The Last Human Decision
The first arms-control test of the artificial intelligence age is no longer theoretical. The question is whether the final decision to take a human life can disappear into a system no citizen can question.
The meeting rooms in Geneva rarely look like the future. They look like translation booths, nameplates, legal phrases and patient diplomats moving through brackets in a draft text. But this week, as Le Monde reported, consultations under the Convention on Certain Conventional Weapons have become the place where the world is trying to decide whether military artificial intelligence should have a legal boundary before it becomes normal.
The timing matters. At the G7 summit in Evian-les-Bains, political leaders and technology executives were discussing frontier AI, security and economic power. In Geneva, diplomats and civil-society groups were asking a narrower and more severe question: what must never be delegated to software in war?
The old phrase is "lethal autonomous weapons systems." It is accurate and insufficient. The issue is not only a robot firing without a person touching a button. It is the whole chain of war becoming partially automated: surveillance, pattern recognition, target selection, collateral-damage estimation, route planning, drone coordination, escalation analysis, kill recommendations and after-action justification. The decision may still formally belong to a human, but the human may be staring at a machine-shaped answer inside a machine-shaped tempo.
That is why the last human decision matters.
The speed trap
Modern militaries do not adopt AI because they are careless. They adopt it because war is getting faster, more data-rich and more distributed. A commander cannot manually interpret every drone feed, satellite pass, intercepted signal and battlefield sensor. AI promises speed, pattern detection and coordination across systems that already exceed human attention.
There are real defensive uses. A system can identify incoming missiles faster than a human watch officer. It can help clear mines, monitor airspace, detect cyber intrusions or warn civilians of attacks. A blanket fear of "military AI" is too crude to be useful.
The danger begins when speed becomes an argument against judgment.
If one side believes it must automate to avoid being outpaced, the other side will do the same. If both sides build systems that compress warning times and recommend action inside seconds, restraint becomes harder. The human is not removed all at once. The human is rushed, surrounded and subordinated. The machine does not need to seize command. It only needs to make hesitation feel irresponsible.
That is the speed trap. It turns oversight into theatre.
The accountability gap
War law rests on human responsibility. A soldier, commander, minister or head of state can be asked why a target was selected, whether civilians were protected, whether proportionality was assessed, whether distinction was respected, whether alternatives existed.
An autonomous system cannot answer morally. It can produce logs, probabilities, model outputs, confidence scores and audit trails. Those may help. They are not responsibility.
The accountability problem becomes acute when several actors share the chain: a private model developer, a defence contractor, an intelligence agency, a field commander, an allied military, and an automated targeting platform trained on data no single person fully understands. When the wrong building is hit, who is answerable? The officer who accepted the recommendation? The engineer who built the classifier? The procurement official who certified the tool? The minister who authorised deployment? The model that cannot be punished?
If responsibility becomes too distributed to locate, impunity becomes technical.
That is not a legal detail. It is the difference between a rule-governed war and a violence system that can apologise without anyone being guilty.
Geneva's narrow chance
The Convention on Certain Conventional Weapons is not glamorous, but it has precedent. It has been used to restrict or prohibit weapons judged excessively injurious or indiscriminate. The framework is imperfect, slow and consensus-heavy. Yet it is one of the few places where militarily powerful states, smaller states, humanitarian organisations and legal experts can place weapons technology inside a common rulebook.
Le Monde reported that more than half of the roughly 130 participating states support moving toward formal talks on military AI. That is not yet a treaty. It is not even a guarantee of negotiations. But it is movement.
The minimum standard should be clear. A system must not select and engage human targets without meaningful human judgment. The phrase "meaningful" matters. A person clicking approve after a machine has already framed the battlefield is not enough. A commander must have time, information, authority and the practical ability to refuse. Otherwise the human is not in control. The human is decorative.
The second standard should be explainability at the level of law, not marketing. If a state cannot explain why a system identified a target, how it assessed civilian risk and who had authority to halt the strike, that system should not be used for lethal force.
The third standard should be traceability. Every AI-enabled military decision must leave an audit path usable by courts, investigators and democratic oversight bodies. War already tempts secrecy. AI cannot be allowed to deepen it.
The company problem
The military AI debate is also a corporate-governance debate. Frontier AI systems are not built only inside defence ministries. They come from companies whose models, infrastructure and talent sit in civilian markets but can be adapted for military use. This creates a difficult triangle: states want capability, companies want contracts and control, publics want safety, and no one has yet settled who gets to draw the red lines.
It is not enough to rely on corporate terms of service. A company can refuse one military use today and accept another tomorrow. It can be pressured by a government. It can be replaced by a rival with fewer scruples. Ethical branding is not arms control.
Nor is it enough for states to say "trust us." Democratic oversight of national-security technology is already weak in many countries. Classified procurement and wartime urgency can hide decisions until after norms have been set by use.
The only durable answer is public law. Companies need clear boundaries. Militaries need clear obligations. Citizens need to know that the authority to kill has not been quietly privatised or automated.
Why this is a democratic issue
Military AI is often framed as a strategic issue, and it is. It is also a democratic issue.
In a democracy, the use of force is supposed to be politically accountable. Citizens may not vote on each operation, but they can judge leaders, challenge policy, protest wars, investigate abuses and demand legal constraint. If lethal decisions become opaque machine processes, democratic control weakens.
The citizen cannot deliberate about a system that is described only as "advanced." The court cannot test a targeting process hidden behind proprietary code. The parliament cannot oversee a war machine if the machine's behaviour is explained after the fact by people who themselves do not fully understand it.
The danger is not that AI makes war emotionless. War has always found ways to dull human feeling. The danger is that AI makes war administratively frictionless. A strike becomes a workflow. A pattern becomes a target. A recommendation becomes action. A death becomes an output.
Peace requires friction. It requires hesitation, argument, burden of proof and the knowledge that a person will be answerable for what is done.
The line to draw
The world will not ban all military AI. It will not stop armies from using algorithms for logistics, defence, analysis or reconnaissance. Pretending otherwise would make the law irrelevant before it begins.
But the world can still draw a line around the final decision to kill. It can say that human beings may use tools, but tools may not become moral agents. It can say that speed is not a substitute for legality. It can say that accountability cannot be automated away.
That line will be difficult to enforce. Some states will resist. Some will hide. Some will argue that their adversaries cannot be trusted, and therefore restraint is naive. That argument has accompanied every arms-control effort in history. It is sometimes sincere. It is also how humanity talks itself into every race it later regrets.
The Geneva process may fail. It may produce only soft language. It may be weakened by powers that want freedom of action. But the attempt matters because the default is not neutrality. The default is deployment first, norms later, grief last.
The last human decision must remain human. Not because humans are wise enough, but because only humans can be responsible enough. A machine can calculate. It cannot answer for the dead.
The Global Federation treats peace as a design problem: how to build institutions, laws and habits that keep force answerable to humanity before force learns to answer only to itself.