Towards Multilateral Policy on Autonomous Weapon Systems

Executive summary

Autonomous weapon systems (AWS)—weapons capable of selecting and applying force to targets without human intervention—raise profound legal, ethical and security questions. Over a decade of multilateral deliberations on AWS has yielded limited progress, with states divided on definitions of AWS, regulatory approaches and pathways for action. The resulting landscape is one of institutional complexity, political sensitivity and growing urgency.

This report examines possible directions for multilateral policy on AWS. Drawing on confidential interviews with state representatives from all United Nations regional groups and with non-governmental organizations, as well as analysis of official documents and processes, it maps the evolution of international discussions, identifies the key factors shaping states’ positions, and evaluates foreseeable pathways. The report does not advocate a single solution; instead, it provides policymakers with a structured overview of options, trade-offs and feasibility considerations, to assist them to advance multilateral efforts on AWS.

The report makes three findings:

1. States are weighing nuanced factors when developing their respective national positions on policy directions on AWS. However, these nuances are rarely expressed in the multilateral debate. The result is that the multilateral debate may appear (falsely) to involve only a dichotomous choice between whether or not to negotiate a legally binding instrument on AWS.
2. States share some common ground with respect to AWS policy directions but progress towards outputs will require more political will and leadership.
3. States may be missing an opportunity to align AWS policy discussions with broader discussions on military applications of artificial intelligence (AI).

Looking ahead, the end of 2026 represents a critical juncture: the expiry of the current mandate of the Convention on Certain Conventional Weapons (CCW) Group of Governmental Experts on Emerging Technologies in the Area of Lethal Autonomous Weapons Systems (GGE on LAWS), the convening of the CCW Review Conference, and parallel developments in the UN General Assembly and plurilateral initiatives on military AI. States will need to decide whether to extend, adapt or replace existing mechanisms, and how to balance the pursuit of legally binding rules, political commitments and new governance frameworks.

The report makes three recommendations:

1. States should ensure that their national policy positions are holistic, taking into account the range of possible policy directions. States should be as transparent as possible about their views on the various policy directions.
2. States wishing to pursue an instrument on AWS—no matter their views on its status, forum, procedure or timing—should build on the rolling text of the CCW GGE on LAWS.
3. States’ policy decisions with respect to AWS should take into account the common risks and challenges posed by the use of AI in decisions to use force.

In sum, multilateral policy on AWS remains a test of the international community’s ability to collectively govern emerging technologies of warfare. While the CCW retains pre-eminence, new processes and hybrid approaches are gaining traction. The choices made in the coming years will shape not only the governance of AWS, but also the broader trajectory of military AI regulation.

1. Introduction

Autonomous weapon systems (AWS)—weapons capable of selecting and applying force to targets without human intervention—have put into question both the capacity and willingness of states to collectively govern the means of violence.¹ Over a decade of deliberations have yielded a policy landscape marked by slow progress, persistent divergence and growing urgency. This results from the challenges states face in conceptualizing their concerns around AWS and the use of technology in decisions to apply force; limitations in the processes and tools available for multilateral efforts; and eroded trust among states.²

This report explores what states can do to advance multilateral efforts on AWS. It clarifies and assesses the different policy directions (see box 1.1)—continuing the status quo, converging related processes, initiating new ones—and outlines the trade-offs, implications and feasibility considerations attached to each. The goal is to equip policymakers with a structured and realistic overview of possible next steps, grounded in the perspectives of key participants and a clear understanding of the institutional terrain.

Methodology

The report uses a mixed-methods approach. Primary data was collected through confidential interviews with state and non-state actors involved in multilateral discussions on AWS. Representatives from all United Nations General Assembly regional groupings—Africa, Asia-Pacific, Eastern Europe, Latin America and Caribbean, and Western and other states—took part in interviews.³ Data from policy documents, statements and past reports supplemented these interviews.

Structure of the report

Chapter 2 maps the evolution since 2010 of the multilateral policy ecosystem on AWS, highlighting recurring themes and the growing institutional complexity of the policy process. Chapter 3 unpacks the key variables that states consider in weighing policy directions, including scope, timing, legitimacy and impact. Chapter 4 outlines and evaluates foreseeable policy pathways in light of these variables. The final chapter offers key findings and recommendations to help stakeholders chart a path forward.

2. A history of multilateral policy efforts on AWS

States consider and discuss options to address AWS within a structure of multilateral policy discussions and plurilateral initiatives that has developed over time. This existing structure shapes the options that states consider open to them and how they weigh their implications and trade-offs.

This chapter charts the development of this structure by summarizing the policy discussions on AWS since 2010 (see figure 2.1): from the first report of the UN special rapporteur on extrajudicial, summary or arbitrary executions on this topic; to the work of UN member states since 2013 under the auspices of the Convention on Certain Conventional Weapons (CCW); to the more recent addition of AWS to the agenda of the UN General Assembly in 2022; and, finally, to the additional forums in which states are discussing AWS and related technologies, both within the UN and outside it. This potted history underlines the conceptual, substantive and procedural difficulties that states face as they try to address the challenges and risks posed by AWS.

https://dev.sipri.org/sites/default/files/Figure%201_0.png

Figure 2.1. A selected history of developments in multilateral policy efforts on autonomous weapon systems

AWS = autonomous weapon systems; CCW = Convention on Certain Conventional Weapons; GGE = Group of Governmental Experts; IHL = international humanitarian law; LAWS = lethal autonomous weapons systems; UNGA = United Nations General Assembly.

Special rapporteurs and civil society are the first to publicly raise concerns about AWS

Between 2010 and 2013, UN special rapporteurs and civil society organizations voiced early warnings about the development of AWS and their implications for compliance with international humanitarian law (IHL) and international human rights law (IHRL). In 2010 the special rapporteur on extrajudicial, summary or arbitrary executions, Philip Alston, submitted an interim report to the UN General Assembly on the relevance of new technologies, such as ‘lethal robotic technology’, to human rights fact-finding, targeted killings and extrajudicial executions.⁴ Alston’s recommendations focused on safety standards, testing of reliability and performance, and investigations and accountability.⁵

In 2013 the next special rapporteur on extrajudicial, summary or arbitrary executions, Christoph Heyns, voiced a more urgent need for action: ‘If left too long to its own devices, the matter will, quite literally, be taken out of human hands.’⁶ He recommended immediate steps, including national moratoriums on lethal autonomous robotics, national measures to ensure compliance with IHL and IHRL, and transparency in legal reviews of robotic systems.⁷ He also recommended that the High Commissioner for Human Rights convene ‘a High Level Panel on [lethal autonomous robotics] consisting of experts from different fields such as law, robotics, computer science, military operations, diplomacy, conflict management, ethics and philosophy’, and for the panel to, within a year, propose ‘a framework to enable the international community to address effectively the legal and policy issues’.⁸

In 2012 Human Rights Watch (HRW) and the Harvard Law School International Human Rights Clinic (IHRC), having anticipated Heyns’s report, called for an international treaty to prohibit the development, production and use of ‘fully autonomous weapons’, namely robots that are capable of selecting targets and delivering force without any human input or interaction, or where the human supervision is limited.⁹ The HRW/IHRC report considered that the ‘escalating use of aerial drones’ and the use of ‘automatic weapons defense systems’, such as close-in weapon systems and sentry systems, presaged future reliance on more autonomous systems.¹⁰

The reports by successive special rapporteurs on extrajudicial, summary or arbitrary executions, and by HRW/IHRC, contain themes that continue to influence multilateral policy discussions on AWS today. These include concerns regarding compliance with IHL and IHRL, as well as transparency and accountability; calls for prohibitions at the international and national level; and promotion of testing, legal reviews and standards. These calls for action were based on predicted developments in technology because the AWS that existed at the time had not raised significant legal or ethical concerns, particularly among the public. A lack of public outrage may explain why urgent calls for regulation between 2010 and 2013 were not met with an equal level of urgency in the response. As discussed below, a view that (fully) AWS were ‘future weapons’ influenced the shape of states’ response to Heyns’s 2013 report.¹¹

Formal discussions among states commence in Geneva

Following the publication of Heyns’s report, and at the suggestion of Brazil and France, the High Contracting Parties (HCPs) to the CCW decided to convene a meeting of experts to discuss ‘the questions related to emerging technologies in the area of lethal autonomous weapons systems’.¹² Experts met and adopted reports between 2014 and 2016. At the Fifth Review Conference of the CCW, the HCPs decided to establish an open-ended group of governmental experts (GGE) to build on the work of the meetings of experts and to explore and agree on possible recommendations for options ‘related to emerging technologies in the area of lethal autonomous weapons systems [LAWS], in the context of the objectives and purposes of the Convention’.¹³ The GGE on LAWS is open to all CCW HCPs and signatories, as well as other states, international organizations and non-governmental organizations.

Adoption of guiding principles

The work of the GGE on LAWS began by addressing four broad themes: (a) characterization of AWS; (b) potential legal and ethical challenges posed by AWS; (c) military implications of AWS and related technologies; and (d) aspects of human–machine interaction in the development and use of AWS. In 2019 the GGE adopted 11 ‘guiding principles’ as a basis for its future work. These guiding principles affirmed, among other things, that IHL continues to apply to all weapon systems including AWS; that human responsibility for decisions on the use of weapon systems must be retained because accountability cannot be transferred to machines; and that human–machine interaction should ensure that the use of AWS complies with applicable international law.¹⁴ The GGE adopted these principles without prejudice to the possible policy options for addressing challenges posed by AWS, including a legally binding instrument, a political declaration, and clarity on the implementation of existing obligations under international law, in particular IHL.¹⁵ On this question, no consensus was reached.

Steps towards a normative and operational framework

Tasked with clarifying and developing ‘the normative and operational framework’, the GGE canvassed a wide range of issues without agreeing on substantive outcomes. The GGE chair, Ambassador Amandeep Singh Gill of India, grouped proposals related to the normative and operational framework into three categories: (a) a legally binding instrument (the subject of a number of proposals); (b) a less-than-treaty status instrument (such as proposals for a political declaration); and (c) clarification of existing international law obligations (such as agreed ‘principles and good practices’, or a document on the application of IHL to AWS).¹⁶

While states expressed different views on the form by which the normative and operational framework should be developed, convergence emerged on the approach for such development. During this period a growing number of states began (and continue) to express support for a so-called ‘two-tiered’ approach. This approach—common in arms control—refers to a structure that distinguishes between, in the first tier, types and uses of AWS that would be unlawful in all circumstances and, in the second tier, limits and requirements on the development and use of all other AWS.¹⁷ This approach is relevant to instruments describing international law as it currently is, as well as instruments that create new rules under international law.

The ‘rolling text’

Since 2024 the GGE has worked on a ‘rolling text’, a term the current chair, Robert in den Bosch of the Netherlands, uses to refer to a document that aims to capture elements that have attracted provisional consensus and is the subject of repeated readings throughout GGE meetings, which might become an instrument but ‘without prejudging its nature, and other possible measures’.¹⁸ This rolling text covers characterization of ‘lethal’ AWS (including the meaning of ‘lethal’); the application and interpretation of IHL (including the CCW itself, see box 2.1) with respect to AWS; prohibitions and regulations regarding AWS necessary for compliance with IHL; measures to be taken prior to, during and after the use of AWS; and matters related to individual and state responsibility and accountability for the development and use of AWS. According to 39 of the HCPs (and three observer states), the rolling text ‘is a sufficient basis to fulfil the mandate of this GGE in its current [May 2025] form’.¹⁹

Additional efforts in New York and in regional groupings

Since 2022, states have also discussed AWS in meetings in New York of the First Committee of the UN General Assembly, dealing with disarmament and international security. To date, these discussions have emphasized the need for urgent action to address the challenges posed by AWS and have considered challenges not addressed by the CCW GGE on LAWS, such as compliance with IHRL and risks related to stability (e.g. unintended escalation) and security (e.g. proliferation). This section outlines the developments in New York and the relationship between these efforts and those in the GGE.

In 2022 a group of 70 states led by Austria delivered a joint statement calling on UN member states to ‘intensify consideration’ of AWS and to focus efforts on ‘elaborating the normative and operational framework regulating, where appropriate and necessary, autonomous weapons including through internationally agreed rules and limits’.²⁰ In 2023 Austria led an effort to adopt a resolution requesting the UN secretary-general to seek the views of member states, observer states and others on ways to address the challenges and concerns raised about AWS.²¹

The secretary-general, in his report published in 2024, noted the ‘widespread recognition . . . and concern’ about AWS.²² Indeed, AWS had been discussed in a number of regional meetings during 2023 and 2024.²³ The secretary-general observed that ‘time is running out’ and reiterated his call for the conclusion, by 2026, of a legally binding instrument to prohibit AWS that function without human control or oversight and that cannot be used in compliance with IHL, and to regulate all other types of AWS.²⁴ While the secretary-general noted and encouraged the ongoing work of the CCW GGE on LAWS, he also encouraged the General Assembly to continue its consideration of the matter.²⁵

In 2024 the General Assembly adopted a follow-up resolution that invited the GGE to take into account the secretary-general’s report and decided to convene informal consultations to consider the report ‘in full complementarity with and in a manner that supports the fulfilment of the mandate’ of the GGE.²⁶ These informal consultations were held in May 2025 with the aim of updating New York–based delegates on the work of the GGE and furthering the international community’s understanding of AWS by discussing the content of the secretary-general’s report, ‘including the various proposals from States and other matters raised that may thus far not been discussed in detail in the GGE [on] LAWS’.²⁷

Elsewhere in the UN

States have continued to address the theme of AWS in UN human rights and disarmament forums other than the CCW and the First Committee, though the relationships between the discussions in these forums is not yet apparent.

Human rights forums

The Human Rights Council (HRC) in 2019 tasked its Advisory Committee to prepare a report on the ‘Possible impacts, opportunities and challenges of new and emerging digital technologies with regard to the promotion and protection of human rights’.²⁸ The focus of the resulting 2021 report was on civilian applications of new technologies, but the Advisory Committee noted that ‘various salient human rights concerns surround [the use of AWS] in armed conflict’.²⁹ In 2022 the HRC requested the Advisory Committee to prepare a study examining the human rights implications of new and emerging technologies in the military domain, to be presented at the HRC’s 60th session in 2025.³⁰

The UN special rapporteur on extrajudicial, summary or arbitrary executions also returned to the theme of AWS in 2024, with Morris Tidball-Binz recommending that, in addition to continuing work within the GGE on LAWS, states address the human rights implications of AWS, namely by: identifying measures to ensure that attribution of, and accountability for, all uses of AWS is possible; monitoring AWS use; implementing prohibitions on certain anti-personnel uses of AWS (including through both voluntary measures and the adoption of a legally binding instrument); and refraining from domestic use of AWS.³¹

Disarmament forums

The UN Disarmament Commission established a working group to address ‘Recommendations on common understandings related to emerging technologies in the context of international security’, in which states have discussed technologies including AWS and artificial intelligence (AI).³² In 2023 Pakistan proposed that the Conference on Disarmament (CD) continue deliberations on the security aspects of the ‘development, deployment, integration and use of AI for military purposes and autonomous weapon systems’.³³ These security aspects are seen by most states as outside the remit of the CCW GGE on LAWS. Belarus and Japan had already in 2021 drawn a connection between AWS and the CD’s agenda item on ‘new weapons of mass destruction and new systems of such weapons; radiological weapons’.³⁴

Discussions of AWS in forums related to AI in the military domain

The policy discussions on AWS and those on military AI have a shared conceptual focus. Like AWS, the military AI applications of most interest and concern, such as the integration of AI in weapons systems and systems that contribute to offensive capabilities, impact the role of humans in decisions to use force.³⁵ This conceptual similarity stands in contrast to the categorical scope of the GGE on LAWS discussions, which is limited to emerging technologies in the area of lethal AWS. Despite this, states have introduced into the GGE themes related to AI, such as algorithmic bias.³⁶

The conceptual similarity is also relevant to how states have structured discussions about governance of military AI within the UN. In 2024 the UN General Assembly adopted a resolution on AI in the military domain, sponsored by the Netherlands and the Republic of Korea (South Korea), these two countries also jointly initiated the Summit on Responsible AI in the Military Domain (REAIM).³⁷ This resolution focused on the implications for peace and security of increasing integration of AI in military applications, including weapons. The resolution acknowledged the importance of the work of the GGE on LAWS and the ‘need to ensure complementarity between discussions in this regard and in discussions on the broader security implications of artificial intelligence in the military domain’.³⁸ According to interviewees, states that wish to see progress on a legally binding instrument appear to have adopted a deliberate strategy aimed at preserving buy-in from a small group of mostly Western states that view the CCW as the only forum in which discussions of AWS can take place, while continuing to deepen discussions of a broader range of challenges associated with AWS. The implications and trade-offs of the discussion of AWS in the General Assembly are explored further in chapter 4.

The UN Security Council debated the use of AI in conflicts in December 2024, when the UN secretary-general warned that AI threatens human control over weapon systems, and some council members called for regulation on autonomous weapons.³⁹

States have also discussed AWS in plurilateral forums that are focused on AI in the military domain. These initiatives include the REAIM summit, the related REAIM Global Commission established by the Netherlands following the first REAIM summit, the United States–led meetings of states endorsing the ‘Political declaration on responsible military use of artificial intelligence and autonomy’ (US political declaration), and the discussions on maintaining human control in AI-enabled weapon systems under the auspices of the AI Action Summit hosted by France in 2025.⁴⁰

Though it is not yet possible to assess the effect of these initiatives, it is already apparent that each has contributed to an increased level of awareness among policymakers of the topic of AI in the military domain, including AWS. However, their outcomes remain general and high-level, lacking the level of detail required for normative effect.⁴¹ For example, in the REAIM Blueprint for Action, adopted in 2024, stakeholders affirmed that AI applications must be developed and used in accordance with international law, but committed only to ‘engaging in further discussions and to promoting dialogue’ on developing measures to ensure responsible AI in the military domain at the national, regional and international level.⁴² These initiatives have also been criticized for not being inclusive of all states. For example, Russia was not invited to the 2023 and 2024 REAIM summits, and the US political declaration process is open only to endorsers of the declaration.

An inflection point in 2026

Many participants in the AWS policy discussion perceive that the end of 2026 will be an inflection point. This is because the CCW GGE on LAWS is required to complete its work under its current mandate by the end of 2026, with the chair to submit a report for consideration at the Seventh Review Conference of the CCW, in late 2026.⁴³ At that conference, the HCPs will likely debate the future work of the GGE (if it continues at all), and indeed the future of efforts to develop an international treaty on AWS (such as the adoption of a negotiation mandate).

This debate on future international regulation of AWS will likely take place against the background of further resolutions in the UN General Assembly on AWS and on AI in the military domain. The next resolution on AWS will follow up on the informal consultations of May 2025, while the next resolution on military AI will likely respond to the report of the UN secretary-general following the submission of views by states and others.⁴⁴ In addition, states will have received the report of the HRC Advisory Committee, as well as the REAIM Global Commission, an independent body tasked by the REAIM Summit to produce ‘a strategic guidance report which identifies short and long-term recommendations for governments and the wider multi-stakeholder community on responsible AI in the military domain’.⁴⁵

Conclusion

Themes that continue to dominate states’ discussions on AWS in 2025 have been present since 2010, notwithstanding the rapid (though not altogether unprecedented) technological developments related to AWS in the intervening years. These recurring themes include: the urgent need for action to maintain the human element in decisions to use force; the centrality of compliance with existing international law including IHL; and the importance of practical measures such as standards for developers, testing of reliability, legal reviews and measures to ensure accountability for violations.

For now, opportunities to discuss policy responses to AWS outside Geneva have been crafted deliberately as modest and complementary to the CCW GGE on LAWS. Similarly, plurilateral initiatives and initiatives in the UN relating to the use of AI in the military domain have contributed to an increased awareness of the relevant policy issues, but the direction of these policy discussions and their impact on the GGE is not yet apparent.

The end of 2026 is viewed as a possible inflection point where states will be required to decide on the future of efforts to regulate AWS. Chapter 3 describes the factors that states are weighing up in considering these future efforts. Chapter 4 unpacks the options for future policy directions and how they relate to these factors.

3. Weighing up policy directions

Policymakers are weighing up a range of factors in developing national positions on AWS. This chapter describes eight such factors, drawing on confidential interviews and open-source literature. Although not all states’ representatives describe each factor in the same way, and not all states are weighing all factors, the aim is to capture, in a nuanced way, how states are approaching multilateral policy approaches to the governance of AWS and unpack how these factors influence assessments of ‘feasible’ policy options.

The eight factors relate to: the scope of technology and potential policy approaches; the impact of policies in the world; the timing of next policy steps; the coherence between broader governance frameworks; whether policy efforts and outcomes should be assessed wholesale (i.e. on their own merits) or as part of a path of incremental change; the influence of geopolitics and national contexts; issues of legitimacy of a process; and the breadth and depth of participation in a process.

Scope of obligation

Interviewees highlighted that the scope of the obligation imposed by a possible policy output is an important factor for weighing up future policy directions. Scope of obligation entails both breadth and depth. Breadth of obligation relates to the range of issues encompassed by an output, including human rights and security considerations. Depth of obligation is concerned with both the granularity of output—the level of detail—and the restrictiveness of the obligation—how much it prohibits. For instance, a legally binding instrument is understood to give rise to more meaningful obligations than an output that is not legally binding. Depth of obligation can also relate to the substance of an output, including whether a legally binding instrument offers only a restatement of IHL or contains novel substance.

Type and degree of impact

Another important factor for interviewees is the expected impact of policy outcomes. Impact is conceived primarily as the effect of an outcome on state behaviour, but also includes effects on civilians and accountability for violations of international law. This consideration influences how interviewees view both the form and the content of potential policy outputs.

For example, some states favour the negotiation and adoption of a legally binding instrument on AWS because, in their view, such an instrument is likely to have greatest impact on state behaviour. Others have in mind potential impact when highlighting the importance of support for any policy output from states currently using or likely to use AWS, arguing that any output (including a legally binding instrument) would be less impactful without their support. Along these lines, some interviewees pointed out that the GGE’s discussions are already having an impact in two areas—national positions on the development, acquisition and use of AWS (and therefore presumably on the behaviour of national militaries) and the development of some common understandings of how IHL applies to the development and use of AWS—and that this should be borne in mind when assessing the success of the GGE and its next steps.

Timing of action

Precisely what constitutes ‘good timing’ for taking next policy steps on AWS is subjective and disputed. Interviewees gave weight to different factors in determining the ‘right moment’ for action. For some, good timing is related to developments in the technologies as well as in understanding about possible use-cases. There were concerns that moving to the adoption of a legally binding instrument too soon, before consequential or concerning use-cases of AWS occur, may lead to ineffective regulation that fails to adequately address those consequences or concerns. Some expressed a similar view that, settled too quickly, regulation on AWS may fail to capture the likeliest and most common use-cases—for example, the use of AWS for close-in weapon support could be more common than anti-personnel uses.

Some interviewees used the 2008 Convention on Cluster Munitions to demonstrate that progress on arms control sometimes requires a significant humanitarian incident.⁴⁶ Israel’s use of cluster munitions in its conflict with Hezbollah in Lebanon in 2006, which the UN Mine Action Coordination Centre of South Lebanon said exacted a huge human and economic toll, is considered to have spurred progress towards adoption of the Convention.⁴⁷ Some interviewees considered that only once a clear, high-profile humanitarian incident entailing AWS had occurred could normative development be achieved, particularly within the GGE on LAWS.

Aside from when policy steps should be taken, interviewees also indicated that they are weighing the duration of developing policy outputs and achieving policy goals. Many interviewees expect, and are prepared for, the AWS policy process to continue for some years yet—some estimate another decade for the negotiation and adoption of a legally binding instrument. Many indicated the necessarily incremental nature of developing policy in the international domain and that normative change takes time.

Across the range of views expressed there was one point of convergence: the wrong moment to commence or decide the future policy direction is prior to the end of 2026, when the current mandate of the GGE on LAWS expires. For many, the risk of prejudging the outcome of the GGE provides the greatest deterrent against action.

Coherence across process and outputs

In weighing up policy directions, states are considering coherence across processes and outputs. Coherence has two aspects. The first aspect means ensuring and maintaining a clear and consistent relationship between different processes and outputs, both on AWS and military AI. The importance of this aspect of coherence is, foremost, to ensure that the total combination of international policy efforts on AWS and military AI are effective and do not result in inconsistent governance or regulatory regimes that would undermine those efforts. For some states, the importance of this aspect of coherence is a motivation to prioritize one policy process or forum, such as the GGE on LAWS, over others, to provide a benchmark against which to assess the consistency of other processes.

In this respect, the UN General Assembly resolution on AI in the military domain acknowledges ‘the need to ensure complementarity between discussions [at the GGE on LAWS] and discussions on the broader security implications of artificial intelligence in the military domain’.⁴⁸ However, our research shows that state representatives mostly equate coherence with non-duplication—avoiding replication of work across discrete forums and processes—and non-contradiction—that efforts and content in respective processes not oppose one another—rather than ‘complementarity’. The importance of non-duplication is to preserve finite time and resources, particularly given the busy disarmament schedule in Geneva. The importance of non-contradiction is to ensure that respective processes do not cancel each other out and lead to contradictory outputs.

Complementarity, by contrast, means two or more different things augmenting or emphasizing each other’s qualities. For instance, complementarity between multilateral processes on AWS might be achieved through one forum or process providing detailed, technical groundwork on AWS (e.g. the GGE on LAWS), particularly on developing treaty language, while other forums (e.g. the General Assembly) foster more inclusive involvement and momentum.⁴⁹ More broadly, the desire to avoid duplicating or contradicting efforts on AWS in military AI multilateral initiatives has meant that the question of complementarity between these respective processes has been given limited attention—despite the fact that language and concepts related to discussions on military applications of AI are increasingly present in discussions on AWS, including in the GGE. Broadly speaking, opportunities to harmonize various processes dedicated to AWS and military applications of AI are not currently being sought.

The second aspect of coherence is consistency with the overall aims and precedents of disarmament, the regulation of weapons and IHL. This includes consistency of processes and outcomes with already established frameworks and international governance regimes for arms control and disarmament. In the international policy discussions on AWS so far, one of the more prominent issues has been consistency with IHL of language and concepts in the GGE on LAWS (its rolling text). States also have historic positions on disarmament and arms control, as well as existing practice within the CCW and elsewhere, that they seek to ensure consistency with.

Notably, interviewees did not mention the effects of different policy outputs—such as the negotiation of a legally binding instrument or the adoption of a statement on the application of IHL to AWS—on normative development within IHL or international law more broadly. That is, the risk that pursuing such policy outputs might lead to fragmentation—with states holding different obligations or different interpretations of their obligations with respect to means and methods of warfare and the conduct of hostilities—was not raised as a factor being weighed.

Wholesale versus incremental change

Interviewees held different perspectives on whether possible policy outputs should be assessed on their own merits, or as one step in a journey of incremental normative development. For some, policy outputs that are viewed as too modest, too narrow or otherwise falling below expectations are unacceptable. For example, one interviewee indicated that adopting a prohibition on the development and use of ‘non-existent’ AWS would be unacceptable. This perspective may be explained by a fear that the opportunities for developing the normative framework relating to AWS are reducing or disappearing altogether, or that states with modest ambitions will use any output, no matter how lacklustre, to argue that they have discharged their responsibility to act to address the risks posed by AWS, further limiting the options or appetite for progress.

Others take a long-term perspective, viewing any output, no matter how modest or narrow, as forming part of an incremental approach to norm development. For these participants in the policy debate, ‘something is better than nothing’, and even an output that represents agreement on a minimal set of issues or is not enforceable (such as a political declaration) provides a basis on which to build further policy developments, and could influence state behaviour in the meantime.

Geopolitical context

Global politics affect how participants engage in the AWS policy process and the development of national positions. Interviewees cited various aspects of the deteriorating geostrategic environment—increasing polarization and instability, and worsening regional security—as factors affecting their view of the policy process. These factors reinforce and justify the view that the development and use of AWS are a necessary part of states’ response to their security concerns. This reflects a view that has affected adherence to existing weapons treaties. For example, in announcing their withdrawal from the Anti-Personnel Mines Convention, Estonia, Finland, Latvia, Lithuania, Poland and Ukraine cited security concerns.⁵⁰ According to some interviewees, this view of the arms–security nexus is making it more difficult to attract support for restraints on the development and use of AWS which, in turn, is affecting states’ perceptions of those options, with some states not wishing to constrain themselves more than their (current or future) adversaries.

Domestic context

At the national level, two factors affect the way participants view possible policy outcomes for AWS: the influence of national values and experiences, and the influence of the executive and legislative branches of government.

For many, policy responses to AWS are explicitly grounded in national values—such as respect for human dignity—and reflect historical experiences of weapons use and armed conflict on their territory or against their people.⁵¹ These factors militate for policy outputs that result in greater restraint on the development and use of AWS.

For others, executive and legislative branches of government have been central to shaping the development of a national position. For example, in 2019 the Dutch Parliament adopted a resolution calling for binding international rules on AWS.⁵² National policy positions on AWS have also appeared in the political platforms on which governments have been elected.⁵³ In such instances, there is a constituency that bears directly on the position that a state can take in their multilateral policy response to AWS. This can be shaped by ideas and events that have an electoral impact such as public dismay over military use of an autonomous weapon, or concerns regarding defence readiness or capability.

Less visible but nonetheless relevant is how states’ views on policy options are shaped by dynamics between relevant government ministries, particularly between ministries of foreign affairs and ministries of defence or armed forces. Interviewees who cited this as a relevant policy factor indicated that national positions led by a ministry of foreign affairs are more likely to prioritize humanitarian concerns or risks and challenges related to AWS development and use, while those led by a ministry of defence are likelier to prioritize the potential military utility of AWS. Reservations on the part of ministries of defence or armed forces about a particular policy direction may also result from perceived difficulties in implementing new regulation domestically.

Legitimacy of the policy process

Legitimacy of political institutions and of the decisions made within them is a key consideration of policymakers when evaluating future directions on AWS. Legitimacy pertains to two principal factors and a third subsidiary factor.

First, legitimacy relates to the universality of the given process or output. Simply, the greater the number of states committed to a process or output, the greater the legitimacy that process or output commands. This aspect of legitimacy can be parsed as both process and output. For instance, the UN General Assembly is seen as an ideal forum in terms of legitimacy of process, since it entails near-universal representation of states. However, some interviewees expressed scepticism about General Assembly resolutions which, needing only a majority vote, can give a false impression of global consensus about an issue. Some states indicated that an output adopted by consensus—that is, the absence of any express opposition, or the appearance of unanimity—is more legitimate than an output that has been adopted by a majority vote, especially an output that some states have voted against. This is cited as reason for preferring discussions under the auspices of the CCW. The mandate of the GGE on LAWS requires it to consider and formulate elements of an instrument by consensus, and the group conducts its discussions under rules of procedure that, in general, require decisions to be adopted in the same manner as the CCW—that is, by consensus.⁵⁴ Others, however, indicated that the ‘consensus rule’ in the CCW leads to deadlock and weakens the forum’s ability to discharge its mandate, perhaps undermining its legitimacy.

Second, legitimacy relates to the authority of a given process. That the UN is seen as the most legitimate forum in which to pursue policy outcomes reflects not only states’ numerical representation. It also reflects the authority of the UN resulting from its historic position and mandate to make recommendations with regard to the principles governing the regulation of armaments and to promote the development and codification of international law.⁵⁵ A number of officials indicated that their state had decided not to participate in certain international non-UN governance initiatives because they did not want these initiatives to substitute for the role they believed the UN ought to play. A number of states expressed scepticism—some of it deep—and wariness about normative development outside of UN channels, typically giving the REAIM initiative as an example.

Third, legitimacy relates to a forum’s specialization, which is seen as providing a type of authority. For example, the GGE on LAWS holds legitimacy partly for its being a process organized by the UN and partly because it has accrued a significant pool of specialization on AWS.

Participation in policy outputs

In weighing possible policy directions, participants in the multilateral policy debate are considering the potential breadth of support and opposition for a policy output. This consideration has two facets: the first relates to the identity of supporters and detractors, and the second relates to the diversity of support.

First, many interviewees emphasized the importance of achieving support from states likeliest to acquire and develop AWS. These are sometimes referred to as ‘major military powers’ or ‘militarily advanced states’. The grouping’s composition is not always expressed, but China, India, Russia and the USA have been mentioned in this context. The desired ‘support’ from these states encompasses formal and informal, active and passive support. For example, these states may show support by joining consensus in a formal policy process (i.e. not detracting from the vote), voting in favour of an output in the UN General Assembly, adhering to a legally binding instrument, or abstaining from voting against a policy output or breaking consensus in a formal process.

The support of these states is viewed as an essential ingredient in the success of any policy output, and necessary for any output to have a meaningful impact on addressing the risks posed by AWS. This is because one measure for the success of a norm on AWS is the extent to which it restrains (or appears to restrain) states that develop and use such weapons. Examples include the important role of the United Kingdom’s decision to support the draft text of the Convention on Cluster Munitions in attracting the support of other states with stockpiles of cluster munitions; and the role that the US withdrawal of its textual proposals on the draft Anti-Personnel Mines Convention played in the adoption of that treaty.⁵⁶ However, some interviewees cautioned against permitting a small group of major military powers or militarily advanced states to have an outsized influence on the development of IHL. For some, the perceived poor record of compliance with and support for IHL among this group was a further reason for not weighting their influence on the policy debate too heavily.

The second facet of participation is the diversity of states that express support for an option based on, for example, geographic and economic indicators. This is one rationale some interviewees gave for devoting time and resources to consulting with a broad group of states in a wide range of forums, including among permanent missions to the UN in New York and regional bodies such as the Association of Southeast Asian Nations and the African Union. This process of ‘democratization’ or ‘mainstreaming’ of AWS as an agenda item is seen as helpful for building consensus, as well as for promoting coherence.

Conclusion

While it may not be evident on its face, the multilateral policy debate on AWS is shaped by a complex interplay of considerations at the national level. For policymakers, it appears the process is as important as the outcome. Securing legitimacy, ensuring broad participation and maintaining coherence with existing legal frameworks are likely to influence whether future policy directions are viewed as sustainable and widely supported. Divergent geopolitical interests and national priorities are affecting whether incremental progress or wholesale change is more likely.

Research indicates that policymakers are operating under an awareness of the need to balance urgency with caution, inclusivity with effectiveness, and ambition with pragmatism. Decisions taken now will not only shape policymaking on AWS, but also set precedents for the regulation and governance of other technologies of warfare, notably military AI. Further, AWS policy cannot be addressed in isolation; it is embedded in the wider fabric of international security, IHL and trust in multilateralism. Chapter 4 discusses the challenge for policymakers—when and how to take action in a way that meaningfully addresses the risks and challenges posed by the development and use of AWS, while fostering consensus.

4. Future policy directions

As with politics generally, international policymaking on AWS is about the art of the possible (see figure 4.1). After a decade of effort, stakeholders—policymakers in particular—have now before them a range of policy directions for advancing international efforts on AWS. States must now make decisions about the outputs they wish to achieve and the process or processes through which these will be achieved, in light of the overall outcome they desire.

https://dev.sipri.org/sites/default/files/Figure%202_0.png

Figure 4.1. A summary of some of the policy directions for advancing international efforts on autonomous weapon systems

AI = artificial intelligence; AWS = autonomous weapon systems; CCW = Convention on Certain Conventional Weapons; GGE = Group of Governmental Experts.

This chapter outlines foreseeable policy directions and considers them with respect to the factors discussed in chapter 3. These policy directions have been collated in consultation with representatives of states and non-state organizations. They are divided into both policy processes—procedural rules, participation and substantive issues—and the types of outputs that might be produced through these processes.

Processes

A policy process refers to a particular configuration of types of procedural rules, participation and substantive issues by which a policy response is deliberated and determined. This section discusses process in terms of international forums because international policymaking takes place within and alongside a set of historically given institutions (e.g. the UN) that constrain choices about the design of a process. The processes discussed are not mutually exclusive, and they can be pursued either simultaneously or consecutively.

The Convention on Certain Conventional Weapons

The CCW GGE on LAWS has been the predominant forum for much of the history of international policy efforts on AWS. The GGE will arrive at a critical juncture in 2026 when its current mandate expires and its existence is decided at a review conference of the HCPs to the CCW. A renewed mandate—if it is renewed—could take a number of forms, including a rollover of the existing mandate, a move to a negotiating mandate, or a mandate broadening or narrowing the terms or scope of discussion.

Context, coherence and legitimacy were key considerations voiced by interviewees with respect to the CCW GGE on LAWS as a forum. In the context of a period marked by geopolitical instability, competition and dwindling trust among states, and where security and humanitarian considerations jostle for preponderance, the CCW’s contested status as both an IHL and a disarmament forum was seen as advantageous. On the one hand, its track record on prohibiting the use of certain weapons (e.g. laser weapons that cause permanent blindness, weapons that injure by non-detectable fragments) demonstrates it to be a forum for progressing weapons regulation. On the other hand, its history of taking a pragmatic approach to regulating the use of weapons that are not prohibited in line with IHL and involving those developing and using the weapons, means it can accommodate those that see strategic and military utility in AWS.

The CCW GGE on LAWS is also seen as pivotal to the issue of coherence. Because it has been de facto treated as the international forum on AWS policy discussion, it provides the benchmark against which the coherence of other policy initiatives on AWS will be assessed. For example, a number of interviewees saw it as important that future initiatives neither duplicate nor contradict what has already been achieved at the GGE.

The CCW GGE on LAWS was also seen as commanding considerable legitimacy, both as a UN process and because it has accrued considerable expertise on AWS. The active participation of the military powers in the GGE is also seen as important for the effectiveness of any outcome resulting from efforts in that forum.

The First Committee of the UN General Assembly on disarmament and international security

There is an open question about what can be done next at the UN General Assembly (which meets in New York), and what its role is meant to be in the governance landscape around both AWS and AI in the military domain. As yet, there is no clear articulation of a long-term vision of what can and ought to be achieved regarding AWS at the First Committee dealing with disarmament and international security. For example, there are still open questions about whether a new process at the First Committee would be meant to extend the work of the Geneva-based CCW GGE on LAWS and its existing goals, or whether it would be meant to have a purpose distinct from that of the CCW GGE on LAWS. This goes to the heart of the question of coherence and complementarity between these processes.

The current lack of long-term vision around a New York–based process reflects that states are hesitant to pre-empt the success or failure of the CCW process by formalizing a process in New York. While understandable, this has had the effect of militating against longer-term thinking on AWS multilateral governance. For instance, there are questions about how New York fits within incremental, longer-term norm development on AWS, particularly as many see any output from the CCW GGE on LAWS as constituting only a partial contribution to an AWS policy response.

That said, there are a number of shorter-term advantages for a process on AWS in New York. Many policymakers supportive of a process in New York impute tactical value to it insofar as it puts external pressure on the CCW process to achieve a substantive result. A New York–based process was also seen as an opportunity for states to develop expertise on a range of issues other than those related to IHL. For others still, the value of a process in New York is that it permits greater representation of states than the CCW, strengthening the legitimacy of the process. Indeed, the overwhelming support for the two UN General Assembly resolutions on AWS have been taken by some as proof that the CCW process is allowing a handful of states to block the development of a treaty that the majority of states appear to urgently want. Among those willing to countenance the failure of the CCW GGE on LAWS, a New York–based process offers the possibility of preserving and extending efforts at the CCW, for instance through continuing work on the rolling text. In any case, any further informal consultations in New York could be used to discuss, identify and develop the format and scope of any future process, either in the First Committee or within the wider UN system.

There have also been diplomatic talks at the UN General Assembly on AI in the military domain. These talks could represent an opportunity to fold AWS—perhaps the most well-known use case of military AI—into wider military AI discussions. The feasibility, advantages and disadvantages of folding AWS into a military AI process generally are discussed below. For the moment, in the context of the First Committee, the military AI resolution is explicit about excluding AWS from such talks.⁵⁷

Existing efforts at New York provide the opportunity also to fold discussions on AWS into a process on military applications of AI, namely initiatives associated with the UN General Assembly’s resolution on military AI. The CCW GGE on LAWS is in principle—if not always in practice—neutral regarding the technologies that can enable AWS.⁵⁸ This is despite the fact that the sets of concerns that have motivated contemporary regulatory discussions on AWS relate to the capabilities and functions enabled by AI, and that in the broader policy and academic literature, and in public discourse, there has been a tendency to treat AWS as foremost an AI issue, and AWS as one of the clearest and dramatic instances of military AI. Folding AWS into military AI discussions generally would therefore present policymakers with an opportunity to treat and respond to AI-specific AWS and the particular issues they raise. Taking AWS into military AI discussions could also help ensure the coherence of governance and regulatory discussions on military AI technologies, and mitigate the possibility of contradiction between AWS-specific initiatives and military AI initiatives. In this regard, states’ responses to the UN secretary-general’s recommendation that they establish a ‘dedicated and inclusive process to comprehensively tackle the issue of AI in the military domain and its implications for international peace and security’ may serve as an indicator of the feasibility of folding discussions on AWS into a process on military AI.⁵⁹

For some, however, folding AWS into military AI discussions generally would constitute a tacit acknowledgement that regulatory efforts at the CCW are not progressing adequately, and it is only by maintaining a focus on AWS as a discrete issue, in a discrete forum, that the hope of a legally binding instrument can be kept alive. This option is therefore unlikely to be countenanced in the short term by those seeking a legally binding instrument on AWS.

Elsewhere in the United Nations

The CCW and the First Committee are not the only possible forums within the UN for a process on AWS. There is also the possibility of establishing a policy process within the HRC, returning deliberations on AWS to their initial starting point within the UN. Interviewees viewed establishing a policy process within the HRC foremost as an opportunity to discuss sets of concerns that are widely believed to be pertinent to the development and use of AWS but which cannot be given sufficient treatment within the framework of the CCW, such as implications for IHRL.

As demonstrated by its 2024 debate on the use of AI in conflict (see chapter 2), there is a role for the UN Security Council in debate on AWS, at least in a general sense, through either a debate or (less likely) a resolution addressing the peace and security implications of the development and use of AWS.

Outside the United Nations

Illustrative examples given of a state-led process outside the UN are the Ottawa process—referring to the diplomatic, political and civil society efforts that led to the creation of the Anti-Personnel Mine Convention—and the similar Oslo process that led to the Convention on Cluster Munitions. Any new process outside the UN would require states to decide on both the procedural and substantive elements of that process.

Another option is to fold discussions on AWS into a policy process on military AI outside the UN, such as existing initiatives (described in chapter 2) or any future initiatives. This would have the advantage of providing a forum in which issues of AWS as a specifically AI-based technology can be addressed. It could also provide the opportunity to develop a new framing around these technologies, namely a focus on the role of the human in the decision to use force.

On the whole, processes outside the UN were thought likely to command less legitimacy than a UN-based process and as possibly lacking in effectiveness through failing to achieve buy-in from relevant states. The possibility of a process outside the UN was, on the whole, seen as unlikely or undesirable until all UN options had been exhausted.

Outputs

Legally binding instrument

Perhaps the most widely recognized output that could emerge from an AWS policy process is a legally binding instrument. This term refers to an international agreement on AWS concluded between states and governed by international law. States and non-state groups advocating for a robust international legal framework on AWS favour a comprehensive treaty that contains prohibitions and restrictions on the development and use of AWS.⁶⁰ A legally binding instrument could exist as a new protocol under the CCW, or as a standalone treaty.

Policymakers view a legally binding instrument as the highest-impact output—both symbolically and practically. Many interviewees noted that without legal commitments, norm internalization may be weak, particularly in conflict-prone or authoritarian contexts. However, some caution that impact depends on scope and enforcement mechanisms, and that delayed negotiations could mean any instrument will have less relevance if it is outpaced by technological advances.

Legitimacy of a legally binding instrument is a key factor. An instrument negotiated in a UN setting carries greatest legitimacy. Several interviewees emphasized that a treaty negotiated through the General Assembly or the CCW would carry unmatched normative weight. There was a concern that, if a small group of states push forward with a legally binding instrument without wider support, the instrument risks being dismissed as politically motivated or exclusionary. Another key consideration was participation. Many highlighted the difficulty of achieving broad participation in reaching a legally binding instrument due to divergent national positions. Some states, especially major military powers, are opposed or ambivalent about an instrument that is legally binding, which could weaken its global coverage.

A few interviewees supported adopting the approach used for the 2017 Treaty on the Prohibition of Nuclear Weapons (TPNW).⁶¹ This approach would begin with like-minded states exerting normative pressure outside of UN processes to establish a legally binding instrument on AWS. However, a number of interviewees expressed apprehension at a TPNW-style approach to AWS. There is regret among some at the speed with which the TPNW was taken outside of the UN, meaning a treaty drafted in haste, and discontent among a number of states (not just nuclear-armed states) that their interests are not served by the treaty. Nonetheless, some interviewees noted that the TPNW approach indicated a general lesson: that disarmament has typically worked through a minority of states or civil society groups building pressure and generating greater support on that basis. It has also worked the other way round, with major powers deciding it is time to eliminate a category of weapon—for example, biological and chemical weapons—and bringing everyone along with them.⁶²

A legally binding instrument, according to several interviewees, could potentially introduce greater coherence to the international governance landscape, especially if linked to existing IHL obligations. Others worry that a new treaty might complicate or even fragment existing discussions if it is pursued outside the CCW framework or without integration into broader military AI debates.

However, political context was widely recognized as a major barrier to a legally binding instrument. Interviewees cited geopolitical tensions and CCW consensus rules as making a legally binding instrument unlikely in the short term. However, some noted that future crisis scenarios or technological misuse could make such an instrument more feasible, especially as part of a reactive shift in the international climate.

Outputs of a less-than-treaty status

Outputs of a less-than-treaty status might include high-level political declarations or statements of commitment—non-binding instruments signalling intent, setting norms or coordinating behaviour among states. High-level documents that already set out shared principles for both AWS in particular and military AI generally include the CCW GGE on LAWS guiding principles, the US political declaration and the REAIM Blueprint for Action (see chapter 2). These existing documents could form the basis of a broader political commitment on the use of AWS. A global political declaration on AWS could also be built around consensus language already developed in a UN-based forum, such as the CCW GGE on LAWS. This might take the form of a group of states, a regional forum or a General Assembly resolution issuing a declaration of standards or expectations for AWS governance.

Political declarations are generally seen as flexible in scope, which can be both an advantage and a limitation. Their breadth allows them to capture emerging consensus language but they often avoid specificity to maintain broad support. This vagueness can dilute normative clarity, with political declarations signalling intent but failing to lock in details.

Interviewees generally agreed that political declarations have limited legal impact but can shape expectations, frame state behaviour and lay the groundwork for future legally binding instruments. They were described as tools for setting agendas, achieving buy-in for norms, and mobilizing coalitions. However, several stakeholders warned that declarations are sometimes treated as endpoints rather than stepping stones—that is, a political declaration might be used to establish the baseline, but there needs to be a clear pathway beyond a political declaration to avoid the risk of stagnation.

Outputs of a less-than-treaty status could also include interpretive guidance intended to clarify how existing legal obligations apply to AWS. In 2022 the United Kingdom proposed such an output on AWS.⁶³ An example is a manual on the application of IHL to AWS, similar in style and approach to those on international law applicable to armed conflicts at sea and to cyber warfare, among others.⁶⁴ Such documents could set out expert interpretations of how IHL governs AWS use. Normative guidance could also include a best practice compendium on (for instance) how different states conduct legal reviews, how they formally interpret and understand key legal provisions, and how they realize higher-level principles in practice.

Guidance documents have interpretive and normative impact, rather than regulatory impact. That is, they do not directly create obligations (though they can contribute to the crystallization of a customary rule of international law), but instead influence how existing obligations are understood and how militaries review AWS development and use.⁶⁵ They are especially valued by states that wish to clarify IHL obligations with respect to AWS without negotiating new law. However, some stakeholders note their limited policy traction outside legal circles. The legitimacy of these types of outputs is also a key factor, with particular concerns around the idea that such manuals would exist as elite exercises or reflect only Global North practice. However, there can be value in military-to-military dialogue for the discussion of best practices.

Governance frameworks

Governance frameworks comprise institutional or procedural mechanisms to support ongoing regulation of AWS. A possible AWS governance framework could be a new standing panel or advisory board composed of governmental or independent experts, or a multi-stakeholder group reporting to the UN General Assembly or the UN secretary-general. Some interviewees pointed out the value of the secretary-general’s report collating views of states on AWS, and how such a useful exercise should not be a standalone effort. A standing body or expert group could track developments in AWS technology; testing, evaluation, validation and verification practices; and assurance practices, certification regimes and standards. State submissions to a standing body or panel could be modelled on transparency measures in existing instruments, such as the reporting requirement under other arms control treaties, such as the Anti-Personnel Mine Convention, the Convention on Cluster Munitions or the Arms Trade Treaty.⁶⁶ One interviewee highlighted the value of a programme for AWS similar to the UN Programme of Action on small arms and light weapons, in which states agreed to strengthen and report regularly on their domestic laws, regulations and policies with respect to small arms and light weapons, and to cooperate with and assist each other in eradicating the illicit trade in these weapons.⁶⁷ The multidisciplinary Independent International Scientific Panel on AI and Global Dialogue on AI Governance, while limited to the non-military domain, will likely influence how states view the feasibility of this kind of governance initiative for military AI.⁶⁸

A contention around governance frameworks for AWS is their mandate and scope. Stakeholders generally see institutional mechanisms as having a procedural rather than normative scope—they usually do not set hard rules but facilitate national or international regulation through governance processes. Their remit can be narrow (e.g. monitoring technical developments) or broad (e.g. convening regular dialogues across states and disciplines). Interviewees highlighted their flexibility, which allows them to evolve over time, though some worry that an ill-defined mandate or scope risks mission creep or institutional redundancy. The impact of governance frameworks is seen as indirect but structurally significant—they can shape the policy environment by aggregating knowledge, ensuring continuity and enabling informal coordination. They may not impose obligations but can help translate political intent into practice.

In the current multilateral diplomatic context, institutional mechanisms are viewed as feasible, low-risk options. They are politically easier to agree on than substantive rules, and can serve as interim solutions while consensus builds. Several interviewees pointed to analogies with other arms control domains, where review conferences or expert groups have kept discussions going during stalemates.

Such governance frameworks also offer opportunities for broader participation, especially for non-state actors, technical experts and civil society. Many interviewees valued them for lowering the entry barrier for those unable to engage in formal diplomatic negotiations. However, true inclusivity depends on open consultation, language access and representation balance being built into the framework structure.

Conclusion

Foreseeable policy directions on AWS reflect a dynamic interplay between outputs, processes and forums. These variables are not independent; rather, they co-constitute the policy directions available to states. Different combinations—such as binding legal instruments negotiated within the CCW, political declarations developed in the UN General Assembly, or new institutional frameworks housed elsewhere—carry distinct implications for states in developing their national positions on AWS, particularly when weighing the impact of policies in the world, issues of legitimacy of a process, and the breadth and depth of participation. The choice of forum shapes the range of outputs deemed feasible, just as the process determines the procedural norms and coalitions that can be built, and whether wholesale or incremental change can be expected.

The CCW remains the preeminent and broadly preferred forum for international discussions on AWS, commanding not only legitimacy but also a level of expertise seen as necessary for addressing the scope of issues states wish to see discussed and the impact they see as desirable. Yet, stakeholders increasingly recognize the value of pursuing alternative outputs and processes, including those emerging in the UN General Assembly or outside the UN. This reflects a broader willingness to explore new procedural formats and hybrid strategies, when the timing is advantageous. In contrast, on military AI more broadly, no clear policy pathway has yet emerged. The lack of consensus on goals, forums and outputs for military AI governance remains a major constraint—and a reminder that the AWS debate, while complex, is relatively more advanced.

5. Findings and recommendations

Difficulties in conceptualizing concerns around AWS, limitations in the processes and tools available for multilateral efforts, and eroded trust among states have stymied multilateral efforts to develop policies responding to the risks and challenges posed by AWS. The previous chapters in this report have equipped policymakers with information about the factors being weighed in states’ consideration of possible policy directions, and outlined the layered trade-offs and implications that these directions entail, beyond the yes/no question of whether states should negotiate a legally binding instrument regulating AWS. This information is of use to states not only for considering future directions with AWS but also for establishing policy responses to other technologies of warfare, such as military applications of AI. This chapter sets out the report’s three findings and proposes three recommendations for states that are developing their national positions in relation to the multilateral policy processes on AWS and AI in the military domain.

Findings

States are weighing nuanced factors when developing their respective national positions on policy directions on AWS. However, these nuances are rarely expressed in the multilateral debate. The result is that the multilateral debate may appear (falsely) to involve only a dichotomous choice between whether or not to negotiate a legally binding instrument on AWS.

Participants in the policy debate have nuanced approaches to assessing the feasibility of potential policy directions. States’ consideration of policy directions is dynamic, interconnected and fluid.

Research for this report indicates that state representatives have in mind the range and scope of potential policy approaches; the impact of the technology and the policies in the world; questions of timing; coherence with broader governance frameworks; and whether potential outputs are standalone or represent incremental change. States’ national positions are also influenced by geopolitical and national contexts, questions of legitimacy around both process and output, and the breadth of support that a policy option attracts.

At the multilateral level, however, these nuances are rarely reflected—whether due to procedural strictures (e.g. time limits on statements at the CCW GGE on LAWS), sensitivity or confidentiality around the process of national policymaking, or deliberate or unintentional omission. As a result, the multilateral debate may appear, falsely, to be between those who support the negotiation of a legally binding instrument on AWS and those who do not. This false dichotomy masks two possible realities. On one hand, states may be able to take more steps to progress the policy process than is apparent; on the other hand, the nuanced reality may make it more difficult for states to coalesce around a way forward.

States share some common ground with respect to AWS policy directions but progress towards outputs will require more political will and leadership.

The national approaches to policy directions belie a broad convergence among states that the CCW is an appropriate forum for continuing policy discussions. Support for the CCW as a forum can be understood as a proxy for common ground among states on questions of forum, procedure and focus (e.g. the ‘two-tier approach’ and IHL as the primary legal framework).

Against this background, the lack of progress in policy discussions should be attributed not to an absence of common ground but, rather, to a deficit in political will and leadership. Political will and leadership are largely linked to context—both geopolitical and domestic. The recommendations below address what states can do to progress policy discussions within the current context.

States may be missing an opportunity to align AWS policy discussions with broader discussions on military applications of AI.

States’ goal of seeking to ensure that the totality of international policy efforts on AWS and military AI are coherent and effective—and do not result in inconsistent governance or regulatory regimes—has flattened into one where states are seeking to avoid duplication.

The result is that states may be missing opportunities to harmonize and coordinate their efforts, despite the fact that integration of AI into decisions on the use of force and AWS (whether enabled by AI or not) raise overlapping risks and challenges. This is especially so with regard to the use of AI-enabled decision support systems in military targeting. The military use of AI to enable AWS or other military applications related to the use of force alters the human role in decision making—an issue that has been central, though not always explicit, in the technology-agnostic discussions in the CCW GGE on LAWS.

Recommendations

States should ensure that their national policy positions are holistic, taking into account the range of possible policy directions. States should be as transparent as possible about their views on the various policy directions.

Addressing risks and challenges posed by AWS requires states to use all means at their disposal. This is because no process or output alone is capable of addressing all concerns regarding the development and use of AWS. At the same time as they are developing the elements of an instrument on AWS under the auspices of the CCW GGE on LAWS, states should pursue other policy processes and outputs that might help them achieve the outcomes they seek. These include joint statements agreed at the political level, such as declarations; guidance on the interpretation and application of international law relevant to the development and use of AWS; translation of relevant international law rules for use by groups such as militaries or developers of AWS; coordinated integration of AWS and related topics on the agendas of relevant international forums; and frameworks that support transparency and confidence building among states.

For this reason, states should ensure that their national policy positions account for the range of possible policy directions and how these might assist them in reaching their desired outcomes. To develop a holistic national position, states could consider the outcomes they wish to achieve with respect to AWS, including the legal, ethical and strategic concerns that underpin these desired outcomes; how various policy directions align with or contribute to the achievement of their desired outcomes; the trade-offs and implications of pursuing the various policy directions (including sequencing); and the resources (financial, human, institutional) that they can draw on in order to achieve their desired outcomes.

Further, openness about their respective national policy positions on the range of policy directions, to the greatest extent possible, could facilitate coordination within and among the various extant policy processes. This could aid in the development of a coherent approach to regulation.

States wishing to pursue an instrument on AWS—no matter their views on its status, forum, procedure or timing—should build on the rolling text of the CCW GGE on LAWS.

The rolling text in the GGE on LAWS is substantive and is the result of constructive participation from a large and diverse group of states. Regardless of whether consensus exists on the document as a whole, or the level of support that each paragraph attracts, the rolling text has achieved a certain status as a source document. The result is that this document is—and is considered by policymakers as—the repository of states’ views. Given this status, this document will likely be the first and most important reference point for the development of any instrument on AWS, of any status, in any forum, at any time. For this reason, states wishing to pursue an instrument—no matter their views on its status, forum, procedure or timing—should build on and preserve the rolling text.

Measures that could build on the rolling text include detailed and constructive formal statements at the GGE on LAWS (and in informal consultations convened by the chair of the GGE); development and sharing of detailed national positions on matters dealt with by the rolling text, including the interpretation and application of international law with respect to the development and use of AWS; and convening of discussions among relevant officials in open or regional settings to support the development of national positions and facilitate understandings among states of their respective views.

States’ policy decisions with respect to AWS should take into account the common risks and challenges posed by the use of AI in decisions to use force.

States should ensure that their policy decisions with respect to AWS are informed by understanding of the overlapping risk and challenges posed by the integration of AI into military decision making, focusing on the effect on the role of the human in the use of these technologies.

Opportunities for sharing of views and facilitating increased awareness and understanding of the overlapping issues related to the role of the human in decisions to use force exist in the context of the GGE on LAWS and the First Committee of the UN General Assembly, as well as in other forums such as REAIM. States should harness these opportunities by expressly addressing in their statements common legal, ethical and policy issues raised by AWS and military AI, and identifying and explaining distinctions where appropriate. States should also have in mind these overlapping risks and challenges when framing mandates for future discussions, for example in drafting resolutions on these themes in the General Assembly or the CCW.

Endnotes

1. 1 Blanchard, A. et al., ‘Dilemmas in the policy debate on autonomous weapon systems’, SIPRI Backgrounder, Feb. 2025; Boulanin, V. and Verbruggen, M., Mapping the Development of Autonomy in Weapon Systems (SIPRI: Stockholm, Nov. 2017), pp. 24–27; and International Committee of the Red Cross (ICRC), ‘ICRC position on autonomous weapon systems’, 12 May 2021.
2. 2 Sauer, F., ‘Stepping back from the brink: Why multilateral regulation of autonomy in weapon systems is difficult, yet imperative and feasible’, International Review of the Red Cross, vol. 913 (Mar. 2021); Bode, I. and Qiao-Franco, G., ‘The geopolitics of AI in warfare: Contested conceptions of human control’, eds R. Paul, E. Carmel and J. Cobbe, Handbook of Public Policy and Artificial Intelligence (Edward Elgar: Cheltenham, 2024); Kmentt, A., ‘Geopolitics and the regulation of autonomous weapons systems’, Arms Control Today, Jan./Feb. 2025; and Taddeo, M. and Blanchard, A., ‘A comparative analysis of the definitions of autonomous weapons systems’, Science and Engineering Ethics, vol. 28 (2022).
3. 3 United Nations, Department for General Assembly and Conference Management, ‘Regional groups of member states’, [n.d.].
4. 4 United Nations, General Assembly, ‘Interim report of the Special Rapporteur on extrajudicial, summary or arbitrary executions’, A/65/321, 23 Aug. 2010.
5. 5 United Nations, General Assembly, A/65/321 (note 4), para. 48.
6. 6 United Nations, Human Rights Council, ‘Report of the Special Rapporteur on extrajudicial, summary or arbitrary executions, Christoph Heyns’, A/HRC/23/47, 9 Apr. 2013, para. 110.
7. 7 United Nations, Human Rights Council, A/HRC/23/47 (note 6), para. 118.
8. 8 United Nations, Human Rights Council, A/HRC/23/47 (note 6), para. 114.
9. 9 Human Rights Watch (HRW) and Harvard Law School International Human Rights Clinic (IHRC), Losing Humanity: The Case Against Killer Robots (HRW/IHRC, Nov. 2012), p. 2.
10. 10 HRW and IHRC (note 9), pp. 6–14.
11. 11 See e.g. French delegate to the CCW GGE on LAWS, Speech to the Meeting of States Parties to the CCW, Geneva, 13–14 Nov. 2014 (in French).
12. 12 CCW, Meeting of the High Contracting Parties, Final report, CCW/MSP/2013/10, 16 Dec. 2013, para. 32; and Campaign to Stop Killer Robots, ‘Report on outreach on the UN report on “lethal autonomous robotics”’, 31 July 2013, pp. 11–12 (Brazil) and pp. 14–15 (France).
13. 13 CCW, Fifth Review Conference, ‘Final document of the Fifth Review Conference’, CCW/CONF.V/10, 23 Dec. 2016, p. 9.
14. 14 CCW, Group of Governmental Experts on Emerging Technologies in the Area of Lethal Autonomous Weapon Systems (GGE on LAWS), Report of the 2019 session, CCW/GGE.1/2019/3, 25 Sep. 2019, annex IV.
15. 15 CCW, GGE on LAWS, CCW/GGE.1/2019/3 (note 14), p. 7.
16. 16 CCW, GGE on LAWS, Report of the 2018 session, CCW/GGE.1/2018/3, 23 Oct. 2018. For examples of the proposals see United Nations Office for Disarmament Affairs (UNODA), ‘Convention on Certain Conventional Weapons—Group of Government Experts on Lethal Autonomous Weapons (2022): Documents’, 2022.
17. 17 See Bruun, L., Towards a Two-tiered Approach to Regulation of Autonomous Weapon Systems: Identifying Pathways and Possible Elements (SIPRI: Stockholm, Aug. 2024).
18. 18 For the most recent version of this rolling text see CCW, GGE on LAWS, ‘Rolling text, status date: 12 May 2025’, 2025. (The term ‘rolling text’ in the CCW is not used in its formal sense within the UN context, where it generally refers to a text that goes through several readings during the negotiation of a resolution by UN organs. United Nations, ‘Drafting resolutions’, United Nations Guide to Model UN (UN: Geneva, 2020).)
19. 19 CCW, GGE on LAWS, ‘Joint statement to the September 2025 session of the CCW GGE LAWS’, Sep. 2025.
20. 20 United Nations, General Assembly, First Committee, 77th Session, Joint statement on lethal autonomous weapons systems delivered by Austria et al., 21 Oct. 2022, p. 2.
21. 21 UN General Assembly Resolution 78/241, 23 Dec., para. 2.
22. 22 United Nations, General Assembly, ‘Lethal autonomous weapon systems: Report of the Secretary-General’, A/79/88, 1 July 2024, para. 88.
23. 23 See Communiqué of the Latin American and the Caribbean Conference of Social and Humanitarian Impact of Autonomous Weapons, adopted on 24 Feb. 2023; Economic Community of West African States (ECOWAS), ‘Communiqué of the Regional Conference on the Peace and Security Aspects of Autonomous Weapon Systems: An ECOWAS perspective’, 18 Apr. 2024; and Philippines Department of Foreign Affairs, ‘Philippines calls for Indo-Pacific voices to address lethal autonomous weapons system risks’, Press release, 15 Dec. 2023 (reporting on the Manila Meeting on Indo-Pacific Perspectives on Autonomous Weapon Systems on 13–14 Dec. 2023).
24. 24 United Nations, A/79/88 (note 22), para. 90.
25. 25 United Nations, A/79/88 (note 22), paras 91–92.
26. 26 UN General Assembly Resolution 79/62, 10 Dec. 2024, paras 5 and 7.
27. 27 Ebo, A., ‘Welcoming remarks to the informal consultations on lethal autonomous weapon systems’, UNODA, 12 May 2025, p. 1.
28. 28 United Nations, Human Rights Council, ‘Possible impacts, opportunities and challenges of new and emerging digital technologies with regard to the promotion and protection of human rights’, A/HRC/47/52, 19 May 2021.
29. 29 United Nations, Human Rights Council, A/HRC/47/52 (note 28), para. 9.
30. 30 UN Human Rights Council Resolution 51/22, 7 Oct. 2022.
31. 31 United Nations, Human Rights Council, ‘Autonomous weapon systems’, A/HRC/56/CRP.5, 16 Apr. 2024, para. 14.
32. 32 United Nations, Disarmament Commission, Working Group II, ‘Chair’s summary’, 23 Apr. 2025, p. 1.
33. 33 Conference on Disarmament, Note verbale, CD/2334, 26 July 2023, annex, para. 37.
34. 34 Conference on Disarmament, 1582nd Plenary Meeting, Final record, CD/PV.1582, 15 June 2021, pp. 3 and 10.
35. 35 Blanchard, A. and Bruun, L., ‘Autonomous weapon systems and AI-enabled decision support systems in military targeting: A comparison and recommended policy responses’, SIPRI, June 2025.
36. 36 See CCW, GGE on LAWS, ‘Rolling text, status date: 12 May 2025’ (note 18), paras 6 and 7.
37. 37 UN General Assembly Resolution 79/239, 24 Dec. 2024.
38. 38 UN General Assembly Resolution 79/239 (note 37), p. 3.
39. 39 United Nations, ‘Security Council debates use of artificial intelligence in conflicts, hears calls for UN framework to avoid fragmented governance’, Meetings Coverage SC/15946, 19 Dec. 2024.
40. 40 US Department of State, Bureau of Arms Control, Deterrence, and Stability, ‘Political declaration on responsible military use of artificial intelligence and autonomy’, [n.d.]; and Elysée Palace, ‘Paris declaration on maintaining human control in AI enabled weapon systems’, 11 Feb. 2025.
41. 41 With the exception of initiatives associated with the US-led political declaration, which is focused on capacity-building among like-minded states.
42. 42 ‘Full statement: REAIM Blueprint for Action’, The Readable, 10 Sep. 2024, para. 10.
43. 43 CCW, Meeting of the High Contracting Parties, Final report, CCW/MSP/2023/7, 23 Nov. 2023, para. 20.
44. 44 United Nations, General Assembly, ‘Artificial intelligence in the military domain and its implications for international peace and security: Report of the Secretary-General’, A/80/78, 2 June 2025.
45. 45 Global Commission on Responsible Artificial Intelligence in the Military Domain, ‘Activities’, [n.d.].
46. 46 Convention on Cluster Munitions (CCM), opened for signature 3 Dec. 2008, entered into force 1 Aug. 2010.
47. 47 United Nations, ‘Two years on from war in Lebanon, progress on cluster munitions—UN agency’, UN News, 14 Aug. 2008; Human Rights Watch (HRW), Flooding South Lebanon: Israel’s Use of Cluster Munitions in Lebanon in July and August 2006 (HRW: 16 Feb. 2008); and Wiebe, V., Borrie, J. and Smyth, D., ‘Introduction’ in G. Nystuen and S. Casey-Maslen (eds), The Convention on Cluster Munitions: A Commentary (Oxford University Press: Oxford, 2010), paras 0.27–0.29.
48. 48 UN General Assembly Resolution 79/239 (note 37).
49. 49 van Rooijen, N., ‘Enabling machines to make life and death decisions is morally unjustifiable’, Civicus Lens, 20 June 2025.
50. 50 Convention on the Prohibition of the Use, Stockpiling, Production and Transfer of Anti-Personnel Mines and on their Destruction (APM Convention or Ottawa Convention), opened for signature 3–4 Dec. 1997, entered into force 1 Mar. 1999. For statements of withdrawal see, e.g., Ukraine Ministry of Foreign Affairs (MFA), ‘Statement of the MFA of Ukraine on decision to suspension from the Ottawa Convention’, Press release, 29 June 2025; Latvian Ministry of Defence, ‘Ministry of Defence proposes Latvia’s withdrawal from the Ottawa Convention’, Press release, 15 mar. 2025; and ‘Finland prepares to leave the Ottawa Treaty banning anti-personnel landmines’, Yle News, 1 Apr. 2025. See also the statement of Lithuania’s Minister of Defence on 15 July 2024 regarding the policy to withdraw from the Convention on Cluster Munitions (note 46), summarized in English in Žilinskas, J., ‘When security prevails: Lithuania votes to withdraw from the Convention on Cluster Munitions’, Articles of War, 13 Aug. 2024.
51. 51 E.g. the Constitution of the Philippines refers to the right to human dignity (section 11, Article II).
52. 52 ‘Breakthrough: Dutch parliament calls for international rules on killer robots’, Pax, 15 May 2019.
53. 53 Examples include the coalition agreement of the parties that formed Germany’s government in 2021 and the Norwegian government’s election platform for 2021–25. See, respectively, Social Democratic Party of Germany, Alliance 90/The Greens and Free Democrats, ‘Dare more progress: Alliance for freedom, justice and sustainability’, 2021, p. 137; and Labour Party and Centre Party, Hurdalsplattformen, 2021 (in Norwegian).
54. 54 CCW, Meeting of the High Contracting Parties, CCW/MSP/2023/7 (note 43); CCW Art. 8(1)(b) and (3)(b), referred to in the Rules of Procedure adopted by the Sixth Review Conference.
55. 55 UN Charter, Arts 11(1) and 13.
56. 56 Wiebe Borrie and Smyth (note 47), para. 0.69; and Casey-Maslen, S., The Anti-Personnel Mine Ban Convention: A Commentary, 3rd edn (Oxford University Press: Oxford, 2004), paras 0.43–0.44.
57. 57 UN General Assembly Resolution 79/239 (note 37), para. 7.
58. 58 Boulanin, V. et al., Limits on Autonomy in Weapon Systems: Identifying Practical Elements of Human Control (SIPRI and ICRC: Stockholm, June 2020).
59. 59 United Nations, General Assembly, A/80/78 (note 44), para. 66.
60. 60 For a discussion on the possible content of such an instrument see Bruun (note 17).
61. 61 Treaty on the Prohibition of Nuclear Weapons (TPNW), opened for signature 20 Sep. 2017, entered into force 22 Jan. 2021.
62. 62 Goldblat, J., ‘The biological weapons convention: An overview’, International Review of the Red Cross, vol. 37, no. 318 (1997); and Kimball, D. G., ‘The past, present and future of the Chemical Weapons Convention’, Speech at the Conference on Chemical Weapons, Armed Conflict, and Humanitarian Law Queens University, Kingston, Ontario, 29 Oct. 2018.
63. 63 CCW, GGE on LAWS, ‘United Kingdom proposal for a GGE document on the application of international humanitarian law to emerging technologies in the area of lethal autonomous weapon systems (AWS)’, Mar. 2022.
64. 64 See e.g. San Remo Manual on International Law Applicable to Armed Conflicts at Sea, 12 June 1994; Schmitt, M. N. (ed.), Tallinn Manual on the International Law Applicable to Cyber Warfare (Cambridge University Press: Cambridge, 2013); Schmitt, M. N. (ed.), Tallinn Manual 2.0 on the International Law Applicable to Cyber Operations (Cambridge University Press: Cambridge, 2017); Dinstein, Y. and Dahl, A. E., Oslo Manual on Select Topics of the Law of Armed Conflict: Rules and Commentary (Springer: Cham, 2020); Program on Humanitarian Policy and Conflict Research (HPCR) at Harvard University, HPCR Manual on International Law Applicable to Air and Missile Warfare (Cambridge University Press: Cambridge, MA, 2013); Beard J. and Stephens, D., The Woomera Manual on the International Law of Military Space Operations (Oxford University Press: Oxford, 2024); Newport Manual Export Group, ‘Newport Manual on the law of naval warfare, second edition’, International Law Studies, vol. 105 (2025); and The West Point Manual on the International Law Applicable to Artificial Intelligence in Warfare (Lieber Institute, forthcoming).
65. 65 Partington, E. A., ‘Manuals on the law of armed conflict’ in A. Peter and R. Wolfrum (eds), Max Planck Encyclopedia of Public International Law, Aug. 2016, para. 3.
66. 66 APM Convention (note 50), Article 7; CCM (note 46), Article 7; and Arms Trade Treaty, opened for signature 3 June 2013, entered into force 24 Dec. 2014, Article 13.
67. 67 United Nations, Programme of Action to Prevent, Combat and Eradicate the Illicit Trade in Small Arms and Light Weapons in All Its Aspects, adopted 20 July 2001. See United Nations, ‘Report of the United Nations Conference on the Illicit Trade in Small Arms and Light Weapons in All Its Aspects: New York, 9–20 July 2001’, 
    A/CONF.192/15, paras 22–24 and pp. 7–22.
68. 68 United Nations, General Assembly, ‘Terms of reference and modalities for the establishment and functioning of the Independent International Scientific Panel on Artificial Intelligence and the Global Dialogue on Artificial Intelligence Governance’, Draft resolution, A/79/L.118, 18 Aug. 2025.

Acknowledgments

SIPRI and the authors express their gratitude to the ministries for foreign affairs of the Netherlands, Sweden and Switzerland for their generous financial support for this publication.

The authors express their gratitude to interviewees for generously giving their time and expertise during research for this report. The authors thank SIPRI colleagues Vincent Boulanin, Laura Bruun, Dustin Lewis, Jules Palayer and Dan Smith for their invaluable insights and support throughout the writing of this report. The authors are also grateful for comments and feedback provided by Simon Cleobury and Rain Liivoja. Finally, the authors acknowledge the invaluable editorial work of the SIPRI Editorial Department.