.png)
The missing rulebook for military AI: drawing the line before the machines do
What happens to a Saker Scout drone when it loses its communications link mid-mission over Ukraine’s front lines? Nothing. It keeps flying.
Autonomous Weapons Systems (AWS) are generally defined as weapons that can select and engage targets without a human in the loop, with NATO’s Air Power experts going as far as to characterise them by their ‘consciousness and self-determination’. AI is entering the battlefield faster than its creators can predict its behaviour, and far faster than the law can respond.
States are not waiting for permission to deploy it.
Yet no single legally binding multilateral instrument governs the use of autonomous weapons. Oversight remains fragmented, resting on a patchwork of voluntary commitments, alliance-scoped arrangements and governance instruments that exclude the military entirely. This regulatory void is, at best, a temporary feature of a technology still in transition; at worst, it actively threatens crisis stability, sustained by unenforced standards.
Turn to the statutes and the omission is glaring. Both the EU AI Act and the Council of Europe's Framework Convention on Artificial Intelligence explicitly leave defence applications outside their scope. NATO’s principles bind no one outside the alliance; meanwhile, two rival coalitions dominate the debate. The first, a ‘tight regulation’ camp working through the Convention on Certain Conventional Weapons, has spent over a decade in talks to produce only a draft of basic prohibitions that 52 states called a minimum standard. The second, a ‘responsible use’ camp visible at the 2023 REAIM Summit in The Hague, produced a US-led Political Declaration that sought only to build a diplomatic consensus.
It would be reductive to ignore the tactical appeal of military AI. AI can operate in communications-denied environments, analyse copious quantities of data in seconds, and identify vulnerabilities in an adversary’s systems. Each advantage, however, carries a weighty risk. When black-box decision-making obscures how a weapon behaves, the scale and speed of automated decision support can reduce human judgement to a rubber stamp.
Artem Martynenko, a lieutenant colonel in the Ukrainian military, notes that the “hybrid” and “grey zone” nature of modern battlefields demands “processing the context”, which remains a challenge for AI-driven systems. A former Palantir employee describes analysts who would rather “go home or go get a cheeseburger” than keep up with reviewing “list[s] of 1000 targets” churned out by the machine. In the first 38 days of the Iran war, the US struck 13,000 targets with the help of AI-supported targeting. One was a school in Minab where more than 150 people, most of them children, were killed. Whether an algorithm pulled the trigger is unproven; the more pressing question is what happens to the margin of error when systems built on statistical correlation process targets at a speed no human review can match.
Targeting errors aren’t new. But AI multiplies their pace and volume, and with them the risk of indiscriminate attacks, damage to civilian infrastructure, and unintended escalation.
The fragmentation is not incidental: in a climate of escalated great-power tension, states see military AI as a decisive strategic asset, one Putin has claimed would make its leader “the ruler of the world”; fewer governments are willing to forfeit an asset that could decide the future balance of power. Treaties, moreover, bind only those governments that sign them, leaving private actors unbound. Anthropic’s threat-intelligence report this month found its model Claude being misused across China, Russia, Ukraine, Iran, Yemen, Sudan, and Taiwan to develop weapons towards autonomous drone swarms. Then there is a technical objection: if arms control cannot verify what these systems will do in every scenario, what is the point of a binding law that cannot be enforced?
I do not dismiss these realist arguments. But my question is: what now? Are we going to let military AI advance unmanned and unchecked until it triggers an escalation that no human can contain?
Anything is better than sitting idly by, and history shows us that claims of impossibility can be overcome. The Strategic Arms Limitation Talks highlight how rival superpowers could agree to incremental caps and oversight on chemical weapons at the height of the Cold War. Anthropic CEO Dario Amodei has proposed permanently embedded third-party evaluators inside frontier labs, urging the US and China to pace their AI competition in an approach comparable to SALT treaties.
We cannot aim for perfection on the first try, but we can build on what already exists. The UN’s existing governance ecosystem can move from voluntary ethical guidance into a unified, legally binding framework. A proposed International Artificial Intelligence Agency - independent, science-based and representative - would give these efforts a global home, a central evidence base for common standards, and a means of building state regulatory capacity across states worldwide.
The Saker Scout will not be the last drone to fly without human input. Those sceptical are right to believe that verification is difficult, that definitions keep evolving and that states are reluctant to sign away their advantages. None of this changes the fact that no algorithm should be able to carry the moral weight of the decision to kill.
States face a choice: wait for a perfect agreement or accept the status quo, where a regulatory void leaves the door open to unchecked military technology. It is between building an institution while it is still possible, and living under whatever a machine decides instead.
No image changes made.
Comments