Written by Adele Klassen, a recent graduate of the University of Ottawa’s MA in Security and Defence Studies

Under the Convention on Certain Conventional Weapons (CCW), an international treaty framework through which states prohibit or restrict certain conventional weapons, states began discussing lethal autonomous weapon systems in 2014 and established a formal Group of Governmental Experts (GGE) on the issue in 2016. In September, states participating in the GGE agreed on elements that could inform a future international instrument on lethal autonomous weapon systems (LAWS), including a shared description of which systems fall within this category, along with provisions on human control, legal review, testing, accountability, and other safeguards.

The central concern lies in paragraph 25, which describes LAWS while also excluding some systems from that category. A system falls outside the characterization when people determine the specific target, or specific group of targets, and the timing of the attack, even if autonomous functions continue to shape how the weapon identifies, tracks, approaches, or engages that target.

Canada supported the final consensus, but its own earlier work at the GGE took a broader approach, focusing directly on autonomous functions rather than excluding systems when people have already chosen the target and timing of the attack. If states move to negotiations, they will need to decide whether the systems excluded by paragraph 25 should also be excluded from any additional rules on autonomous weapons.

What paragraph 25 leaves outside

Paragraph 25 states: “Within the scope of the application of the CCW, 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.” Identification can involve recognizing that a person or object matches relevant characteristics, selection involves determining which target will be acted against, and engagement involves applying force against the selected target.

The report makes clear that people can program a system with potential targets, or with characteristics that potential targets must meet, without automatically removing it from the LAWS category. A system instructed to locate and attack vehicles matching particular characteristics, for example, could remain within the description if the system determines which specific vehicles to select and engage.

Paragraph 25(a), however, excludes a system when people determine the specific target, or specific group of targets, and the timing of the attack. During the final session, the Chair explained that, for the purposes of the agreed wording, a “specific group of targets” referred to a group already identified by people, such as a particular military formation, rather than a broad category from which the system later chooses individual targets. The clarification narrows the exclusion, but systems still fall outside the characterization when humans make those decisions in advance.

Canada participated directly in shaping this wording, supporting the UK proposal to refer to the circumstances of engagement, while also indicating that it could accept the US proposal to refer to the timing of the attack. The final understanding addresses concerns about overly broad target categories, but it does not settle how systems should be treated when autonomous functions remain important after the target and timing have been chosen by people.

What the exclusion could leave out

Choosing a target is an important exercise of human judgement, but it does not necessarily determine everything a weapon does after activation. Depending on the system, autonomous functions may still recognize or reacquire a target, track it as conditions change, navigate toward it, distinguish it from surrounding objects, or influence how force is applied.

Later provisions in the report address limits on targets, duration and geographical scope, whether people can intervene in a system’s operation, how systems should be tested, and whether their capabilities and limitations are understood. The report also addresses unintended engagements, bias in data or artificial intelligence models, and automation bias, which occurs when people place too much trust in an automated system. The 3 September revised draft went further, requiring autonomous identification, selection and engagement functions to operate with adequate reliability, and their functioning and anticipated effects to be predictable and traceable, but this language was removed from the final report. These concerns depend on how autonomy operates in practice, including after an initial human targeting decision has been made.

Stop Killer Robots has warned that exclusions in the characterization could create loopholes or support narrow interpretations of future rules, particularly if paragraph 25 is treated as the final limit of what regulation can cover. Paragraph 25(c) leaves room to address this concern, explicitly allowing for future changes to the characterization, including changes to which systems are included or excluded.

The consequences of a narrower scope

Weapons outside paragraph 25 remain governed by international humanitarian law, which regulates how armed conflicts are fought. Article 36 of Additional Protocol I to the Geneva Conventions requires states that are party to the treaty, including Canada, when studying, developing, acquiring or adopting a new weapon, means or method of warfare, to determine whether its use would, in some or all circumstances, be prohibited by applicable international law.

The broader international process is now considering whether, and how, autonomous weapons should be subject to additional rules. Depending on what states eventually negotiate, those rules could address prohibited systems or uses, required forms of human control, limits on operating environments or targets, testing and predictability, or measures related to accountability and transparency. If paragraph 25 is eventually used to determine which systems a future instrument covers, weapons excluded from the characterization could also fall outside those additional requirements.

Canada has previously supported a more function-focused approach

A 2025 GGE proposal co-sponsored by Canada shows that it has previously supported a broader way of assessing autonomous weapons, one that focused directly on autonomous functions rather than only on whether a system fell within a particular category. In September 2025, Canada joined ten other states in submitting proposed draft articles on autonomous weapon systems. The proposal addressed weapons that, once activated, could identify, select and engage targets with lethal force without further intervention by an operator, and it did not contain the express target-and-timing exclusion now found in paragraph 25(a).

The proposed articles also applied legal and operational requirements directly to autonomous functions, stating that the lawfulness of relying on those functions could depend on their expected performance, available alternatives consistent with customary military practice, and the urgency of the situation. When assessing proportionality, commanders were also expected to consider relevant and available information, which could include how the weapon’s autonomous identification, selection and engagement functions would perform.

Canada’s 2025 submission also reproduced earlier GGE conclusions recognizing that autonomy can exist at different points in the targeting process, that different functions within the same weapon can have different degrees of autonomy, and that a working characterization should not predetermine policy choices. These earlier conclusions placed greater emphasis on how autonomy operates across different functions and stages of the targeting process, an approach that could capture concerns not fully addressed by paragraph 25(a)’s target-and-timing exclusion.

The 2026 report still contains safeguards concerning testing, understanding system capabilities and limitations, and human control, but those safeguards apply within the LAWS category created by paragraph 25. Canada’s previous position therefore provides a clear basis for asking whether future negotiations should retain a wider functional focus, including for systems that fall outside paragraph 25(a).

What comes next

The final report recommends that the Seventh Review Conference of the Convention on Certain Conventional Weapons, scheduled for 16 to 20 November 2026, consider and endorse the agreed elements and decide what form future work should take. Treating paragraph 25 as settled could carry its exclusions into whatever rules follow.

One option available to Canada is to support future work that preserves paragraph 25(c)’s flexibility, allowing states to revisit the characterization as they negotiate substantive rules. This could include examining whether systems excluded under paragraph 25(a) still rely on autonomous functions in ways that raise similar humanitarian concerns addressed elsewhere in the report, and whether some future restrictions or safeguards should apply to them.

Canada’s 2025 proposal already provides a basis for a broader, function-based approach, while paragraph 25(c) leaves the current characterization open to revision. At the November Review Conference, Canada could support future work that keeps this flexibility intact, particularly where systems outside paragraph 25 continue to rely on autonomous functions that raise similar humanitarian concerns.