Symposium on the CCW GGE LAWS: Once in a Lifecycle-Same as It Ever Was in the GGE’s Elements

[Netta Goussac is an Associate Senior Researcher at the Stockholm International Peace Research Institute and Associate Partner at Lexbridge. Netta has taken part in the GGE since 2017 and was a member of the Australian delegation during the second session in 2026. The views expressed in this post are her own and do not represent the Australian Government or other entity. This post is the fifth in a series of analyses on aspects of the recent adoption of the 2024-2026 CCW GGE LAWS report.] 

Paragraphs 36 to 39 of the final report of the 2024-2026 sessions  of the Group of Governmental Experts on lethal autonomous weapons systems (GGE on LAWS) set out ‘lifecycle measures’. These measures – legal reviews, testing and evaluation, training, bias mitigation – facilitate a state’s adherence to the prohibitions and constraints on use of LAWS set out in preceding paragraphs. States are to take such measures in the study, design, development, acquisition and adoption – that is, at earlier stages of the ‘lifecycle’ of an autonomous weapon than the point of use, which is the primary focus of the preceding paragraphs.

Like the others in this series, this piece explains the provenance and significance of these elements with the aim of assisting states and others in the policy debate. 

The ‘lifecycle’ of autonomous weapons

The concept of a weapon’s lifecycle ‘dawned’ in this debate in 2018 and was reflected in the GGE’s 2019 Guiding Principles. Principle (b) states that human responsibility for decisions on the use of autonomous weapons should be considered across the lifecycle, and principle (c) states that human-machine interaction may be implemented at various stages. The term was not defined, but principle (g) indicates states’ thinking, specifying that risk assessments and mitigation measures should be part of the ‘design, development, testing and deployment cycle of emerging technologies in any weapons systems’ (emphasis added). Australia made a similar contribution to the Group in 2019, describing a nine-stage ‘system of control’ from the legal and policy framework to after-action evaluation. 

The elements under discussion in the GGE followed a lifecycle of their own. The July 2024 rolling text carried a short list under the heading ‘other measures to ensure compliance with [international humanitarian law]’. By November 2024 this had become a discrete block introduced by a chapeau: measures to be taken ‘prior to potential employment and as applicable throughout the entire life cycle of LAWS’. By June 2026 the block had been split, with two things states ‘must’ do and the remainder qualified by ‘should… as appropriate, across the life cycle of LAWS’.

The adopted elements contemplate four kinds of lifecycle measures: legal reviews (paragraph 36), the consideration of risks (paragraph 37), measures to ensure that capabilities and limitations can be understood (paragraph 38), and measures against bias (paragraph 39). A fifth measure, in paragraph 42, is taken up at the end of this piece.

Legal reviews

The legal review of new weapons has been consistently present in this part of the text, appearing in every iteration since July 2024. The function of legal reviews has never been in doubt: it is how the prohibition on developing and using systems incapable of lawful use is given practical effect.

The first sentence of paragraph 36 reads:

In accordance with States’ obligations under applicable international law, in the study, development, acquisition, or adoption of a new LAWS, determination must be made whether its employment would, in some or all circumstances, be prohibited by international law.

The sentence is the same as guiding principle (e) as agreed in 2019, but for the substitution of ‘a new LAWS’ for ‘a new weapon, means or method of warfare’ and the insertion of ‘applicable’.

The consistency in language from 2019 to 2026 reflects diversity in the applicable legal framework. Not every State Party to the Convention on Certain Conventional Weapons (CCW) is party to Additional Protocol I of the Geneva Conventions, whose article 36 contains an obligation to conduct legal reviews. At the same time, some non-parties to Additional Protocol I review capabilities as a matter of domestic policy – the United States and Israel most prominently. Principle (e) was drafted in 2019 precisely to state the obligation only for states that already have it. The qualifier ‘in accordance with their obligations under international law’ moved in and out of the rolling text, and the verb oscillated between ‘must’ and ‘should’, but in the end the GGE returned to the 2019 formula.

This does not mean that paragraph 36 is not innovative. No CCW protocol requires or recommends legal review as a measure of implementation, and neither does the Anti-Personnel Mine Ban Convention nor the Convention on Cluster Munitions. Any instrument incorporating the element in paragraph 36 would be the first weapon-specific instrument to restate this general rule of international humanitarian law (IHL). 

Whether it does much work is another matter. Few states are known to maintain a standing review mechanism for any capability, and LAWS are a demanding means of warfare to review.

A practical difficulty discussed by the GGE is when to review a LAWS that has been modified. Software is easier to change than hardware, and these systems are software-controlled. Sub-paragraph 39(b) of the Chair’s suggestions had addressed this directly, requiring review of any further development or modification with a ‘substantial impact on the operational functioning’ of the identification, selection and engagement functions, or on anticipated effects (emphasis added). That paragraph did not survive adoption; nor did the related sub-paragraph 38(c) that would have prevented mission parameters being substantially modified without human authorisation. During discussion of this aspect of the rolling text, some states preferred ‘material impact’ over ‘substantial’. The United States sought either to delete the paragraph or to replace it with its own formulation, confined to modifications that could render use prohibited or substantially change the circumstances of lawful use. This was despite the US Department of Defense Directive 3000.09 describing a domestic policy-based approval process that requires senior review of autonomous weapon systems whose ‘system algorithms, intended mission sets, intended operational environments, intended target sets, or expected adversarial countermeasures substantially differ from those applicable to the previously approved weapon systems’. Similarly, US Army Regulation 27-53 requires a legal review if there is ‘a substantial change to the intended use or anticipated effects’ of a weapon or ‘substantial enhancements or modifications’ to a weapon.

That the adopted text says nothing about modification at all is a consequential deletion, and a missed opportunity to provide helpful guidance to states, the need for which others and I have identified through expert consultations.

What did survive is the encouragement of voluntary exchange of relevant practices, bearing in mind national security considerations and commercial restrictions on proprietary information. It is the only part of paragraph 36 that is new relative to 2019, and lends support to exchange initiatives – like those run by the International Committee of the Red Cross, the Stockholm International Peace Research Institute, the Asia-Pacific Institute for Law and Security, and the UN Office for Disarmament Affairs  – that have so far drawn modest participation.

Considering and mitigating the risks of LAWS

Paragraph 37 requires that during design, development, testing and deployment, the risks of civilian casualties and the precautions that minimize incidental loss of civilian life, injury and damage ‘must be considered’, with other risks – unintended engagements, loss of control, proliferation, acquisition by terrorist groups – to be considered ‘as appropriate’. The internal split is between risks of non-compliance with IHL, where consideration is framed as mandatory, and security risks, framed as discretionary.

The purpose of considering these risks is to mitigate them. Understanding a system’s capabilities, limitations and anticipated effects is what IHL rules on the conduct of hostilities presuppose: the prohibition on indiscriminate means of warfare (article 51(4)(c) of Additional Protocol I), the proportionality rule (article 51(5)(b)) and the obligation to take precautions in attack (article 57). Paragraphs 31, 33 and 34 of the elements rest on that assumption. The two measures listed in paragraph 38 are how those obligations are made operational. Notably, paragraph 38 is expressed as risk mitigation, and as ensuring that capabilities and limitations can be understood rather than that they are.

Testing and evaluation (in sub-paragraph 38(a)) has been reflected in the text throughout the drafting process, but its stated purpose has shifted. In July 2024 it was to give a human operator a reliable expectation of how the system would perform. That purpose was deleted in December 2025, and testing became a means of ensuring understanding. The adopted paragraph 38(a) restores a performance-assessment purpose, though not the original one: testing and evaluation, ‘including adequately realistic simulations’, is now deemed necessary ‘to assess how the LAWS will perform in the expected operational environment’. ‘Adequately’ qualifies ‘realistic’, presumably because perfect simulation of dynamic operating environments is not available. ‘[T]he expected operational environment’ is in the singular, replacing the Chair’s ‘variable operational environments’. The adopted formulation narrows the purpose of testing back towards the anticipated circumstances of use, rather than the variability of the operational environment that makes such systems difficult to test.

The reference to ‘guidance, training and instructions’ in subparagraph 38(b) highlights the resources that lawful use of LAWS requires. The addressees of this element have changed throughout the GGE’s mandate: from human operators, to operators and those in the responsible chain of command, to those in control of LAWS, and now to the broadest formulation: ‘relevant personnel, including those who bear responsibility in the chain of human command and control’.

Paragraph 38 no longer requires, as the Chair’s suggestions would have, that LAWS operate in a predictable and reliable manner and that their functioning and effects be explainable and traceable. Delegations divided on whether these were legal or policy terms; the adopted text resolves the argument by omission.

Bias, machine and human

Paragraph 39 asks states to implement measures to prevent and mitigate unintended bias in the functioning of LAWS, including bias in data sets, artificial intelligence models and other technological components, and, separately, automation bias in the use of LAWS.

Bias is where the broader military artificial intelligence (AI) conversation has most visibly reshaped the GGE conversation on LAWS. It was raised in the Group as early as 2019, but the vocabulary now in the text (data sets, models, automation bias) is imported from AI governance and appears in documents such as the US-led Political Declaration on Responsible Military Use of Artificial Intelligence and Autonomy and the Pathways to Action document of the 2026 Responsible AI in the Military Domain (REAIM) Summit.  

The paragraph also reflects a small drafting change with large implications. The qualifier changed from ‘harmful’ bias to ‘unintended’ bias in May 2025. The Chair’s summary records that most delegations welcomed the change as more neutral, while some observed that a degree of bias is intrinsic to distinguishing targets. IHL does not ask for a system without discriminating tendencies. It demands certain discriminations – between combatants and civilians, between military objectives and civilian objects – and tolerates others so long as they do not produce unlawful effects.

(An annotation on the somewhat odd formulation in the chapeaux of paragraphs 38 and 39: ‘In order to mitigate the risks related to LAWS, including those related to the implementation of applicable IHL…’ (emphasis added). The formulation was drafted in the informal consultations of 3 September 2026, which are not on the record. Its provenance is not completely obscure, however. That day the United States expressed reservations about the three paragraphs that became paragraphs 38 and 39, on the ground that they ‘lack a clear nexus showing the purpose for which we are doing these kinds of measures’, and Pakistan supported linking them to ‘the true purpose, ensuring compliance with IHL’.)

Closing the loop between pre- and post-use measures

A further set of lifecycle measures sits outside paragraphs 36 to 39. Paragraph 42 invites states to promote human responsibility and accountability in the use of LAWS, and to mitigate the risk of unintended engagements, through understandable human-machine interfaces, guidance and instructions consistent with applicable IHL, and training of personnel to understand both those instructions and the capabilities and limitations of the system’s autonomous functions.

These are measures decided before use: an interface is a design decision, doctrine and rules of engagement are written before deployment, and the training in paragraph 42(c) mirrors the training in paragraph 38(b). Their location in the accountability cluster of elements is instructive, because it shows the same measures can do two different kinds of work. Leading up to an attack, they are what enables those who plan and decide upon it to comply with the requirements of distinction, proportionality and precautions. After one, they are what makes it possible to establish what happened. They are what a state needs if it is to investigate and, where warranted, prosecute war crimes, whether under the grave breaches regime of the Geneva Conventions and Additional Protocol I or through the national arrangements on which the complementarity regime of the Rome Statute depends.

What states do before use determines what they are able to know afterwards. The internal mechanisms for addressing incidents encouraged by paragraph 44 of the report  are where what they learn should return to the beginning: to the next policy framework, the next training package, the next set of rules of engagement.

Lifecycle measures as benchmarks for implementation

That the lifecycle measures serve aims established earlier in the text does not make them any less important. They convert a legal obligation or policy commitment into practical and tangible actions by states. They are the standards by which the effectiveness of any instrument on LAWS will be judged, and the starting point for any actor wishing to drive behavioural change at the national level.

They also reflect an outcome with which the GGE is rarely credited. Much of what these paragraphs describe is already happening in the handful of states that develop, acquire or use autonomous weapons. In the years since the GGE commenced its work in 2017, these states have elaborated policies, guidelines and standards to guide their militaries’ development and use of autonomous weapons. These and other states have in some cases publicised their interpretation of IHL and other relevant legal obligations with respect to autonomous weapons. The GGE’s role in prompting this elaboration of national practice and informing states’ interpretation and application of international law is one of its durable achievements.