
Governments have ended a major round of talks in Geneva without agreeing to begin negotiations on binding rules for autonomous weapons. The outcome leaves a widening gap between the speed at which militaries are developing systems that can select and attack targets and the pace of international diplomacy.
In a 21 November statement, the International Committee of the Red Cross (ICRC) urged states to establish legally binding limits. Its warning followed the review conference of the Convention on Certain Conventional Weapons, where governments did not agree on a mandate to negotiate an instrument. The failure to start treaty negotiations is not the same as an agreement that existing law does not apply, or a decision to ban the systems.
The conference discussed how to address autonomous weapons but stopped short of launching formal negotiations on new binding rules. States can continue discussing the subject through the convention’s processes. The ICRC argues that talks alone are no longer enough as the technology moves towards real deployment.
The organisation noted that 163 states supported a resolution in the UN General Assembly’s First Committee calling for negotiations on a legally binding instrument. That figure signals broad support for action, but it does not mean 163 countries have signed a treaty or agreed on its wording. The unresolved questions include which systems should be prohibited outright and what level of human control should be required for others.
The ICRC’s explanation of autonomous weapon systems focuses on weapons that, once activated, can select and engage targets without a person choosing each individual target. That distinction matters. A remotely piloted drone whose operator decides when and what to strike is not automatically an autonomous weapon. Nor does every autonomous weapon need to use modern generative AI.
AI can nevertheless make these systems more capable, including by helping them interpret sensor data or identify objects in complex surroundings. It can also make their behaviour harder to anticipate or explain. The policy issue is therefore not simply whether a weapon carries an AI label. It is whether a human can understand, limit and take responsibility for the choice to use force.
Humanitarian law already requires parties to armed conflict to distinguish civilians from military targets, avoid indiscriminate attacks and take precautions to limit civilian harm. An autonomous system does not remove those obligations. Yet if a machine makes a target selection in an environment that changes after activation, commanders and operators may struggle to predict a specific strike or explain why it happened.
The ICRC is pressing for rules that prohibit unpredictable autonomous weapons and systems designed to target people, alongside strict limits on other systems. Critics of a binding treaty tend to favour applying existing law and developing operational guidance, while some states also worry that broad definitions could restrict defensive systems. The gap is as much about where to draw the line as whether human judgement is important.
Military AI is wider than targeting. TechBooky has reported on AI deployment in classified US defence networks, which is a separate use from allowing a weapon to select a target. Treating every military AI programme as a lethal autonomous weapon would obscure the specific legal and technical problem negotiators are trying to solve.
Governments can pursue further work under the convention, build support through the United Nations or develop national limits. For now, there is no newly agreed global treaty text from the Geneva meeting. The meaningful milestones to watch are a formal negotiating mandate, agreed definitions, proposed prohibitions and rules that specify how people retain control over the use of force.
The absence of a new agreement does not pause research or procurement. That is why the ICRC says the diplomatic timetable matters. Each year without binding rules raises the risk that weapons will be fielded before states have settled what decisions should never be delegated to machines.







