At 4:30 in the morning on Saturday in Geneva, diplomats from 128 countries announced their conclusion: two more years to think about it. After a week of negotiations on autonomous weapons at the UN Convention on Certain Conventional Weapons, the CCW's seventh review conference closed without any commitment to legally binding rules on AI-powered systems that can select and kill human targets on their own.
The outcome followed a script critics said they had seen before. Talks that Secretary-General António Guterres opened by calling "the most complex and consequential challenge in the history of warfare" produced a document that lets countries suggest future regulations for the next two years, binding or otherwise. The systems the talks were meant to govern keep deploying faster than the diplomacy designed to slow them.
How 128 Countries Spent a Week Disagreeing in Geneva
The CCW's seventh review conference ran November 16 to 21, 2026, at the Palais des Nations in Geneva. The formal agenda was straightforward: should states open negotiations on a legally binding treaty to restrict lethal autonomous weapons systems, or on something less? In practice, negotiations dragged past Friday and into the small hours of Saturday morning before delegates found language everyone could accept.
According to UN News, the meeting addressed more than 120 member states and all 128 High Contracting Parties to the CCW, a 40-year-old treaty framework covering weapons that cause unnecessary suffering. Guterres told the opening session that autonomous weapons systems capable of making life-and-death decisions without human oversight represent a fundamental break from every conflict before them, adding a call for "clear prohibitions and restrictions" on such systems. UN High Commissioner for Human Rights Volker Türk went further: "A machine must never have the power to decide whether a human being survives or dies."
The language in those opening speeches had little bearing on what followed. The NPR report on the outcome describes an impasse that forced delegates to work through the night. The final text gives states two years to propose possible regulations, which the agreement specifies "could be legally binding, but could also be optional." The ICRC (International Committee of the Red Cross) had entered the talks demanding binding rules within a single year. What it got was an open-ended suggestion process with an uncertain destination.
Why the US, Russia and Israel Blocked a Treaty
The talks began badly on day one. Al-Monitor reported that Russia objected to the meeting's agenda before substantive negotiations could begin, consuming hours that delegates had expected to use on the central question. Its formal position on autonomous weapons has never committed to legally binding constraints, and state officials have previously described the moment as premature for new international legal obligations.
The United States entered the conference with its own proposal: a five-year window for discussions toward nonbinding, optional guidelines. That position, reported by Newsweek in advance of the talks, reflected Washington's standing argument that existing international law already covers how nations conduct warfare, making a bespoke AI weapons treaty redundant. Military lawyers in the US Department of Defense have long maintained that Directive 3000.09, which requires appropriate levels of human judgment for autonomous weapons, satisfies the obligation.
Israel's position aligned with the US and Russia. Its Geneva mission stated that it supports multilateral discussion and deploys data technologies in full accordance with international law. The scrutiny on Israel intensified after its AI-assisted targeting systems in Gaza drew public criticism, though representatives disputed that any system removes human decision-making from lethal choices entirely. Alaska Dispatch News documented an earlier phase of this dynamic in September, when the US and Russia spent roughly 15 hours modifying a CCW draft document to soften language on human accountability and ethical safeguards, removing passages the original text's proponents argued were essential minimums.
Against the three major opponents, support for a binding treaty is broad but fragmented. The campaign group Stop Killer Robots counts approximately 130 countries in favor of binding rules, with only about a dozen opposed. France, Germany, Ireland, Italy, Brazil and Mexico were among those pressing for a firm timetable. The structural problem is that the CCW operates by consensus: a single state can prevent any outcome. That architecture, designed for the 1980s, was not built to withstand a moment when the countries with the most advanced autonomous weapons development programs also have the clearest incentive to slow regulation of them.
80% of Sudan's Civilian Deaths Fell to AI-Targeted Drones This Year
The gap between the Geneva process and the actual use of autonomous weapons is not abstract. A PBS NewsHour report on ICRC warnings before the talks cited UN figures showing that in Sudan's ongoing conflict, 80% of civilian deaths this year were caused by drones. Many of those drones use AI-assisted target selection. The figure comes from the UN Office of the High Commissioner for Human Rights and was entered into the Geneva record as evidence for why the ICRC's urgency was grounded in observable casualties, not theoretical risk.
Ukraine and Gaza present a parallel track. Both conflicts have produced documented use of AI-assisted target recognition, and both have surfaced failures: systems misidentifying civilian objects, loitering munitions operating beyond the operational control of human commanders, and AI-powered electronic warfare tools that cannot be recalled once launched. Military researchers have noted publicly that the 72-hour processing latency between target identification and strike approval, which exists in some command systems, is incompatible with the kind of real-time autonomous systems that are now commercially available to smaller states and non-state actors at a fraction of earlier costs.
The scale of the market underlines why time matters. Polaris Market Research values the autonomous weapons systems market at $12.3 billion in 2025, rising to $29.8 billion by 2034. Military drone deployment is growing faster: MarketsandMarkets estimates that segment alone at $34.85 billion in 2026, expanding to $109.22 billion by 2031. Those figures represent systems already in procurement pipelines. A two-year discussion process that produces optional guidelines in 2028 means those systems will be in service, tested in multiple conflicts, and deeply embedded in defense doctrine long before any binding rule applies to them.
The Two-Year Clock and What It Actually Commits to
The practical content of the Geneva outcome is thinner than it sounds. Countries have until approximately the end of 2028 to submit proposals on how autonomous weapons might be regulated. Those proposals could be legally binding, but could also be optional. There is no mechanism to compel any state to submit anything, no secretariat to review submissions for consistency, and no agreed definition of what counts as an autonomous weapons system in the first place. That last gap has paralyzed CCW discussions since at least 2014, when the topic first appeared on the agenda, more than 12 years ago.
Anna Hehir, a military AI governance expert at the Future of Life Institute, told NPR before the talks concluded that absent a binding agreement, "We're likely to see the systems proliferate." The absence of a binding definition means any country can describe its AI-assisted weapons as not autonomous, since "autonomous" has no agreed technical threshold. That definitional vacuum is not a bug in the CCW process. For the states opposed to binding rules, it is a feature.
The ICRC's Georgia Hinds described the accountability gap in precise terms: "You cannot hold a machine accountable for a violation of the law." The Geneva Conventions create liability for individuals who order unlawful attacks. No existing legal framework creates comparable liability for a software system that selects targets, nor for the chain of developers, procurement officers and commanders through whom it reached the battlefield. The CCW framework was the most plausible venue to begin addressing that gap. The two-year deferral leaves it open.
An unnamed delegate at the talks, speaking to NPR after the conclusion, said the outcome amounted to "just kicking the can down the road." Campaigners from Stop Killer Robots called it "a real disappointment" and "a nothing burger." Negotiators defended the text as proof that the UN disarmament process could still function. Those two framings cannot both be true if the measure of success is fewer civilian deaths from autonomous targeting systems in the next two years.
The Non-Obvious Problem: Commercial Drones Already Outpace the CCW
The CCW was built to address weapons that states develop and deploy. The harder challenge for autonomous weapons governance is that the underlying technology is now commercial. Consumer drones modified with computer-vision AI for target identification have appeared in Ukraine. Loitering munitions based on open-source flight controllers and off-the-shelf image-recognition models are being assembled by non-state actors for under $500. The CCW has no jurisdiction over commercial technology manufacturers, and it has no mechanism to address non-state actors at all.
This gap explains why the ICRC's demand for a binding treaty, while morally coherent, was also insufficient even on its own terms. A CCW agreement binds the 128 High Contracting Parties. It does not bind private arms dealers, technology exporters operating outside treaty jurisdictions, or the growing number of semi-state militias that now operate advanced drone fleets. The proliferation Hehir described is not primarily a state procurement problem. It is a commercial technology diffusion problem, and the CCW is the wrong instrument to solve it alone.
The more productive governance frame that received little attention in Geneva is export controls. The US Bureau of Industry and Security issued an interim rule in November requiring export licenses for advanced AI chips when there is knowledge they will supply compute to companies in restricted destinations. That rule, narrow as it is, does something the CCW cannot: it establishes a licensing mechanism enforceable through penalties on US exporters, regardless of whether the recipient state has signed any treaty. It is imperfect and covers only one part of the technology stack. But it is enforceable in a way the Geneva outcome is not.
The Bear Case for Doing Nothing Has a Name: Military Advantage
The bear case for the status quo, however, is not simply obstruction. The US Department of Defense position is that autonomous systems reduce civilian casualties by improving accuracy and removing the fatigue-related targeting errors that human operators accumulate over long campaigns. That argument is contested by humanitarian researchers but is not fabricated: in some controlled military studies, AI-assisted target recognition has achieved lower false-positive rates than exhausted human operators working under time pressure.
Critics argue that those studies measure performance under conditions that do not resemble actual conflict: controlled imagery, single-object targets, and operators not simultaneously managing communications, logistics and incoming fire. The risk the Geneva talks were trying to address is not that AI systems will perform worse than humans in ideal conditions. It is that they will perform worse than humans in the chaotic, context-rich, rapidly shifting environment of real combat, where the difference between a combatant and a civilian may hinge on a behavior no training dataset ever labeled.
Skeptics point out that the CCW process itself carries a hidden cost. By providing a forum that appears to be addressing the problem, it absorbs diplomatic energy that might otherwise go toward the smaller, more tractable steps that could actually reduce harm: a ban on autonomous weapons with no meaningful human control in populated areas, mandatory incident reporting for autonomous targeting failures, or a moratorium on weapons whose effects their developers cannot fully model. None of those required a treaty. All of them died in Geneva alongside the binding mandate.
The next CCW meeting is not yet scheduled. The states that blocked a binding mandate this week will return in two years to review whatever proposals emerge from an open-ended suggestion process. By then, the military drone market will be closer to $60 billion. The systems will be more capable, more widely deployed and more deeply integrated into military doctrine. The window for establishing binding norms before that hardware becomes legacy infrastructure closes a little more each time the CCW agrees to meet again.
A machine must never have the power to decide whether a human being survives or dies, and two more years of optional discussions changes nothing about whether one already does.
Key Takeaways
- No binding rules: The CCW's seventh review conference ended November 21 without any commitment to legally binding restrictions on autonomous weapons, deferring the question for at least two years.
- US proposed five years, got two: Washington entered with a proposal for a five-year nonbinding discussion window; the final text gives states two years to propose regulations that could be binding or optional.
- 80% of Sudan's civilian deaths: UN figures cited in the ICRC's pre-conference submission show that 80% of civilian casualties in Sudan's 2026 conflict were caused by drones, most using AI-assisted targeting.
- $109B drone market by 2031: MarketsandMarkets estimates the military drone segment will reach $109.22 billion by 2031, outpacing any governance timetable from Geneva.
- No agreed definition after 12 years: The term "autonomous weapons system" has no binding international definition despite CCW talks running since 2014, leaving every state free to classify its own systems as it chooses.
Questions Worth Asking
- If the states that oppose binding rules are also the states developing the most advanced autonomous systems, can any UN consensus-based treaty process ever produce meaningful constraints on this technology?
- The US argument is that existing international law already covers autonomous weapons. If that is true, why has no state ever been prosecuted under it for an AI-assisted targeting failure?
- Commercial autonomous drone technology costs under $500 for a non-state actor to assemble with AI target recognition. What governance mechanism, short of controlling the underlying components, could make that weapon illegal to use?