The Convention on Certain Conventional Weapons (CCW) has always been a place where governments talk about the future of warfare. That’s part of its purpose: to anticipate new categories of weapons and decide, collectively, what rules should apply before those weapons become routine. But when Branka Marijan attended the CCW sessions in Geneva in November 2017, she expected the meeting to feel like the familiar rhythm of international diplomacy—careful, cautious, and largely hypothetical.
Instead, she found something more unsettling. The discussion that year was still framed around lethal autonomous systems, often described in the language of “killer robots” and speculative scenarios. Yet the atmosphere shifted quickly from abstract concern to practical urgency. What had seemed like a distant possibility—autonomous or semi-autonomous systems making life-and-death decisions—was being treated as something that could arrive sooner than many participants wanted to admit.
That change in tone matters, because it reveals how quickly the policy conversation can lag behind technological capability. It also shows how international law debates don’t just respond to inventions; they respond to momentum. When the momentum becomes visible inside government delegations—when attendees begin to talk less about whether autonomous weapons will exist and more about how they might be used—the entire structure of the debate changes. The question stops being “Could this happen?” and becomes “How do we prevent the worst outcomes if it already is happening, or is about to?”
At the CCW, the setting is formal and procedural: twice-yearly sessions in Geneva under the UN umbrella, with working groups, statements, and negotiated language that moves at the speed of consensus. But beneath that surface, the underlying contest is about control—control over what counts as acceptable autonomy, control over how responsibility is assigned when machines act, and control over the pace at which militaries adopt systems that can operate with reduced human oversight.
In 2017, Marijan’s experience captured a turning point. The meeting was not merely discussing a theoretical future. It was confronting a present reality: military AI systems were already improving, already being integrated into sensing, targeting, and decision-support tools, and already being tested in ways that blurred the line between assistance and action. Even when delegates spoke in careful terms, the direction of travel was hard to miss. The “red lines” people referenced—what should never be allowed, what must remain under meaningful human control—were no longer purely rhetorical. They were being stress-tested by the capabilities that were emerging outside the conference rooms.
To understand why the CCW’s discussions have become so consequential, it helps to recognize what the forum is trying to do. The CCW is not a single treaty that covers everything. It is a framework for specific protocols and norms related to conventional weapons, including those that raise humanitarian concerns. Lethal autonomous systems sit at the intersection of two difficult problems: the ethics of delegating lethal force to machines, and the legal challenge of ensuring accountability when decisions are made by systems that may not be fully interpretable by humans in the moment.
Autonomy is not a binary switch. In practice, military systems can range from highly supervised tools—where humans select targets and authorize engagement—to systems that can identify, track, and propose actions with varying degrees of independence. The policy challenge is that the more autonomy increases, the harder it becomes to guarantee that human judgment remains meaningful rather than ceremonial. A commander might technically “approve” an action, but if the system has already narrowed options, predicted behavior, and selected targets based on patterns humans cannot verify in real time, then approval risks becoming a formality.
This is where the CCW debate becomes more than a philosophical argument. It becomes a question of operational design. Delegates are effectively asking: what level of human involvement is required for compliance with international humanitarian law? How do you ensure distinction between combatants and civilians when the system’s perception and classification are probabilistic? How do you assess proportionality when the machine’s estimate of collateral effects is opaque or incomplete? And if something goes wrong, who is responsible—the developer, the operator, the commander, the state?
Those questions are not new, but AI changes their texture. Traditional weapons can be evaluated through relatively stable performance characteristics. AI-enabled systems, by contrast, can behave differently across environments, can be sensitive to data quality, and can produce outputs that are difficult to trace back to a simple rule. Even when a system is designed to follow constraints, the real-world conditions of conflict—uncertain signals, degraded communications, adversarial interference—can push it into edge cases.
That’s why the CCW’s shift from hypotheticals to practical capability is so important. When the conversation stays hypothetical, it can afford to focus on worst-case fears and broad principles. When it becomes practical, it must grapple with implementation details: how systems are trained, how they are tested, how they are deployed, and how they are constrained. It must also confront the fact that militaries do not adopt technology in a vacuum. They adopt it within doctrines, procurement cycles, and strategic incentives. If one state believes it can gain advantage by moving faster, others face pressure to keep up—even if they are uneasy about the ethical implications.
In that sense, the CCW debate is also about deterrence and arms control dynamics. Autonomous weapons raise a particular kind of escalation risk. Faster decision loops can compress the time available for human review. If a system can detect and engage without waiting for extended authorization, then the window for de-escalation shrinks. That doesn’t automatically mean autonomous weapons are inherently uncontrollable, but it does mean that the traditional escalation ladder—where humans can pause, verify, and respond—may be altered by machine speed.
There is also the issue of ambiguity. If multiple systems can perform similar functions—surveillance, tracking, target suggestion—then distinguishing between “assistance” and “engagement” becomes harder for observers. That ambiguity can complicate verification and trust-building. Arms control agreements rely on shared understanding of what is being limited and how compliance is measured. With AI systems, compliance can be difficult to verify because the same hardware and software architecture can be configured differently depending on mission parameters.
The CCW’s work, therefore, is not only about banning certain categories of weapons. It is about shaping norms around how lethal autonomy should be used, and about building a common vocabulary for what counts as unacceptable autonomy. That vocabulary is contested. Some states emphasize the need for a ban on fully autonomous lethal systems. Others argue for regulation rather than prohibition, focusing on safeguards, transparency, and human control. Many positions fall somewhere in between, reflecting different threat perceptions and different views of what “human control” can realistically mean in modern warfare.
What makes the 2017 shift notable is that it suggests the debate was arriving at the same conclusion many technologists and defense analysts were reaching: the relevant question is not whether AI will be used in war, but how quickly it will be integrated into the kill chain. Once AI becomes part of sensing and decision support, it can gradually expand its role. A system that begins by recommending targets can evolve into a system that selects targets under certain conditions. A system that begins by operating in constrained environments can be adapted for more complex ones. The boundary between “tool” and “actor” can move.
This is why the CCW’s discussions have become a kind of early warning system. They show how governments interpret the trajectory of AI in military contexts. They also show how quickly international forums can be forced to confront realities that domestic politics and procurement timelines may already be accelerating.
The Verge’s reporting on this topic underscores that the CCW sessions are not merely academic. Attendees watch demonstrations, discuss technical capabilities, and listen to arguments that increasingly reference real-world systems rather than purely imagined ones. Even when the language remains cautious, the subtext is clear: the “future” is arriving in increments, and those increments are enough to test legal and ethical boundaries.
One unique aspect of the CCW debate is that it takes place in a space where humanitarian law is assumed to be the baseline. International humanitarian law already requires distinction, proportionality, and precautions in attack. The CCW’s challenge is to translate those principles into guidance that fits the behavior of autonomous or semi-autonomous systems. That translation is difficult because the law was written for human decision-making and human accountability, while AI systems can make decisions that are not easily explainable in the moment.
This is where the concept of “meaningful human control” becomes central—and controversial. Meaningful control is not just about having a human in the loop. It is about whether the human can understand what the system is doing, predict its behavior, and intervene effectively. If a system operates at machine speed, the human may not have time to evaluate whether the system’s target identification is correct. If the system’s reasoning is opaque, the human may not know why it selected a target. If the system’s training data is biased or incomplete, the human may not be able to compensate for those limitations during deployment.
In other words, meaningful control is both technical and procedural. It depends on interface design, on training and doctrine, on testing regimes, and on the ability to override or abort actions. It also depends on whether the system’s autonomy is bounded by reliable constraints. A system that can be overridden instantly and reliably is different from a system that continues to act even after a human attempts to stop it.
The CCW’s discussions reflect these complexities. Delegations are essentially negotiating what kinds of autonomy are compatible with humanitarian law and what kinds are not. But because AI systems can be configured in many ways, the debate can become a moving target. A safeguard that works for one system might not work for another. A definition that seems clear in principle can become ambiguous in practice.
That ambiguity is one reason why some advocates push for stronger restrictions. They argue that the unpredictability of AI in dynamic environments makes it impossible to guarantee compliance with humanitarian law at scale. They also argue that the more autonomy increases, the more responsibility becomes diffuse. If a system makes the final decision, then the human operator may be reduced to a supervisor