AI-Powered Killer Robots, Who Is Responsible When an Algorithm Chooses the Target?

Military technology is entering a new phase in which artificial intelligence and autonomous systems are increasingly capable of performing tasks once reserved for human operators. Some weapons systems already possess varying degrees of autonomy, ranging from air-defense platforms capable of detecting, tracking and engaging specific threats automatically to drones and loitering munitions that rely on computer vision, automated sensing and autonomous navigation.

The major technological leap, however, comes when autonomy moves beyond navigation and movement into the targeting process itself — identifying a potential target, selecting it and deciding whether to engage.

That shift raises a fundamental legal, ethical and political question: what happens when part of the decision-making process involved in the use of lethal force moves from a human being to a machine capable of acting independently?

The issue returned to the international agenda on September 5, 2026, when 128 states participating in the Convention on Certain Conventional Weapons reached consensus on key elements for defining lethal autonomous weapons systems and addressing their use.

The document does not create new legally binding obligations, nor does it establish new legal duties for states. However, it could provide a basis for further discussions, including possible negotiations on an international instrument during the convention's review conference in November.

From Automation to Autonomy, What Has Changed?

Modern weapons can be categorized according to the degree of autonomy they possess at different stages of what military analysts often call the “kill chain” — the sequence involving detection, identification, selection, engagement and execution.

The use of artificial intelligence does not automatically make a system a lethal autonomous weapon. There is an important distinction between an AI system that assists a human decision-maker by analyzing images or prioritizing potential targets and a system that, once activated, can independently select and engage a target without additional human intervention.

Vincent Boulanin, director of the Stockholm International Peace Research Institute's AI governance programme, has explained that the crucial technical distinction concerns where human intervention occurs in relation to the system's ability to use force.

In decision-support systems, AI generates a recommendation that a human interprets before making a specific targeting decision. Autonomous systems, by contrast, can execute a broader human decision established in advance and adapt that decision during an attack according to the environment and operational parameters in which they function.

SIPRI has likewise distinguished autonomous weapons from AI-enabled decision-support systems, particularly by examining the role of humans in the targeting process.

Military autonomy itself predates modern artificial intelligence. Certain naval mines and anti-ship missiles have long incorporated autonomous functions, while air-defense systems have used radar and computing technologies for decades to detect, track and engage particular threats under tightly defined conditions.

The US Phalanx close-in weapon system, first deployed in 1980, is a prominent example of constrained autonomy. It can automatically detect, evaluate, track and engage certain airborne threats, but it operates against defined categories of targets and within a restricted operational environment.

More recent developments in machine learning, computer vision, advanced sensors and high-performance computing have expanded the range of tasks that military systems can potentially perform without continuous human intervention.

Boulanin has nevertheless warned that the distinction between autonomous weapons and AI-assisted decision-support systems can become blurred under certain circumstances.

If an AI decision-support system is connected to automated fire-control mechanisms, or if recommendations are generated so quickly that an operator has no realistic opportunity to examine them, the overall system could effectively become part of an autonomous targeting chain.

In such circumstances, a commander or operator could end up following the system's recommendations almost automatically because of time pressure or a lack of tools needed to critically assess them.

More Than a Decade of International Debate

International concern over lethal autonomous weapons gained significant momentum in 2013 following a report by UN Special Rapporteur Christof Heyns on so-called “killer robots.” The issue subsequently moved into a formal diplomatic process under the Convention on Certain Conventional Weapons.

In 2014, states held their first expert meeting under the convention to examine the emerging technology. The International Committee of the Red Cross also convened specialists from governmental, technical, legal and ethical backgrounds.

By 2017, lethal autonomous weapons had become part of the work of a governmental group of experts within the convention, with states reaffirming the relevance of international humanitarian law to the development and use of such systems.

The central disagreement has remained remarkably consistent: how much autonomy should be permitted, and what international rules should govern systems capable of selecting and engaging targets?

The ICRC describes autonomous weapons as systems that select and apply force against targets without human intervention. While some autonomous systems may operate against inherently military targets, concerns become considerably more complex when autonomy extends to the selection of human targets in unpredictable and densely populated environments.

On August 25, 2026, UN Secretary-General António Guterres and ICRC President Mirjana Spoljaric warned that the world was approaching what they described as a dangerous ethical threshold involving autonomous weapons targeting human beings.

During recent negotiations, 76 states supported moving toward negotiations on an international legal instrument, while other countries favored focusing on national rules or operational guidelines. The split illustrates the continuing disagreement over what form future regulation should take.

From Libya to Ukraine and Gaza, The Battlefield Is Changing

For years, lethal autonomous weapons appeared largely as a future possibility or as limited prototypes. Current conflicts, however, have pushed the debate closer to the battlefield.

In Libya, a UN panel of experts reported the use of the Kargu-2 system against retreating forces and pointed to the loitering munition's ability to operate without a continuous data connection to its operator. The case subsequently became an important reference point in discussions about autonomous combat capabilities.

In the war in Ukraine, the ability of drones and other systems to continue operating despite communications disruptions has become increasingly significant as electronic warfare has expanded.

Ukraine's Defense Ministry has said that an AI platform used on the battlefield can identify approximately 70% of enemy military targets by processing more than 100,000 drone video feeds each month.

In Gaza, a journalistic investigation reported that the Israeli military used an AI-assisted system known as Lavender to help identify and generate assassination targets.

According to that investigation, the system classified approximately 37,000 people as suspected militants. The report also alleged that human reviews of some targets lasted no more than 20 seconds, potentially reducing human approval to a largely formal step. The Israeli military has disputed some of the investigation's characterizations.

The broader issue goes beyond the technical accuracy of an algorithm. It concerns what happens when the volume of recommendations produced by an automated system exceeds an operator's practical ability to independently examine each one.

Ingvild Bode, director of the Center for War Studies at the University of Southern Denmark and a researcher specializing in military AI governance, has argued that the distinction between genuine human judgment and nominal approval depends on whether an operator has the ability to exercise meaningful agency and influence the final outcome.

When an operator lacks sufficient time, information or cognitive capacity, excessive reliance on automated recommendations can emerge. This phenomenon is commonly described as automation bias, in which humans place disproportionate trust in machine-generated recommendations.

When the System Makes a Mistake, Who Is Responsible?

If an autonomous system strikes the wrong target, international law does not simply assign responsibility to the algorithm.

Legal responsibility can involve states and individuals connected to the system's design, development, testing, deployment and operation.

International law researcher Jonathan Quick of the Asser Institute has argued that responsibility should not necessarily be viewed as a burden divided between a state, commander, operator and developer. Depending on the circumstances, several actors could bear responsibility simultaneously.

The state generally carries the broadest responsibility for the legal framework, military doctrine, procedures, training, conditions of use and relationship with technology developers.

Laura Bruun, a researcher with SIPRI's AI governance programme, has noted that determining responsibility for an erroneous strike depends on the source of the failure, whether the failure was foreseeable and the scale of the resulting harm.

The legal standards also differ between states and individuals. A state may incur responsibility for violating international humanitarian law without proving criminal intent, while establishing individual criminal responsibility for a war crime requires proof of the relevant elements of the offense, including knowledge and intent where applicable.

Holding one party responsible does not automatically absolve others. A state, commander, operator and developer could potentially face different forms of responsibility if an investigation establishes that their respective actions or failures contributed to the violation.

This creates what experts often describe as an accountability gap — a difficulty in identifying responsibility when decision-making is distributed across multiple actors.

The challenge becomes even greater when machine-learning systems are involved because their behavior may be less predictable than that of conventional rule-based software.

For that reason, investigating a wrongful targeting incident may require much more than reviewing the final attack order. Investigators may need to examine the entire lifecycle of the system, including its design, programming, training data, testing, legal review, procurement, deployment and operational use.

Documentation is therefore critical. Records of training data, test results, known failure rates and documented limitations could help determine whether a failure originated in the system's design, training or testing — or in the decision to deploy it despite known weaknesses.

Meaningful Human Control, Where Does the Human Remain?

At the center of the international debate is the concept of meaningful human control, which seeks to establish a minimum level of human involvement in the use of lethal force.

Simply having a human somewhere in the decision-making chain does not automatically guarantee meaningful control.

Effective human control depends on whether the operator understands the target and operational context, knows the system's limitations, can intervene in time and possesses a genuine ability to reject or stop an attack.

Bode has argued that the targeting decision remains important, but it is not the only relevant moment. Human control is shaped from the earliest stages of system design and testing and is influenced by permitted targets, geographical and temporal restrictions, safety mechanisms, exclusion zones and operational procedures.

Beatrice Fihn, executive director of Lex International and an advocate for restrictions on autonomous weapons, has argued that approving large numbers of targets generated within seconds could turn human authorization into little more than a symbolic stamp.

From this perspective, meaningful human control requires not only a human decision-maker but also reliability, predictability, traceability and the genuine ability to reject, modify or stop an automated recommendation.

The issue even extends to interface design.

Boulanin has suggested that the interface through which an operator receives an AI recommendation could itself become part of the control mechanism. Systems might display confidence levels, explain how a recommendation was produced or present information in ways designed to encourage critical evaluation.

The wording of a recommendation could also matter. Presenting an AI output as an open question may encourage more scrutiny than presenting it as a completed decision. Some researchers are examining whether systems could monitor operator behavior and trigger additional confirmation requirements if a user begins approving recommendations unusually quickly.

Such approaches remain an area of research rather than a universally adopted military standard.

Quick has argued that human control should ultimately be translated into operational requirements that can be tested. These could include restricting the geographical and temporal scope of operations, defining the range of possible outcomes, providing sufficient test results and system information, training operators on known weaknesses and designing interfaces that provide necessary information without overwhelming the user.

Red Lines, What Should Be Prohibited?

After more than a decade of debate, the central question is increasingly less about whether autonomous weapons can be developed and more about which categories should be permitted and which should be prohibited.

Some states and humanitarian organizations have called for prohibitions on systems that can target human beings without adequate human control or that cannot reliably be operated in compliance with international humanitarian law.

Other positions emphasize strict regulation and operational limitations. Systems that retain meaningful human oversight but incorporate advanced automation are often treated as candidates for regulation rather than an outright ban.

Fihn has argued that fully autonomous systems capable of selecting and engaging human targets without genuine human intervention should be explicitly prohibited. Her position is that the fundamental issue is not simply whether an algorithm can be made more accurate, but whether the decision to kill a human being should be delegated to a machine.

Systems retaining meaningful human involvement could instead be subject to strict rules governing their operational environment, targets, timing, geographical scope and mechanisms for human intervention.

Bode has emphasized that meaningful control begins early in the system's lifecycle. Development, testing, evaluation, certification and procurement decisions all influence how much an operator knows about a system's reliability and how predictable its behavior will be.

Bruun similarly argues that responsibility extends across the lifecycle of an autonomous system, including design, procurement, training and testing. A state's obligations under international humanitarian law do not begin only when a weapon is fired.

Article 36 of Additional Protocol I to the Geneva Conventions requires states to determine, when studying, developing, acquiring or adopting a new weapon, means or method of warfare, whether its use would be prohibited under international law in some or all circumstances.

Legal review, however, is only one part of the equation. Traceability, adequate information for operators, and clear restrictions on environment, timing, target type and the number of engagements a system can conduct are also central to accountability.

Digital records can help investigators reconstruct events after an incident, but they cannot guarantee a complete reconstruction. Complex military systems, technical uncertainty and operational conditions may still leave gaps.

That means traceability has to be designed into the system from the beginning, with procedures ensuring that critical events are recorded and remain accessible for subsequent investigations.

In fast-moving combat environments, meaningful human supervision can become particularly difficult. Simply placing an operator somewhere in the system may therefore be insufficient. More stringent restrictions may be required on system design, deployment environments and operating conditions.

The Question Is No Longer Purely Technological

The debate over autonomous weapons ultimately reaches beyond artificial intelligence.

The fundamental question is whether humans will retain enough knowledge, time and practical authority to prevent a machine from making an irreversible decision.

Recent international discussions show that governments remain divided over whether the answer should come through a binding international treaty, national rules, operational restrictions or a combination of these approaches.

The technological trajectory is also moving rapidly. Recent reporting on AI-enabled warfare has highlighted continuing disagreements over how much autonomy military systems should receive and what level of human oversight should remain mandatory.

The real challenge is therefore not simply whether an algorithm can detect a target. It is what happens when that algorithm can determine which target should be attacked, while the human who is ultimately expected to bear responsibility may have only seconds — or insufficient information — to understand, challenge or stop the decision.

As autonomous military systems become more sophisticated, the defining test for international rules may not be whether a human being technically remains somewhere in the kill chain.

It may be whether that human still possesses the knowledge, time, authority and practical ability to exercise meaningful control over a decision that cannot be undone.

The future of autonomous warfare could ultimately depend on where governments draw that line — and whether those limits can be enforced before technology moves beyond them.