Disagreement persisted for years even over whether an in-depth discussion of characterisation was necessary. Some delegations insisted on coming up with at least a working definition, taking the view that clarity on what technology was being discussed was essential for a meaningful debate about regulation. Other delegations took the opposite view, noting that in previous disarmament discussions (such as the one on cluster munitions), definitions were left to a later stage of treaty negotiations, and this did not hamper progress on developing a substantive regulatory framework.
The truth was probably somewhere in the middle. Insistence on definitional clarity overlooked the fact that the participants had some kernel of understanding of what kind of weapon systems were being discussed. After all, examples of specific systems periodically surfaced in the discussions, facilitated by the GGE bringing together diplomatic, legal and military expertise. At the same time, claims about the option to leave the definitional problem untouched ignored the fact that weapon systems subject to previous arms control discussions did not suffer from the same level of conceptual uncertainty. For example, the definition of prohibited cluster munitions could be finalised in the later stages of negotiations on the Convention on Cluster Munitions because there was no general uncertainty about what cluster munitions were, even though there was disagreement over how to regulate them and which cluster munitions should come under a prohibition.
The Nature of the Problem
In terms of the substantive discussion on the characterisation of LAWS that nonetheless ensued, there were two major challenges.
The first was technical. Autonomous functioning in weapon systems can have a high degree of technological complexity. Admittedly, other weapons subject to disarmament and arms control measures can also be very complicated. This would, for example, be true of nuclear weapons. But even with nuclear weapons, their overall effect, humanitarian consequences and challenges for international humanitarian law (IHL) compliance have not been especially difficult to articulate. Autonomous weapon systems, on the other hand, are not really a discrete class of weapons, unlike weapons that have previously been regulated internationally. Rather, autonomy is an emergent feature that can potentially characterise any weapon system. So, rather than dealing with a specific type of gadget, the GGE was confronted with a functionality—or perhaps even a method of warfare—which could not be easily defined or delimited.
The second challenge was political. Early proposals for a characterisation or definition fell broadly into two categories.
The first category essentially sought to define LAWS narrowly so as to capture, more or less, futuristic weapon systems where autonomy was potentially a source of serious concern, and thus required regulation. Some definitions contemplated a total lack of human intervention in the operation of the weapon system, for example: ‘a total absence of human supervision, meaning there is absolutely no link (communication or control) with the military chain of command’ (France, 2016), ‘absence of human intervention and control during the entire process of executing a task’ and ‘impossibility of termination’ (China, 2018) or ‘systems that completely exclude the human factor from decisions about their employment’ (Germany 2020). Some narrow definitions also underscored the ‘evolution’ of the system (China, 2018) or required the system to be ‘capable of understanding higher-level intent and direction’ (UK, 2017).
The second category of definitions sought to define LAWS broadly, including existing systems, which might or might not be a source of concern, and might or might not need additional regulation. Thus, these definitions described an autonomous weapon system as one ‘that, once activated, can select and engage targets without further intervention by a human operator’ (US, 2012/2017), ‘that can select (i.e. search for or detect, identify, track, select) and attack (i.e. use force against, neutralize, damage or destroy) targets without human intervention’ (ICRC, 2016), or ‘that, without human intervention, selects and engages targets matching certain predefined criteria, following a human decision to deploy the weapon on the understanding that an attack, once launched, cannot be stopped by human intervention’ (the Netherlands, 2017).
The States Parties favouring a narrow definition were keen to limit potential emerging regulation to systems that were not already in use, or which had no conceivable military utility. Some of them underscored repeatedly that LAWS, as understood by them, did not yet exist. The delegations favouring broad definitions were more concerned with setting some limits or contours for the debate than with defining the scope of a specific regulation.
One attempt to stop delegations talking at cross-purposes came in August 2021 with the Chair’s proposal to distinguish between ‘partially’ and ‘fully’ autonomous weapon systems, as had been suggested by France. This proposal did not resolve the problem, as the degree of autonomy did not map precisely onto risks and regulatory proposals.
The shift that allowed the characterisation problem to move towards a solution came in 2021–2022 with the introduction and articulation of the two-tier approach. This approach contemplated the prohibition of some LAWS (tier 1) and the regulation of other LAWS (tier 2), while leaving the exact demarcation between the tiers open. The broad support for this approach, even among States that generally favoured new prohibitions, potentially reassured other States that a comprehensive ban on all LAWS was no longer on the agenda. Thus, identifying a baseline characterisation for all LAWS lost much of its political sensitivity.
The Characterisation
The characterisation of LAWS in the elements is a complex, multi-layered affair. The first layer consists of the chapeau of paragraph 25 of the GGE’s 2024–2026 Report, which states that:
a lethal autonomous weapon system can be characterized as a combination of one or more weapons and technological components operating together, that is capable of identifying, selecting, and engaging a target in a way that results in loss of life, without intervention by a human being operating the system.
The second layer consists of three clarifications and saving clauses that follow in subparagraphs (a), (b) and (c) of paragraph 25. Like the characterisation itself, these were the result of intense discussions during the three-year mandate of the GGE and were adopted by consensus.
The third layer consists of four understandings, which the Chair delivered orally on the final day of discussions. While these do not appear to have been expressly adopted by consensus in the same way as the text of the elements themselves, the lack of any objection to the understandings by States Parties indicates that they indeed reflect the common understanding of the GGE. Applying an analogy from treaty interpretation—specifically Article 31(2)(a) of the Vienna Convention on the Law of Treaties—these understandings form part of the ‘context’ that would need to be taken into account when reading the GGE’s report.
Some aspects of the characterisation proved difficult not so much because the States Parties disagreed about the objective as because they struggled to find the right combination of words to clearly express what they meant. An example of this was the notion of a ‘weapon system’, which some considered to be unclear, and not without reason. The NATO Terminology Database, for example, defines a weapon system extremely broadly as ‘[a] combination of one or more weapons with all related equipment, materials, services, personnel and means of delivery and deployment, if applicable, required for self-sufficiency’. The inclusion of personnel as an element of a weapon system would create significant complications for the characterisation of LAWS. Alternative understandings of a weapon system appear to exclude personnel. For example, the Australian Guide to the Legal Review of New Weapons suggests that a weapon system includes weapons and ‘all related equipment, materials, and delivery mechanisms necessary for a weapon to function as intended’. This seems to come closer to what the GGE had in mind. The Russian delegation’s concerns with the notion of ‘weapon system’ point to a further linguistic or doctrinal difference. The Russian term that comes reasonably close to the narrower, technology-focused understanding of a weapon system is ‘weapon complex’ (kompleks vooruzheniya), whereas the Russian term ‘weapon system’ (sistema vooruzheniya) may refer to the aggregate of weapon complexes designed to carry out a specific function (such as an air defence system).
In any event, the difficulties around ‘weapon system’ triggered a prolonged search for more descriptive language, ultimately resulting in the phrase ‘a combination of one or more weapons and technological components operating together’. Interestingly, at the very end of the discussions, the notion of weapons and other components ‘operating together’ replaced the earlier phrasing of a ‘functionally integrated’ combination of weapons and other components. This change was intended to address the concerns of delegations who found the idea of functional integration too vague. But it might actually have loosened the level of integration required even further than the ‘functionally integrated’ formulation.
The idea that autonomy entails the ability of the system to perform certain functions without real-time human input similarly proved challenging to articulate. The wording needed to capture both the idea that the characterisation related to what occurs after the activation of the system and the idea that it is the intervention of the user or operator (rather than that of any old human) that matters. Thus, the text evolved from ‘without further intervention by a human operator’ (July 2024) to ‘without intervention by a human user in the execution of these tasks’ (November 2024 and May 2025) to ‘without intervention by a human operator in the execution of these tasks’ (December 2025) to ‘without intervention by a human operating the system’ (June 2026). What finally met consensus was the phrase ‘without intervention by a human being operating the system’.
The two major substantive differences related, however, to the range of ‘critical functions’ of the system, which it needs to be able to perform without human intervention in order to be considered autonomous, and the ‘lethality’ of the system. These two issues merit a closer look.
Identify, Select, Engage
The starting point for the discussion was the idea, reflected in the early US and ICRC definitions of autonomous weapon systems, that the critical functions are, cumulatively, target selection and engagement. The logic here cannot be faulted: if the system cannot select a target without human intervention, it is not really operating autonomously, and if it cannot engage a target without human intervention, it is more of a decision-support system.
Some States Parties, perhaps most notably Singapore, took the view that target identification was a critical function for the purposes of characterising a LAWS. The first question this raised was whether target identification and selection were cumulative or alternative requirements. The Group eventually settled on the former. Again, the rationale appears to be that a weapon system that can identify targets, but not select specific targets for the purposes of engagement, does not operate autonomously.
The introduction of target identification as a cumulative criterion, however, risked unduly narrowing the characterisation. Such a characterisation, it was argued, might only capture systems that can develop their own target profiles, and exclude the more likely case of systems that match persons or objects encountered in the environment against a pre-set target profile. The solution was to include, in what became subparagraph 25(a), a clause explaining that a system can be a LAWS even if ‘a human being programs or otherwise determines a set of potential targets or characteristics thereof’.
Conversely, the same subparagraph now also clarifies, with a view to excluding precision-guided munitions and the like from the characterisation of LAWS, that a system is not a LAWS if ‘the personnel planning, deciding upon, or carrying out the attack determine the specific target or specific group of targets and the timing of the attack’. Instead of a ‘specific group of targets’, multiple delegations proposed referring here to ‘a group of specific targets’. While this would have been both clearer and more in line with existing law (notably the reference to ‘specific military objective’ in Article 51(4) of Additional Protocol I), the proposal was ultimately defeated.
The distinction between a ‘set of potential targets’ and a ‘specific group of targets’ in the first and second sentences respectively can now give rise to confusion, which the Chair sought to dispel with an understanding. As explained by the Chair, the former refers to a ‘category or class of targets, within which the system itself determines which objects to identify, select and engage’, while the latter means ‘a grouping of targets’ that military personnel have identified, such as ‘a particular formation’.
Thus, a loitering munition that can engage, in a particular area, any target meeting a specific profile pre-determined by a human, such as the known electromagnetic signature of radar stations, would fall within the characterisation of LAWS. Conversely, a sensor-fused anti-tank munition that is launched by a human towards a specific column of tanks would be an example where a human determines a ‘specific group of targets and the timing of the attack’, such that the munition falls outside the characterisation of LAWS.
Lethality
The single most controversial aspect of the characterisation was the requirement of lethality. Many delegations took the view, persistently and over many years, that such a requirement would be anomalous. General rules of IHL relating to means and methods of warfare, and the conduct of hostilities, do not apply differently based on the lethal or non-lethal effects of weapons. Moreover, any such distinction would be difficult to apply in practice, as the precise effect of a weapon would depend substantially on the circumstances of use. There was also some concern that the reference to lethality might focus the instrument on anti-personnel weapons and potentially exclude anti-materiel weapons that raise comparable legal and humanitarian concerns.
But some States Parties, notably Russia and China, insisted on the characterisation reflecting lethality. China’s position, although consistently held, did not really offer a rationale. For Russia, the matter was ostensibly linked to the title and mandate of the GGE relating to ‘lethal autonomous weapon systems’. This position, however, overlooked the fact that the Group’s mandate strictly speaking referred to ‘emerging technologies in the area of lethal autonomous weapon systems’, which would have allowed for a broader approach.
The problem of inadvertent exclusion of anti-materiel systems was conclusively resolved first by incorporating subparagraph (b), clarifying that: ‘The fact that a LAWS can be used in a way that does not result in loss of life, such as to damage or destroy objects or to cause injury, does not exclude it from this characterization.’ But the inclusion or otherwise of lethality as an element of the characterisation proved divisive until the very end. The issue was ultimately resolved by adding the words ‘capable of … engaging a target in a way that results in loss of life’, which reportedly occurred during an informal session on the last day of the GGE.
The sensitivity of this issue is also reflected in the set of understandings articulated by the Chair in a formal session late in the afternoon of 4 September. In particular, these understandings emphasise that the overall characterisation and subparagraph (b) ‘are to be read together, as a single package’. Moreover, the characterisation is ‘framed in terms of [the] capability’ of the system—that is, ‘what the system is capable of doing’—and ‘does not turn on the purpose for which a system was designed, or on its primary function’. This stands in contrast to the definitions of an incendiary weapon in CCW Protocol III and of blinding laser weapons in CCW Protocol IV, which refer, respectively, to what the weapon is ‘primarily designed’ or ‘specifically designed’ to do.
This does not, however, fully resolve the interpretive problem. Consider a weapon system, such as a sentry robot, that shoots rubber bullets with the purpose of repelling people. Assume also that the energy transferred by these bullets does not ordinarily cause death or injury, but may do so when hitting a child or hitting an adult in the head. As the GGE’s understanding is agnostic about the purpose or primary function, the matter turns on whether the rubber bullets have the capacity to result in the loss of life. The answer seems to be yes. If so, this aspect of the characterisation might exclude only very limited categories of weapons, such as some types of directed energy weapons.
Conclusion
The ability of the GGE to adopt, by consensus, a characterisation of LAWS is no small feat. The intricacy of the resulting text, and the amount of time and effort dedicated to crafting it, testify to the sensitivity and the complexity of the task at hand. The characterisation adopted now establishes the material scope of application of the elements of an instrument.
That having been said, the achievement should not be blown out of proportion. For one, the question as to which LAWS are prohibited remains an altogether separate matter. Thus, the practical significance of the characterisation lies in determining which weapon systems should be seen as belonging to tier 2 (regulation) in the two-tier approach to governing LAWS. Moreover, many of the measures contemplated in the elements of an instrument would also be useful with respect to weapon systems that fall outside the characterisation (such as risk-mitigation measures contemplated in paragraph 38) or would simply be inherently irrelevant to such systems (for instance, biases addressed in paragraph 39).
