Who Decides What Gets Hit? The Global Battle Over AI Weapons

A Russian drone strike that killed three civilians at a gas station in Zaporizhzhia on July 6 has raised questions about a new stage in warfare: weapons that can identify and select targets without a human operator making the final decision. A subsequent investigation by The New York Times, reported that the drone used artificial intelligence to select a specific target after human operators directed it towards the gas station.

Ukrainian military commanders, drone experts and forensic investigators who examined recovered wreckage identified an Nvidia Jetson Orin computer and software that appeared to use visual recognition to select the target, believed to be propane tanks. Ukraine’s state news agency Ukrinform initially reported two deaths and nine injuries; the later investigation reported that the death toll had risen to three.

The findings illustrate the questions raised when a weapon can proceed from human-directed flight to selecting the specific object it will attack.

The issue will feature in international discussions on lethal autonomous weapons systems at the United Nations Convention on Certain Conventional Weapons (CCW) review conference in Geneva from November 16 to 20. Governments must consider whether existing laws governing armed conflict are sufficient or whether new international restrictions are needed as military AI advances.

What Makes A Weapon Autonomous?

The defining feature of an autonomous weapon is not that it flies without a pilot, but that it can select and engage a target without further human intervention after activation.

The International Committee of the Red Cross (ICRC) defines autonomous weapon systems as weapons that, once activated, can select and apply force to targets without further human intervention. Such systems use sensors to detect characteristics such as heat, movement, shape or speed, then compare them with a predefined target profile.

This distinguishes three broad categories. A remotely piloted drone remains under human control, even if software helps it navigate or identify objects. An AI-assisted system may analyse images, track targets or recommend an engagement while leaving the decision to fire with an operator. An autonomous weapon can complete critical targeting functions without a person selecting the specific target at the time of attack.

Military and policy discussions often describe these arrangements as human-in-the-loop, human-on-the-loop and human-out-of-the-loop systems. However, these are not universally agreed legal classifications. The actual degree of control depends on the system’s design, how it is deployed and whether an operator can intervene before an engagement.

The distinction matters because AI is already used for tasks that do not make a weapon autonomous. The ICRC notes that a system using AI only for navigation or intelligence analysis does not necessarily qualify as an autonomous weapon. The central question is who makes the decision to apply force.

How Are AI And Electronic Warfare Changing Warfare?

Artificial intelligence can process battlefield imagery, identify equipment, track movement and shorten the time between detecting a target and acting on it. These capabilities can help militaries respond more quickly, particularly when large volumes of information must be analysed under combat conditions.

Ukraine’s Ministry of Defence describes AI as a means of improving the military “kill chain”, the sequence from finding and tracking a target to selecting a means of engagement and assessing the result. Its stated approach uses AI to support human operators by improving target recognition and speeding up decisions.

Electronic warfare adds another complication. Jamming can interfere with radio communications and satellite-navigation signals, disrupting drones that depend on those connections. Some systems can use onboard cameras, visual recognition and other navigation methods to continue operating when GPS or communications are degraded.

But resistance to jamming does not itself make a drone an autonomous weapon. A system may navigate towards a destination without human guidance while still requiring an operator to select and approve the target. Conversely, a weapon can have limited autonomy in a specific targeting function without relying on sophisticated AI.

Moreover, the technical issue relies on how much control a human retains over the choice of target, the decision to engage and the weapon’s effects once activated.

How Are Major Militaries Using AI?

The evidence and public disclosures vary across countries, making it important to distinguish battlefield reporting from official policy and systems still under development.

In Russia’s case, the Zaporizhzhia strike is the central reported example. The NYT investigation linked the recovered hardware to a drone system that appeared capable of choosing a precise target through onboard visual recognition. The assessment came from Ukrainian officials, specialists and forensic examination, rather than a public Russian confirmation of autonomous targeting.

Ukraine, meanwhile, has publicly described its development of AI-assisted systems for reconnaissance, interception and decision support. In June, its defence ministry reported that a new generation of interceptor drones could autonomously carry out much of the interception process against incoming Shahed-type uncrewed aircraft. The ministry said the operator selected the target and issued the engagement command, after which the system could guide and engage the incoming drone.

In October, Ukraine’s defence ministry also announced AI tools developed with NATO’s Joint Analysis, Training and Education Centre to predict enemy drone attacks and help plan artillery operations. These applications support military decisions but are not equivalent to a system independently choosing a person or civilian object to attack.

The United States has an established policy framework for autonomous and semi-autonomous weapons. The Department of Defense’s Directive 3000.09 requires appropriate human judgement over the use of force and sets out review requirements for systems that could cause unintended engagements. The existence of that policy, however, does not establish that all US systems operate with the same degree of human involvement.

China’s Ministry of National Defense said in March 2026 that human primacy should be upheld in military AI and that relevant weapons must remain under human control. Beijing has also supported negotiations through the UNCCW.

Israel has said that its military uses data-processing tools to help intelligence analysts cross-reference information and identify material relevant to target assessments. The Israel Defence Forces maintains that human analysts identify targets, authorised officers approve them and operational commanders decide whether to strike.

The comparison shows that military AI covers a range of functions, from analysing intelligence to intercepting incoming threats and selecting targets. Treating all these uses as equivalent would obscure the specific risks posed by systems that apply lethal force without human intervention.

Why Are Countries Debating Global Rules?

Existing international humanitarian law applies to autonomous weapons, including rules requiring parties to distinguish between military objectives and civilians, avoid attacks expected to cause excessive incidental civilian harm, and take feasible precautions to protect civilians.

The difficulty is applying those requirements when a weapon selects its own target. A human operator may activate a system based on a general target profile without knowing precisely which object it will engage or when. In a changing battlefield, a vehicle may be misidentified, civilians may enter the area or a person’s status may change before the weapon strikes.

The ICRC argues that such systems raise serious humanitarian, legal and ethical concerns, especially when their behaviour is difficult to predict or their effects cannot be effectively limited. It has called for legally binding rules to prohibit unpredictable autonomous weapons and systems designed to target humans, alongside strict restrictions on other autonomous weapons.

On August 25, the UN Secretary-General and the ICRC president renewed their call for states to adopt international rules. Their appeal comes as the CCW’s Group of Governmental Experts continues work on possible elements of an international instrument.

The November 16–20 review conference is a significant opportunity for governments to narrow differences over prohibitions, restrictions and the role of human control. It is not, however, a guaranteed deadline for a treaty. States have yet to reach consensus on the form and scope of a binding instrument.

The central disagreement is whether existing humanitarian law, national safeguards and non-binding guidance are sufficient, or whether the risks posed by autonomous targeting require specific international prohibitions and restrictions.

What Is India’s Position?

India has argued for an approach that addresses the use and effects of lethal autonomous weapons rather than placing a blanket prohibition on military AI.

In a working paper submitted to the CCW Group of Governmental Experts in May 2025, India said fully autonomous systems that cannot be used in compliance with international humanitarian law should be prohibited, while other systems should be subject to appropriate regulation.

India’s position emphasises human judgement and oversight, a clear understanding of what constitutes autonomy, and the continuing responsibility of human operators. It also distinguishes fully autonomous weapons from systems operating with human-in-the-loop or human-on-the-loop control, as well as uncrewed vehicles as a category.

This approach reflects the distinction between the use of AI in military operations and delegating the critical decision to use lethal force to a machine. India recognises that AI can support military capabilities, but argues that regulation should focus on whether a system can be employed consistently with international humanitarian law.

As governments prepare for the November conference, the challenge will be to translate broad principles into rules that can be applied across systems with different levels of autonomy. That includes defining which weapons should be prohibited, what forms of human control are necessary and how responsibility should be established when a system causes unlawful harm.

What Is At Stake In November?

The Zaporizhzhia strike illustrates why the distinction between human-supervised and autonomous targeting matters. If a system can select the specific object it attacks, responsibility for ensuring that the target is lawful and that civilian harm is limited becomes more difficult to assess before the weapon is activated.

The November review conference will test whether governments can agree on new prohibitions and restrictions or continue working through existing international humanitarian law and non-binding guidance. Any agreement will need to account for the variety of systems already in use, from defensive interceptors to weapons capable of selecting targets independently.

Military AI is already used in reconnaissance, target recognition, interception and decision support. The unresolved question is where states should draw the line between automating battlefield functions and allowing machines to make the decision to use lethal force.

The outcome of the Geneva discussions will indicate whether governments can establish common limits on that decision-making, even as they continue to develop AI-enabled military systems.