Map of Practices

Map of Practices September 2025 AutoPractices Governing AI Technologies in Military Systems from the Bottom Up: Practices to Sustain and Strengthen Human Agency AutoPractices About the Center for WarStudies The Center for War Studies (CWS), established at the University of Southern Denmark (SDU) in 2012, brings together academics from political science, law, history, and cultural studies to contribute to major debates on the past, present, and future of war, as well as its impact on societies. We strive for interdisciplinary research that is relevant to policymakers and the society at large. We aim to contribute to ongoing debates on war and peace by illuminating their multiple dimensions. Through research excellence and societal relevance, we advance the understanding of the fundamental issue of war and peace. For more information about the CWS and its researchers, see https://www.sdu.dk/en/cws. Funding AutoPractices is a Proof of Concept project funded by the European Research Council (grant number 101156237). It runs from June 2024 to December 2025. It is based on the project “Weaponised Artificial Intelligence, Norms and Order” (AutoNorms), which has received funding from the European Research Council under the European Union’s Horizon 2020 research and innovation programme (grant agreement No. 852123). For more information about AutoNorms and AutoPractices, see www.autonorms.eu. © Center for War Studies, September 2025 All rights reserved. No part of this publication may be reproduced, stored or transmitted, in any form or by any means, electronic and mechanical, without the prior written permission of the copyright holder or as explicitly permitted by law. Acknowledgements The AutoPractices team is grateful to all stakeholders who participated in this research. The AutoPractices team Principal Investigator Prof. Ingvild Bode, Professor of International Politics and Director of the Center for War Studies, University of Southern Denmark (SDU) Research Team • Dr. Anna Nadibaidze, Postdoctoral Researcher, SDU • Dr. Alexander Blanchard, Senior Researcher, Stockholm International Peace Research Institute • Shimona Mohan, in her personal capacity • Ariel Conn, in her personal capacity • Dr. Hendrik Huelss, Assistant Professor, SDU • Dr. Qiaochu Zhang, Max Weber Postdoctoral Fellow, European University Institute • Dr. Guangyu Qiao-Franco, Assistant Professor, Radboud University Nijmegen • Dr. Tom Watts, in his personal capacity This map of practices was prepared by Anna Nadibaidze and Ingvild Bode, with input from the AutoPractices team and the stakeholders. How to cite this publication The AutoPractices Project. (2025). Map of Practices (Odense: Center for War Studies). Cover photo by Jerry Chen on Unsplash (unsplash.com/@jerry_chen_) Published by the Center for War Studies University of Southern Denmark, Campusvej 55, Odense M 5230, Denmark Governing AI Technologies in Military Systems from the Bottom Up 1 Contents Introduction to AutoPractices 2 Data collection 3 Overview of stakeholders 4 Terminology 5 Methodology 8 Map of practices 10 Theme 1 AI systems should not replace human decision-making, especially on the use of force 10 Theme 2 Humans must have a possibility to intervene across the lifecycle 12 Theme 3 Feedback loops between end-users and developers should feature across the lifecycle, including after the deployment of an AI system, and should reinforce human agency 13 Theme 4 Training and education measures should be implemented across the lifecycle 13 Theme 5 Humans must continuously engage in political and policy considerations in the development and use of AI systems 14 Theme 6 AI systems should undergo extensive and ongoing testing procedures 15 Theme 7 Practices across the lifecycle need to ensure human accountability and responsibility 16 Theme 8 The use of AI systems should follow appropriate and detailed risk assessment frameworks 17 Conclusion and overview 19 Appendix – Survey questionnaire 21 List of references 23 About the team 24 2AutoPractices Introduction to AutoPractices As of July 2025, there are no international legally binding regulations specific to the development and deployment of artificial intelligence (AI) technologies in the military domain. Groups of states have agreed on sets of non-legally binding principles guiding the military uses of AI, whether in autonomous weapon systems (AWS)1 or other applications.2 However, top-down, state-led approaches to global governance in this area continue to face challenges such as different regulatory positions, competing interests, and diverging visions of the role of AI technologies in warfare and in society more broadly.3 As the European Research Council-funded project “Weaponised AI, Norms, and Order” (AutoNorms) has found, current practices in the design, training personnel for, and use of AI technologies in military systems have the potential to lead to a reduced exercise of human agency in use-of-force decision-making. A reduced form of human agency in military targeting raises ethical, legal, security, and operational concerns which are insufficiently addressed by sets of broad and often ambiguous principles featuring in current top-down frameworks.4 Considering this global challenge, the purpose of the European Research Councilfunded project “Governing AI Technologies in Military Systems from the Bottom Up: Practices to Sustain and Strengthen Human Agency” (AutoPractices) is to initiate and accompany a process of social innovation to govern AI technologies in military systems from the bottom up. The AutoPractices project aims to co-create a set of ‘best practices’ in the form of a practical toolkit to sustain and strengthen the exercise of human agency when developing and using military systems integrating AI and autonomous technologies. The operational toolkit will be cocreated together with stakeholders who represent different professional backgrounds and geographies. This document outlines a map of practices which represents a step on the way towards the final toolkit. It is therefore meant as a transitional document. The AutoPractices project runs from June 2024 until December 2025, and the operational toolkit is to be finalised by December 2025. 1 The 11 guiding principles adopted by the United Nations Group of Governmental Experts on emerging technologies in the area of Lethal Autonomous Weapon Systems (UN GGE on LAWS) in 2019. 2 TheResponsibleAIintheMilitaryDomain(REAIM)SummitsCalltoAction(2023)andBlueprintforAction(2024);theUnited States Political Declaration on Responsible Military Use of AI and Autonomy (2023). 3 Ingvild Bode, Emerging Norms around Military Applications of AI: The Case of Human Control , GC REAIM Expert Policy Note Series (The Hague: The Hague Centre for Strategic Studies, May 2025), https://hcss.nl/wp-content/uploads/2025/05/Bode-1. pdf; Ingvild Bode et al., “Prospects for the Global Governance of Autonomous Weapons: Comparing Chinese, Russian, and US Practices,” Ethics and Information Technology 25,no.5(2023):1–15,https://doi.org/10.1007/s10676-023-09678-x;Anna Nadibaidze, “Governance of AI in the Military Domain: International Law, Norms, and Ways Forward,” in Oxford Intersections: AI in Society ,ed.DovGreenbaum(Oxford:OxfordUniversityPress,2025),https://doi.org/10.1093/9780198945215.003.0102. 4 IngvildBodeandTomWatts, Meaning-Less Human Control: Lessons from Air Defence Systems on Meaningful Human Control for the Debate on AWS (Oxford&Odense:DroneWarsUK&CenterforWarStudies,2021),https://dronewars.net/2021/02/19/ meaning-less-human-control-lessons-from-air-defence-systems-for-lethal-autonomous-weapons/;IngvildBodeandTomWatts, Loitering Munitions and Unpredictability: Autonomy in Weapon Systems and Challenges to Human Control (Odense & London: Center forWarStudies&RoyalHollowayCentreforInternationalSecurity,2023),https://www.autonorms.eu/loitering-munitions-andunpredictability-autonomy-in-weapon-systems-and-challenges-to-human-control/;AnnaNadibaidze,IngvildBode,andQiaochu Zhang, AI in Military Decision-Support Systems: A Review of Developments and Debates (Odense: Center for War Studies, 2024), https://www.autonorms.eu/ai-in-military-decision-support-systems-a-review-of-developments-and-debates/. Governing AI Technologies in Military Systems from the Bottom Up 3 Data collection This map of practices is based on data collected from stakeholders in two ways: 1. An online survey questionnaire completed by stakeholders who were willing to and gave their consent to participate. The survey was conducted via the SurveyMonkey platform (see the appendix for the full questionnaire). a The questions appeared in the same order for all respondents. b All questions were open-ended. c All questions were optional. 2. One-on-one interviews conducted by a member of the AutoPractices team with stakeholders who gave their consent to participate. The final operational toolkit will be based on 1) the survey responses; 2) the interview responses; and 3) the discussions held during two workshops (one virtual on 26 May 2025 and one in-person on 17 June 2025). Figure 1. The main steps in the AutoPractices project process 10/24 – 03/25 Data collection (surveys and interviews) 04/25 Analysis of the survey and interview process 05/25 – 06/25 Workshops with stakeholders to discuss the draft map of practices 12/25 Dissemination of the operational toolkit 07/25 – 11/25 Preparation of the operational toolkit 4AutoPractices Overview of stakeholders Participation in this research is entirely voluntary and consent based. Stakeholders have been selected based on their knowledge and expertise of the integration of autonomous and AI technologies in military systems, as well as AI technologies more broadly. To ensure interdisciplinarity and diversity of perspectives, stakeholders involved in this project include legal experts, military personnel (former and current), civil society representatives, academics, and researchers with different disciplinary backgrounds (humanities as well as social and natural sciences), as well as industry experts. All stakeholders have been invited to contribute in their personal capacity. Their views do not necessarily represent their states or institutions. As of 1 July 2025, the project involves 47 stakeholders. They include politicalethical; legal; military; technical; and civil society experts representing all continents (except Antarctica). The AutoPractices team has aimed to secure a diverse stakeholder representation across these groups and geographical contexts. We refrain from displaying descriptive statistics about how many stakeholders per our own categorisation have been involved. This is to respect the fact that our categorisation may not align with how the stakeholders would categorise themselves. The research ethics committees at the University of Southern Denmark and the European Research Council have reviewed the AutoPractices project. Governing AI Technologies in Military Systems from the Bottom Up 5 Terminology Systems In the AutoPractices survey or interview questions, we did not define ‘systems’ exclusively as either weapon systems, decision-support systems, or other types of systems integrating AI. This was to allow stakeholders to comment based on their background and expertise, no matter what type of systems they are most knowledgeable about. Some stakeholders explicitly mentioned AI-based (autonomous) weapon systems, others named decision-support systems, while others did not specify which systems they meant. Practices and activities In the AutoPractices survey or interview questions, we did not define ‘practices’ or ‘activities’ to allow stakeholders to interpret these terms according to their background and expertise. In the context of the AutoPractices project, we define practices as organised, linked, patterned activities performed by (groups of) people.5 Activities are simply performances of a certain action by a person or group of people. Agency For the purposes of the AutoPractices project, the exercise of human agency in the context of interacting with AI systems is defined as: The capacity to 1) understand and reasonably foresee a system’s functions and effects in a relevant context; and 2) to deliberate and decide upon suitable actions in a timely manner; and 3) to act in a way that can impact the use of the system. This definition is based on a literature review conducted by the AutoPractices team, which revealed the following common elements in the exercise of human agency: • knowledge of the context (situational awareness) and of the system (its technical characteristics, capabilities, limitations), • based on this knowledge, the ability to foresee how the system would function and its potential effects in the context of use, • the ability to reflect/deliberate upon the effects of this system’s use, and following this deliberation, the ability to decide on an action/response and act upon this response (the ability to intervene in a timely manner), and • the ability to engage in actions that make an impact/change in the world. 5 Theodore R. Schatzki, “A Primer on Practices,” in Practice-Based Education: Perspectives and Strategies , by Joy Higgs et al.(Rotterdam:SensePublishers,2012),13,https://doi.org/10.1007/978-94-6209-128-3_2. 6AutoPractices The lifecycle of AI systems The AutoPractices project takes as a foundation the model of the AI lifecycle6 proposed by the IEEE Standards Association Research Group on Issues of Autonomy and AI in Defense Systems.7 The work of the Research Group builds on other lifecycle frameworks that have been used in relation to military applications of AI,8 but is more fine-grained. To date, the IEEE lifecycle framework is the most comprehensive to have been developed with the specific challenges related to the military context in mind, while combining elements of frameworks from civilian domains. It is also the result of a joint, interdisciplinary exercise involving a group of experts with technical, political, ethical, legal, and military backgrounds. A comprehensive lifecycle framework allows considering both 1) a micro perspective of practices of various groups of humans involved at each stage, and 2) a macro perspective of challenges for the exercise of human agency across the different stages. The framework presents a granular way of thinking about the lifecycle with multiple points of human involvement and opportunities to exercise agency. In this model, the lifecycle of a military system integrating AI technologies includes the following 9 stages (see Figure 2): 1) before AI system development 2) research and development 3) procurement and acquisition 4) Test, Evaluation, Validation and Verification (TEVV) 5) considering the human: education, training, and human-system integration 6) political and strategic considerations 7) operational level command and control 8) tactical employment 9) review, reuse and/or retire. 6 Werecognisethatsomestakeholdersmaynotconsidertheterm‘lifecycle’appropriateinthecontextofwarfare. WehavechosentousethisterminthecontextofAutoPracticesbecausethe‘lifecycle’iscommonlyusedintechnical literature describing the development, use, and post-use review of AI systems. 7 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 (New York, NY: IEEE SA, 2024), https://ieeexplore.ieee.org/document/10707139. 8 MerelEkelhofandGiacomoPersiPaoli, The Human Element in Decisions about the Use of Force (Geneva: United NationsInstituteforDisarmamentResearch,2020),https://unidir.org/wp-content/uploads/2023/05/UNIDIR_Iceberg_ SinglePages_web.pdf. Governing AI Technologies in Military Systems from the Bottom Up 7 Figure 2. Lifecycle framework for systems integrating AI and autonomous technologies in the military domain.9 Before development Considering the human Political and strategic considerations Operational level command and control Tactical employment Review, reuse and/or retire Research and development Procurement and acquisition TEVV Moreover, the IEEE Research Group highlights five activities that are ongoing across the nine stages of the lifecycle: 1) evaluation of legal, ethical, and policy concerns 2) responsibility, accountability, and knowledge transfers 3) considering the human: training, education and human-system integration 4) TEVV, monitoring, hardware system or software updates and interoperability, maintenance 5) risk assessment. 9 Based on IEEE, A Framework for Human Decision-Making through the Lifecycle of Autonomous and Intelligent Systems in Defense Applications . 14 AutoPractices There are three broad areas of training measures to ensure the exercise of human agency: 1 Technical education (digital literacy) • Continuously educating humans involved, especially users and operators, on the evolving technical aspects of AI systems so that they understand the technologies involved, how they function, and importantly, their limitations. Being able to reasonably predict and foresee a system’s behaviour is a key part of exercising human agency. • Educating operators on how to critically assess the output of the system, intervene, or potentially stop the system as well as how to document the issue after the employment stage, or in other words, “quickly cease use and report issues when AI systems do not operate as expected” (S#6). 2 Psychological education (understanding the human factors) • Raising the involved humans’ awareness about how they make decisions both without and with technologies. Personnel should know about various aspects of interaction with AI systems such as automation biases and cognitive biases, assumptions made by humans and integrated into AI systems, and risks of de-skilling, among others. • Considering the diversity of humans involved (in terms of gender, educational background, etc.) when studying human-machine interaction. Education and training should also about being prepared to use the whole socio-technical infrastructure involved, not just one AI system. 3 Legal/ethics training • Training humans involved about applicable legal frameworks, especially international humanitarian law, civilian harm mitigation, as well as military ethics, ethical considerations, and evaluation of broader societal impacts. Theme 5 Humans must continuously engage in political and policy considerations during the development and use of AI systems Prior to the development of AI systems, and throughout the lifecycle, relevant policymakers need to ask critical questions about the purpose of the systems, the intended uses, and whether these systems are politically, legally and ethically appropriate for the context of use. Practices applied across the lifecycle include: • Fostering organisational cultures that promote exchange, feedback and communication that allow engaging in those reflections across the lifecycle. • Delimiting who is responsible for activation, use, setting parameters, setting mission goals, and terminating the systems. • Setting specific instructions for all actors to match the political and policy considerations. • Setting limitations or restrictions, whether geographical, spatial, or not targeting humans, for example, based on evaluations of risks for humans affected by the use of AI systems: “consideration should extend to indirect, long term and reverberating effects” such as “human costs beyond physical effects” or the “impact on the natural environment” (S#33). Governing AI Technologies in Military Systems from the Bottom Up 15 • Framing human-machine interaction within the relevant organisational cultures as well as the complexities of how humans interact with each other and how machines interact with each other within those structures: “it is therefore not just a question of human-machine interaction, but also of human-human-machine interaction” and of “machine-machine interaction in the context of human decision-making processes and their embedding in organisational cultures” (S#14). • Incorporating concerns surrounding international security, including the proliferation of technologies, escalation, lowering the threshold for the use of force, or reinforcement of some narratives about AI, for example that AI inevitably increases efficiency and precision. • Mitigating the risks of erosion of culture of accountability and how decisions can cumulatively affect targeting processes and ultimately also targeting doctrines. Practices applied at specific lifecycle stages include: • At stages 2-3 (R&D, procurement and acquisition), integrating any restrictions on spatial, geographical boundaries or types of targets into an appropriate interface: “this is only possible if these measures are conceived of beforehand” (S#27), at the early stages of the lifecycle. • At stage 9 (post-use), setting post-use legal reviews to evaluate whether political but also strategic, legal, and ethical considerations have been met. Theme 6 AI systems should undergo extensive and ongoing testing procedures Testing, evaluation, validation and verification should be an ongoing set of practices throughout the lifecycle. Practices applied across the lifecycle include: • Involving an appropriate dataset and recording limitations or biases within the training data via clear, transparent documentation. • Conducting independent and transparent testing processes that include decision traceability. • Including input and feedback from end-users to match the planned context of use. • Recognising that updates of systems or adaptations of training models may cause problems. Potential updates and adaptation need to be incorporated into measures and risk assessment frameworks. • Implementing regular reviews with metrics assessing the performance and risks in accordance with international legal obligations, especially international humanitarian law and the principles of distinction, proportionality, and precaution. • Adding parameters that would ensure that, if systems do not meet the requirements, new rounds of testing would be required. Similarly, imposing conditions on how long the system could be used without needing to be reviewed or re-tested. • Ensuring transparent access to the training data and parameters, especially for machine learning systems, in a way that allows humans to “obtain explanations on the causal link between AI inputs and outputs, and the functioning of algorithmic processes” (S#33). 16 AutoPractices • Minimising errors that could lead to unforeseen outcomes, for example via technical audits and failsafe mechanisms, while recognising that malfunctions cannot be fully eliminated. • Including recording methods and monitoring processes such as mechanisms that ensure transparency and facilitate access by relevant parties. • Ensuring data security. • Educating human personnel about the safety risks in case systems are hacked or jammed. • Striving for representative datasets and documenting issues with data such as biases to create records for auditing. • Testing systems in ambiguous contexts and running through how actors would behave. Practices applied at specific lifecycle stages include: • At stages 1-3 (before development, R&D, procurement and acquisition), involving human supervisors who can verify the biases in the systems or the data. • At stages 1-5 (before development, R&D, procurement and acquisition, TEVV, considering the human), ensuring understandability and predictability of AI systems. If this requirement is not fulfilled at the design and testing levels, acquisition should be prevented. • At stage 2 (R&D), applying measures “against potential AI-induced harm” and integrating security measures to ensure a “security-by-design approach during R&D” (S#6).10 • At early stages such as stages 2-3 (R&D, procurement and acquisition), detecting technical malfunctions or uncertainties because “if flaws aren’t caught early, those flaws become baked into the system, undermining human decision-making down the line” (S#25). • At stage 4 (TEVV), considering human-machine interaction in testing, not only technical characteristics but also how humans interact with the AI systems. Testing should sufficiently match the situation of employment (stages 7-8, operational and tactical). Many aspects of testing will therefore depend on the planned contexts of use. • At stages 7-8 (operational and tactical), implementing real-time battlefield monitoring and assessment mechanisms to detect malfunctions and recall the system if needed. Theme 7 Practices across the lifecycle need to ensure human accountability and responsibility Ensuring human agency means keeping track of accountable and responsible humans throughout the stages of the lifecycle, given that only human agents/ natural persons can hold legal accountability for violations of international humanitarian law (I#1). Practices applied across the lifecycle include: • Clearly distributing roles among the human actors involved at each stage. • Applying mechanisms tracing decisions to specific actors to ensure transparency. They should enable human operators to have access to data or information that may help the operators understand what led to this output or recommendation. 10 ZhangLinggaveconsenttobeacknowledgedbynamefordirectquotesfromS#6. Governing AI Technologies in Military Systems from the Bottom Up 17 • Maintaining consistent records and documentation to ensure transparency of how responsible individuals or teams use tools for certain functions. At the same time, this should not involve extensive surveillance, but rather “the security of a team built of trust and excellent communication” (S#8). • Allocating enough resources into building these mechanisms and planning this resource allocation in advance, at the early stages of the lifecycle. • Developing a concept of operations and match these concepts with the computational components so that, at the employment stage, “responsible people can have relied on the practices that went before them” (I#1). • Adopting a verification regime. Practices applied at specific lifecycle stages include: • At stages 1-2 (before development, R&D), incorporating legal guidance (especially on international humanitarian law) and defining accountability measures since the beginning. • At stage 2 (R&D), implementing safeguards or measures to mitigate risks, for example, via agile co-design (S#8). • At stages 7-8 (operational and tactical), ensuring a continuous connection between the systems and human operators. • At stage 9 (post-use), after the use of AI systems and in case of violations of international humanitarian law or other legal frameworks, engaging in a process to apply liability and ensure “judicial agency” (S#14) via “strong after-action mechanisms” (S#27). Theme 8 The use of AI systems should follow appropriate and detailed risk assessment frameworks Actors should implement frameworks to assess risks of using AI systems in the military domain. The type of framework, however, may depend on the context and the systems. Practices applied across the lifecycle include: • Considering the differences and distinctive features of types of conflicts when assessing the risks and necessary frameworks/measures. • Considering the contexts of use, as “specific frameworks will be necessary for high-risk/impact use cases, such as AI-enabled weapons (including AWS), AI in decision support (especially related to use of force), AI in cyber, AI in information operations, among others… this is where the most impactful work could and should be done, especially given all of these use cases are current concerns and in operational use” (S#23). • Engaging in exercises together with other states or organisations to identify best practices and benchmarks for AI systems to meet objectives while assessing risks. • Adopting appropriate (and continuously updated) cybersecurity measures because in case of a cyber-attack, “the performance of the system can utterly change from one day to the next” (S#8).11 • Ensuring that AI systems are interpretable, and that the parameters and weights are used according to broader legal, ethical and strategic considerations. • When it comes to weapon systems, improving traceability and transparency by marking hardware and recording data about operators and tasks. 11 JoannaBrysongaveconsenttobeacknowledgedbynamefordirectquotesfromS#8. 18 AutoPractices • Classifying risks in a tier system (e.g., untenable, high-level, mid-level and low-level), as well as deciding on safeguards to prevent and mitigate different sets of risks. • Classifying how the AI systems relate to the use of force (especially relevant for AI in decision-support systems). Practices applied at specific lifecycle stages include: • At stages 1-2 (before development, R&D), adopting standardised operating procedures (SOPs) across the lifecycle, especially at earlier stages when speed is not such a prominent issue. • At stages 2-3 (R&D, procurement and acquisition), approaching and building an awareness of risks at the earliest phases of the lifecycle as well as adopting a framework to minimise risks during stage 4 (TEVV). • At stages 1-4 (before development, R&D, procurement and acquisition, TEVV), adopting safeguards for online learning. • At stage 9 (post-use), including risk assessments in the form of technological audits or legal reviews conducted after the use of the systems. After-action mechanisms can ensure legal accountability as well as reveal technical malfunctions and vulnerabilities. Risk assessment frameworks that stakeholders consider helpful and important, although these frameworks might currently not be applied to the military domain, include: • The European Union’s AI Act • The US National Institute of Standards and Technology (NIST) • The International Organization for Standardization (ISO), e.g., ISO/IEC 27001, ISO/IEC 23894 • Civil aviation industry standards and the International Civil Aviation Organization (ICAO) • IEEE SA Framework or other IEEE standards, e.g., IEEE 7007-2001 • The International Atomic Energy Agency (IAEA) • The Chemical Weapons Convention • The Organization for Economic Cooperation and Development (OECD) guidelines on AI ethics • The NATO AI Strategy • The NATO Responsible AI toolkit • The US Responsible AI toolkit • The REAIM Summit’s Blueprint for Action • The Australian Voluntary AI Safety Standard • The Trusted Autonomous Systems (Australia) Responsible AI For Defence Toolkit (Consultation) • The UNIDIR taxonomy of risks • The UK Ministry of Defence Dependable Artificial Intelligence (AI) in Defence Directive (JSP 936 V1.1) • Responsible AI principles • Legal reviews of advanced cyber capabilities • The 3D (design, development, deployment) framework • Checklists or flowcharts • National or international confidence building measures • Self-regulation schemes Governing AI Technologies in Military Systems from the Bottom Up 19 Conclusion and overview This map of practices is the foundation for AutoPractices’ work towards the main objective of the project: co-creating, with the stakeholders involved, a practical toolkit to sustain and strengthen the exercise of human agency in the use of military AI systems. The map highlights the diversity of activities that actors, including political and strategic decision-makers, developers, engineers, commanders, operators, lawyers, ethicists, and others can perform to contribute to the exercise of human agency both across the lifecycle of AI systems and at specific stages. The practices mentioned across the eight themes identified above can be grouped into three broad categories (see Table 1, based on terminology in S#3). First, there are technical practices that relate to aspects such as hardware, software, data, and cybersecurity. Second, there are policy practices that relate to establishing operational norms and constraints (including, but not limited to, legal) on how AI systems should be developed and used. Third, there are procedural practices that include adapting organisational cultures, concepts of operations, rules of engagement, special instructions to mitigate risks specifically associated with uses of AI systems. These sets of practices contribute to ensuring the exercise of human agency in human-machine interaction within the military context, especially in decisionmaking on the use of force. The AutoPractices operational toolkit will expand further on some of these practices and their contribution to the exercise of human agency in the military domain. Table 1. General overview of practices Technical practices Policy practices Procedural practices Human-machine interaction Hardware Software Learning techniques Cybersecurity measures Data security Testing and evaluation Design Monitoring Restrictions on use of systems Legal reviews Technological audits Feedback loops and communication mechanisms Organisational culture Concept of operations Rules of engagement Special instructions Education and training 20 AutoPractices Appendix – Survey questionnaire Dear Participant, Thank you for agreeing to take part in this survey. The results of this questionnaire will feed into the project “Governing AI Technologies in Military Systems from the Bottom Up: Practices to Sustain and Strengthen Human Agency” (AutoPractices), funded by the European Research Council (Proof of Concept grant no. 101156237). The purpose of the AutoPractices project is to initiate and accompany a process of social innovation to govern autonomous and AI technologies in the military domain. The project will co-create a set of best practices with stakeholders in the form of a practical toolkit to sustain and strengthen human agency and accountability for the use of AI systems in the military. To co-create this practical toolkit, we approach stakeholders across diverse professional backgrounds and geographies. You have been selected as a research participant based on your knowledge and expertise on AI systems, including in the military domain. Participation is voluntary, and you are free to decline to answer some questions and can leave the data collection activity at any time without giving a reason. The research team will keep your data strictly confidential and anonymised. This questionnaire focuses on your understanding (based on your respective area of expertise) of how human agency should be exercised throughout the lifecycle of systems integrating AI technologies. For the purposes of this survey, the exercise of human agency in the context of interacting with an AI system is defined as: The capacity to 1) understand and reasonably foresee a system’s functions and effects in a relevant context; and 2) to deliberate and decide upon suitable actions in a timely manner; and 3) to act in a way that can impact the use of the system. The lifecycle of a military system integrating AI includes the following stages: 1) before AI system development; 2) research & development; 3) procurement & acquisition; 4) Test, Evaluation, Validation and Verification (TEVV); 5) education & training; 6) political and strategic considerations; 7) operational level command and control; 8) tactical employment; and 9) review, reuse and/or retire. Governing AI Technologies in Military Systems from the Bottom Up 21 Questions Q1. What do you think ensuring the exercise of human agency across the lifecycle of a military system integrating AI means? Please respond in max. 1-2 sentences. Q2. Based on your area of expertise, what are the most important concerns to address at various stages of a system’s lifecycle (listed above) to ensure the exercise of human agency? Please name your top 3 concerns. Q3. What do you consider as key activities contributing to ensuring the exercise of human agency when it comes to human-system integration/human-machine interaction throughout an AI system’s lifecycle? Please name your top 3 activities. Q4. What types of activities would be detrimental to the exercise of human agency across an AI system’s lifecycle? Please name your top 3 activities that you find detrimental. Q5. What do you consider key activities to ensure responsibility and accountability throughout an AI system’s lifecycle? Please name your top 3 activities. Q6. What are some of the key activities needed to ensure the exercise of human agency in Testing, Evaluation, Validation and Verification (TEVV), monitoring, and maintenance? Please name your top 3 activities. Q7. Based on your area of expertise, do you consider certain training and education measures essential to ensure the exercise of human agency across a system’s lifecycle? If yes, which ones? Please name a maximum of 3 measures. Q8. How should risks related to systems integrating AI technologies be assessed? Are there any existing frameworks which you would consider useful or important? 22 AutoPractices Governing AI Technologies in Military Systems from the Bottom Up 23 List of references Bode, Ingvild, Hendrik Huelss, Anna Nadibaidze, Guangyu Qiao-Franco, and Tom Watts. 2023. “Prospects for the Global Governance of Autonomous Weapons: Comparing Chinese, Russian, and US Practices.” Ethics and Information Technology 25 (5): 1–15. https://doi.org/10.1007/s10676-02309678-x. Bode, Ingvild, and Tom Watts. 2021. Meaning-Less Human Control: Lessons from Air Defence Systems on Meaningful Human Control for the Debate on AWS. Oxford & Odense: Drone Wars UK & Center for War Studies. https://dronewars.net/2021/02/19/meaning-less-human-control-lessonsfrom-air-defence-systems-for-lethal-autonomous-weapons/. Bode, Ingvild, and Tom Watts. 2023. Loitering Munitions and Unpredictability: Autonomy in Weapon Systems and Challenges to Human Control. Odense & London: Center for War Studies & Royal Holloway Centre for International Security. https://www.autonorms.eu/loitering-munitions-and-unpredictabilityautonomy-in-weapon-systems-and-challenges-to-human-control/. Ekelhof, Merel, and Giacomo Persi Paoli. 2020. The Human Element in Decisions about the Use of Force. Geneva: United Nations Institute for Disarmament Research. https://unidir.org/wp-content/uploads/2023/05/UNIDIR_Iceberg_ SinglePages_web.pdf. IEEE SA Research Group on Issues of Autonomy and AI in Defense Systems. 2024. A Framework for Human Decision-Making through the Lifecycle of Autonomous and Intelligent Systems in Defense Applications. New York, NY: IEEE SA. https://ieeexplore.ieee.org/document/10707139. Nadibaidze, Anna. 2025. “Governance of AI in the Military Domain: International Law, Norms, and Ways Forward.” In Oxford Intersections: AI in Society, edited by Dov Greenbaum. Oxford: Oxford University Press. https://doi.org/10.1093/978 0198945215.003.0102. Nadibaidze, Anna, Ingvild Bode, and Qiaochu Zhang. 2024. AI in Military Decision Support Systems: A Review of Developments and Debates. Odense: Center for War Studies. https://www.autonorms.eu/ai-in-military-decision-supportsystems-a-review-of-developments-and-debates/. Schatzki, Theodore R. 2012. “A Primer on Practices.” In Practice-Based Education: Perspectives and Strategies, by Joy Higgs, Ronald Barnett, Stephen Billett, Maggie Hutchings, and Franziska Trede, 13–26. Rotterdam: Sense Publishers. https://doi.org/10.1007/978-94-6209-128-3_2.