlich, ‘AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking’, Societies 15, no. 1 (2025): 6, at: https://doi.org/10.3390/soc15010006.

[81] Chan et al., ‘Harms from Increasingly Agentic Algorithmic Systems’.

[82] On the various roles of industry in military AI development, see Alexander Blanchard et al., ‘Mapping the Military AI Industry’, Stockholm International Peace Research Institute (website), 23 April 2026, at: www.sipri.org/commentary/topical-backgrounder/2026/mapping-military-ai-industry.

[83] The Global Commission on Responsible Artificial Intelligence in the Military Domain, Responsible by Design: Strategic Guidance Report on the Risks, Opportunities, and Governance of Artificial Intelligence in the Military Domain (The Hague Centre for Strategic Studies, 2025), p. 57, at: https://hcss.nl/wp-content/uploads/2025/09/GC-REAIM-Strategic-Guidance-Report-Final-WEB.pdf.

[84] Klonowska and Woodcock, ‘Rhetoric and Regulation’; Yiokasti Mouratidi, ‘Armed Groups and International Law—Beyond Compliance Symposium: Before Compliance—Ex-Ante Decision Making in the Design of Military AI Capabilities and the Need to Unravel the Private Technology Sector’s Role’, Armed Groups and International Law, 23 September 2025, at: hwww.armedgroups-internationallaw.org/2025/09/23/beyond-compliance-consortium-before-compliance-ex-ante-decision-making-in-the-design-of-military-ai-capabilities-and-the-need-to-unravel-the-private-technology-sectors-role.

[85] Ioana Puscas, Human-Machine Interfaces in Autonomous Weapon Systems (United Nations Institute for Disarmament Research, 2022), at: https://unidir.org/publication/human-machine-interfaces-in-autonomous-weapon-systems.

[86] Australia, Australia’s System of Control and Applications for Autonomous Weapon Systems, para. 13.

[87] Bode et al., ‘Ensuring the Exercise of Human Agency in AI-Based Military Systems’.

[88] Anna Nadibaidze et al., AI in Military Decision Support Systems: A Review of Developments and Debates (Center for War Studies, 2024), at: www.autonorms.eu/ai-in-military-decision-support-systems-a-review-of-developments-and-debates.

[89] Australia, Australia’s System of Control and Applications for Autonomous Weapon Systems.

[90] HW Meerveld et al., ‘The Irresponsibility of Not Using AI in the Military’, Ethics and Information Technology 25, no. 1 (2023): 14, at: https://doi.org/10.1007/s10676-023-09683-0; Emelia S Probasco et al., AI for Military Decision-Making: Harnessing the Advantages and Avoiding the Risks (Center for Security and Emerging Technology, 2025), at: https://cset.georgetown.edu/publication/ai-for-military-decision-making.

[91] Department of Defence, 2026 National Defence Strategy; Department of Defence, 2026 Integrated Investment Program, p. 79.

[92] Ibid., p. 79.

[93] Wood, ‘Autonomous and AI-Enabled Systems’, p. 10.

[94] Ibid.

[95] Kozlovski, ‘Meaningful Human Control over AI Military Decision Support Systems’, p. 138.

[96] Trabucco, ‘AI-Enabled Autonomous Weapons and Human Control. Part I’, p. 547.

[97] Australian Delegation at the Informal Exchanges on AI in the Military Domain and Its Implications for International Peace and Security, United Nations in Geneva, 15 June 2026. Quote recorded and transcribed by the authors.