Machine warfare is coming, and the problem is no longer the weapons but the processes that govern them

By Anadolu Ajansı · October 21, 2026.

Robin Geiss, director of the United Nations Institute for Disarmament Research (UNIDIR), said on Tuesday that machine warfare is "getting ever closer." He spoke at a press conference at UN headquarters in New York. The question he posed was blunt: "Are we going to allow machines to target and kill human beings?"

According to Geiss, AI and autonomy are already present in intelligence analysis, target identification, drones, cyberattacks and information operations. He added that decision cycles are being compressed "to machine speed" and that the room for human judgment is steadily narrowing, including in the nuclear domain.

The UNIDIR director also noted that the proliferation of these capabilities lowers the barrier to entry for highly sophisticated operations, which could allow non-state actors to carry them out. In his view, a handful of countries and companies are shaping security outcomes in ways that traditional arms control frameworks were not designed to address.

On the institutional front, UNIDIR will launch on Thursday, together with Switzerland and Pakistan, the new forum of its Centre of Excellence on AI, Peace and Security. Geiss described it as the UN's only multi-stakeholder body working on AI in international peace and security and in the military domain. He announced three more initiatives over the course of the week. The aim, he said, is to identify gaps and propose concrete measures without limiting access to peaceful uses of AI.

**Our take.** The most valuable part of his remarks is a line that almost goes unnoticed: the military AI race is "a race for speed," focused on shortening the sensor-to-shooter time (the interval between detecting a target and striking it). Geiss warns that the greater risk is not losing control of a specific weapon, but of entire military processes. It is a sharper thesis than the classic debate about "killer robots."

It matches what we argue about agentic AI: the real danger is near-term and stems from automation and speed, not from some distant superintelligence. If competitive advantage means deciding faster, the human "in the loop" ends up as a formality, just like the user who approves permissions out of habit without controlling anything. Approving is not controlling, whether in a coding agent or in a chain of command.

The geopolitical nuance also matters. Geiss acknowledges that tension between blocs is a serious obstacle, but argues that the concern now transcends those blocs and is moving the major powers. Caution is warranted: this is his assessment, and a UN forum without binding power is not a treaty. Classic arms controls relied on countable objects (warheads, missiles). A model or decision software cannot be counted or inspected the same way, and that is precisely what Geiss admits the current framework does not know how to handle.

In the short term, the winners are those who can integrate AI into their military processes fastest, and the loser is meaningful human oversight. The lower barrier to entry for non-state actors adds a risk that no agreement among powers would solve on its own.

Our position remains the same: evidence-based governance, not panic. That requires concrete, verifiable measures, such as maintaining real human control over the lethal decision and over the processes that precede it, not just over the trigger. It is also a good example of why long-term optimism is not automatic. The abundance AI promises depends on the gravest decisions remaining human during the transition. The UN's attempt to distinguish between preventing misuse and not holding back civilian benefits is a reasonable balance, though it remains to be seen whether it will have teeth.