Who Decides to Kill? The Unresolved Ethics of Autonomous Weapons

The most important legal and ethical question in modern warfare has no settled answer. When an autonomous system selects and engages a human target without a human being in the decision loop at the moment of lethal action, who bears legal and moral responsibility for that kill? The question is not hypothetical. It is active. Autonomous weapons systems have made targeting decisions that resulted in human deaths. Governments are building more of them. And the international community has spent more than a decade failing to agree on basic definitions — let alone binding prohibitions.

This analysis examines the current state of autonomous weapons policy, the incidents that have forced the debate into sharper focus, the positions of major powers, and the substantial gap between where the stated red lines are and where autonomous lethality is actually operating today.

DoD Directive 3000.09: The Foundational Document

The United States Department of Defense established its formal policy on autonomous and semi-autonomous weapons systems through Directive 3000.09, first published in November 2012 and significantly updated in January 2023. The directive is the most detailed official policy statement on autonomous weapons from any major military power, and it serves as the de facto reference point for NATO discussions on the subject.

The directive establishes three categories of weapons systems:

• Autonomous weapons systems: Systems that, once activated, can select and engage targets without further intervention by a human operator. DoD policy requires these systems to apply lethal force only against "materiel targets" (objects rather than humans) unless specific senior-level approval is obtained, and even then only under highly circumscribed conditions.
• Semi-autonomous weapons systems: Systems that require continuous human operator input to select targets or apply lethal force. This is the category in which most current U.S. precision munitions fall.
• Human-supervised autonomous weapons systems: Systems that can select and engage targets, but with a human operator who retains the ability to intervene and halt the engagement. This category — "human on the loop" — is where most advanced current systems operate and where significant policy tension exists.

The 2023 update to Directive 3000.09 was prompted in part by the operational experiences of Ukraine and the accelerating pace of AI capabilities. The updated directive maintained the fundamental requirement for "appropriate levels of human judgment" in the use of lethal force, but acknowledged that defining "appropriate" must account for the speed at which autonomous systems can operate — particularly in defensive counter-UAS scenarios where the engagement timeline is measured in seconds, not minutes.

The KARGU-2 Incident: Libya, 2020

The most thoroughly documented case of an autonomous weapons system potentially engaging a human target without direct human authorization occurred during the Libyan civil war in 2020. A March 2021 UN Panel of Experts report on Libya (document S/2021/229) described an engagement in which autonomous loitering munitions — identified as KARGU-2 systems manufactured by Turkish company STM (Savunma Teknolojileri Muhendislik) — hunted down and attacked retreating Libyan National Army-affiliated fighters and logistics vehicles following a failed assault on Government of National Accord positions.

The UN report stated that the autonomous weapons "were programmed to attack targets without requiring data connectivity between the operator and the munition" and that they "hunted down and remotely engaged" targets. The specific phrasing used in the report — that the lethal engagement may have occurred "without requiring data connectivity" — implies that the targeting decision was made by the weapon's onboard AI, not by a human operator receiving real-time imagery.

STM and the Turkish government disputed the interpretation that the KARGU-2 operated fully autonomously, maintaining that the system's AI capability requires operator activation and that the engagements described may have involved human authorization. The UN report itself was careful to note the limitations of its evidence, acknowledging that the incident was not definitively confirmed as a fully autonomous lethal engagement. However, the episode entered the policy literature as the first case in which a UN expert body assessed that autonomous lethal engagement was "possible" — and has been cited in every subsequent CCW discussion.

Israel's Lavender, Gospel, and Fire Factory: AI at Industrial Scale in Gaza

If the KARGU-2 incident represents the outer edge of documented autonomous lethality, the Israeli Defense Forces' AI-assisted targeting systems deployed during the Gaza campaign of 2023-2024 represent the most extensively reported case of AI integration into targeting decisions at operational scale.

Lavender

Reporting by Israeli outlet +972 Magazine and the Local Call, corroborated by subsequent coverage in The Guardian and The New York Times and based on testimony from multiple IDF intelligence officers, described a system called Lavender that used machine learning to generate target lists. The system reportedly analyzed signals intelligence, social network data, phone and communications metadata, and other data to identify individuals algorithmically assessed as likely Hamas or Palestinian Islamic Jihad operatives, assigning each a probability score. At peak operational tempo, Lavender was reported to have flagged approximately 37,000 individuals as potential targets.

According to the reporting, IDF officers described spending as little as 20 seconds reviewing each Lavender-generated target recommendation before authorizing a strike — a process critics characterized as functionally equivalent to rubber-stamping algorithmic kill lists. The IDF disputed the characterization of how the system was used and maintained that human judgment was applied to each targeting decision.

Gospel and Fire Factory

Gospel (Habsora in Hebrew) is a complementary IDF AI system focused on generating targeting recommendations for physical infrastructure — buildings, tunnels, command nodes — rather than individuals. An earlier IDF publication acknowledged Gospel's operational use, describing it as producing target generation at a pace that would be "impossible for humans to match."

Fire Factory is a strike coordination and logistics AI system that automates the process of assigning munitions loads, calculating strike sequencing, and scheduling missions against target lists generated by Gospel and Lavender. The combination of the three systems creates an end-to-end AI-assisted kill chain in which human involvement occurs primarily at authorization checkpoints — with the cognitive work of target identification, target selection, munitions assignment, and mission sequencing performed by algorithms.

The IDF's AI targeting architecture raises the central question of autonomous weapons ethics in its sharpest form: if AI generates the target, AI assigns the munition, and AI schedules the strike, but a human operator clicks "approve" in a 20-second review window — is meaningful human judgment actually being applied? International humanitarian law requires that targeting decisions be made with individual human judgment applying the principles of distinction, proportionality, and precaution. Whether that standard is met when the human's role is reduced to reviewing an algorithmic recommendation under time and cognitive pressure is the live policy dispute.

The OODA Loop Problem: AI Outpaces Human Decision-Making

The core technical reality driving autonomous weapons policy is the OODA loop — the Observe, Orient, Decide, Act cycle formulated by USAF Colonel John Boyd in the 1970s. Boyd's insight was that the side that could cycle through the OODA loop faster than its opponent would gain decisive advantage. In modern AI-enabled warfare, AI systems can complete an OODA cycle for certain engagement types in milliseconds. Human cognitive processing cannot operate at equivalent speed.

This creates a practical dilemma for "human in the loop" requirements:

• Counter-UAS engagements may require target identification and intercept decisions within 2-10 seconds of threat detection. No human operator reviewing sensor data can consistently make reliable engagement decisions within that window.
• Electronic warfare — particularly cognitive EW systems that adapt waveforms in real time — operates at speeds measured in microseconds. Human oversight of individual engagement decisions is technically impossible.
• Swarm engagements — coordinating simultaneous autonomous operations across dozens or hundreds of platforms — exceed human cognitive capacity to supervise at the individual platform level.
• Cyber offensive operations — automated network attacks and defenses — operate below any practical threshold for human authorization of individual actions.

The practical result is that "human in the loop" requirements, however sincerely stated in policy documents, have been functionally replaced by "human on the loop" architectures — where the human sets parameters and can intervene, but does not approve individual targeting decisions. Every operational system operating at machine speed is, by necessity, autonomous at the action level. The policy debate has not caught up to this operational reality.

The UN CCW Process: A Decade of Non-Resolution

The Convention on Certain Conventional Weapons (CCW) has been the primary international forum for autonomous weapons discussions since 2014, when an informal Group of Governmental Experts (GGE) was first convened to address the topic. The CCW's membership includes most major military powers, including the United States, Russia, China, the United Kingdom, France, Germany, and Israel.

After more than a decade of discussions, the CCW GGE has produced extensive documentation but no binding treaty, no common definition of "lethal autonomous weapons systems," and no enforceable prohibition of any specific capability. The structural reasons for this failure are systemic:

• Definitional disputes: States cannot agree on what constitutes an autonomous weapons system. The distinction between "autonomy" and "automation" remains contested. Most states argue their current systems are automated (following programmed rules) rather than autonomous (making independent judgments) — a distinction that becomes increasingly artificial as AI capabilities advance.
• Verification impossibility: Unlike nuclear weapons, autonomous AI systems are software-defined capabilities that cannot be verified through physical inspection. A country can commit to human-in-the-loop requirements and violate them in operational practice without any external verification mechanism detecting the violation.
• Major power objection: Russia has consistently blocked binding CCW resolutions on autonomous weapons, maintaining that existing international humanitarian law is sufficient and that new instruments are unnecessary. The United States has supported discussions but opposed binding prohibitions that would constrain current or future U.S. programs.

China's Position: Official Restraint, Operational Acceleration

China's stated position on lethal autonomous weapons is nuanced in ways that serve its strategic interests. In CCW discussions, China has supported discussions of autonomous weapons restrictions and in 2018 submitted a working paper proposing a prohibition on fully autonomous weapons that lack human control. This positioning allowed China to present itself as more progressive than the United States on autonomous weapons policy.

However, China's actual weapons development programs tell a different story. The People's Liberation Army has been systematically developing autonomous weapons capabilities across all domains, including:

• The CH-7 stealth combat drone, designed for autonomous deep-penetration strike missions
• Multiple AI-enabled counter-drone systems, including automated CIWS configurations for naval vessels
• Autonomous underwater vehicles for ISR and mine-laying operations in contested waters
• AI-assisted targeting systems for ballistic missile forces, reducing response and launch decision timelines

China's working paper definition of "prohibited" autonomous weapons was carefully constructed to exclude most of these systems. Its stated support for human control requirements applies to a definition of "fully autonomous" so narrow that virtually all current systems would fall outside it. The gap between China's CCW rhetoric and its operational weapons programs represents perhaps the most significant source of strategic instability in the current autonomous weapons environment.

The Campaign to Stop Killer Robots

Civil society opposition to autonomous weapons is organized primarily through the Campaign to Stop Killer Robots, a coalition of more than 270 NGOs from 70 countries launched in 2013. The campaign advocates for a legally binding instrument prohibiting fully autonomous weapons — what campaigners call "killer robots" — before they are deployed at scale.

The campaign has succeeded in elevating public awareness and maintaining autonomous weapons on the CCW agenda, but has not achieved its core legislative objective. A 2022 resolution in the United Nations General Assembly calling for multilateral discussions on autonomous weapons passed with broad support — but resolutions are non-binding and have produced no operational consequences.

The campaign's challenge is structural: it is arguing for prohibition of capabilities that multiple great powers are actively developing and that have demonstrated operational value in Ukraine. The window for prohibition of autonomous weapons systems as a class may have already closed, with the practical policy debate having shifted toward regulation of use rather than categorical prohibition.

NATO Nations: Autonomous Authority Levels in Practice

The actual autonomous engagement authorities across NATO member states vary significantly by domain and system type. The following represents the current operational picture based on open-source national doctrine documents and official statements:

The table reveals a fundamental truth about the autonomous weapons debate: the red line has already been crossed in multiple domains. Phalanx close-in weapon systems have operated in autonomous engagement mode for decades. Active Protection Systems on armored vehicles engage incoming projectiles autonomously. Iron Dome's engagement computer operates faster than any human decision cycle. The framing of autonomous weapons as a future danger to be prevented is substantially incorrect — they are a present reality to be governed.

Where the Red Lines Actually Are

The honest assessment of current autonomous weapons red lines — as opposed to the stated policy positions of major military powers — yields the following conclusions:

The accepted red line: No major military power openly claims the right to deploy AI systems that autonomously select and kill named human individuals without any human authorization step. The targeting of specific persons for lethal action remains nominally within the human authorization requirement under existing IHL frameworks.

The practical reality: When AI systems generate target lists that humans approve in 20-second review windows under cognitive overload conditions, the effective decision-maker is the algorithm. The human authorization step functions as legal cover rather than genuine deliberative judgment. This is the condition critics identify in the IDF's Gaza operations, and it is the trajectory toward which operational pressure is pushing autonomous weapons globally.

The trajectory: As AI targeting systems become more capable and adversary threats become faster, the practical case for genuine human deliberation in targeting decisions weakens. The operational logic of autonomous weapons — faster, more numerous, cheaper, less politically costly — creates structural pressure to reduce meaningful human involvement even as policy documents nominally require it.

The resolution of this tension, if it comes, will not emerge from the UN CCW process or from campaign pressure alone. It will emerge from military defeats — when autonomous weapons used irresponsibly create strategic disasters of sufficient magnitude to shift the cost-benefit calculation — or from negotiated bilateral frameworks between the United States and China that establish mutual constraints, as nuclear arms control eventually did for the preceding generation of weapons of mass effect.

Neither outcome is imminent. Until then, the question of who decides to kill when an AI system pulls the trigger remains legally unresolved, ethically contested, and operationally urgent.