Are killer robots real? Autonomous weapons explained

Do autonomous weapons exist?

Yes, but probably not in the form most people picture. The key question is not whether a weapon uses AI or has no pilot on board. It is who chooses the specific target and the moment of attack.

The International Committee of the Red Cross (ICRC) puts it this way: after a person activates the weapon, it "self-initiates or triggers a strike in response to information from the environment received through sensors and on the basis of a generalized 'target profile'". That means "the user does not choose, or even know, the specific target(s) and the precise timing and/or location" of the strike.1

The UN expert group's 2026 text draws the same line from the other side. A system is not counted when "the personnel planning, deciding upon, or carrying out the attack determine the specific target or specific group of targets and the timing of the attack".3

Three terms you will see:

• -Human in the loop: a person must select and authorise each target. The US Defense Department's 2012 definition, as summarised by the Arms Control Association, is a system that "require[s] human operator selection and authorization to engage specific targets".12
• -Human on the loop: a person supervises and "can intervene to stop an automated action", but their input is not required to choose or attack targets.12
• -Human out of the loop: fully automated; no one is supervising the individual strike.12

Is it legal? No treaty bans these weapons as a class. The 2026 UN expert report says international humanitarian law "applies fully to the development and use of LAWS", that it "imposes obligations on States, parties to armed conflict, and individuals, not on machines", and that such weapons "must not be used" if they are "inherently indiscriminate" or cannot be used within the law of war.3

Autonomous weapons examples

The examples below are the ones named by the sources we opened. We describe what each source says about who decides, not how the weapons work.

Two things these examples share. First, autonomy in weapons is older than today's AI boom; what has changed is that cheap computer vision now fits on small drones.9 Second, nearly every case is described by its makers or users as having a human somewhere in the decision, and critics argue about whether that human role is real or nominal.29

The Libya case: the "first killer robot" headline

In March 2021 a UN Panel of Experts reported on the war in Libya. Describing fighting in 2020, it said retreating forces "were subsequently hunted down and remotely engaged by the unmanned combat aerial vehicles or the lethal autonomous weapons systems such as the STM Kargu-2 … and other loitering munitions", and that these "were programmed to attack targets without requiring data connectivity between the operator and the munition".6

Some coverage called this the first battlefield killing by an autonomous drone. The report itself does not say that: it does not state that a Kargu-2 killed anyone while operating autonomously, and the passage is footnoted to a "confidential source".6 So the honest summary is: a UN panel says these weapons were programmed to work without a data link in that campaign; whether that mode killed anyone is not documented in the report.

Which countries have lethal autonomous weapons?

There is no official list, and lists differ by source and by year. Three reasons:

1. 1.No agreed definition until recently. The UN expert group only agreed its working characterisation in September 2026, and it says that text is "without prejudice to … the exclusion or inclusion of certain types of systems".3 A list built on a broad definition (anything that can fire without a person pulling the trigger) and one built on a narrow definition (offensive weapons that pick their own targets) will name very different countries.
2. 2.Many systems have more than one mode. The same weapon can be used with a human approving each strike or in an autonomous mode, and armies rarely say which mode they use.7
3. 3.Data ages fast. The most detailed public mapping we opened, SIPRI's, is from 2017.7 Newer catalogues exist but cover less ground: a 2023 academic catalogue of 24 loitering munitions by the AutoNorms project (Tom Watts and Ingvild Bode) includes Chinese and Russian systems, and its authors say they "cannot claim to know whether these capabilities have been fielded".30

What that 2017 SIPRI mapping found, as examples rather than a ranking:

• -Automatic air defence is widespread: "at least 89 countries" had automatic air defence systems; the USA (11 systems) and Russia (8) had produced the widest range.7
• -Robotic sentry weapons were rare: three models, developed in Israel and South Korea, with Qatar and the UAE named as alleged users of the South Korean system.7
• -Loitering weapons that can attack with complete autonomy: four operational systems, all from Israel. Germany, the UK and the USA had started similar programmes but none went beyond research and development.7

Since 2017, the UN Libya report adds a Turkish-made loitering munition used in 2020,6 and CSIS describes Ukrainian companies building autonomous guidance for drones.9 Campaign groups and media reports name more countries; we have not listed those, because we could not check them against a primary source.

Autonomous weapons used in Ukraine

The war in Ukraine is where AI in weapons has moved fastest. What the sources we opened report, in date order:

• -March 2025, CSIS study (Kateryna Bondar). Based on interviews in Ukraine, it found that "fully autonomous warfare remains an aspiration", that AI is used for parts of the job (spotting and tracking targets, navigation, including the last part of a drone's approach), and that "engagement decisions remain squarely in the human domain". In the typical case it describes, the pilot finds and confirms the target, then switches on lock-on and the drone completes the approach itself.9
• -July 2025, US-funded guidance kits. The drone-software company Auterion announced that it would deliver 33,000 AI "strike kits" to Ukraine, "backed by a $50 million Pentagon contract". It says the kits turn "manually-operated drones into autonomous weapons systems capable of AI-driven targeting". That is the maker's own description; we found no Pentagon document on the contract.10
• -September 2026, government tests. The Defense Post reported that Ukraine's Defence Ministry and the Brave1 cluster tested guidance systems from seven companies, aiming for drones that can "detect, acquire, and engage moving ground targets without continuous pilot control". The article does not say whether a person still chooses or approves each target.11

So are these autonomous weapons? By the UN expert group's 2026 test, a drone is not one if the people carrying out the attack "determine the specific target … and the timing of the attack".3 On CSIS's account, most current Ukrainian use fits that description: a human picks the target, software flies the last stretch. CSIS also reports that Ukrainian forces' long-term goal is "to maximize autonomy across the battlefield", with human input "limited to engagement decisions".9 Russia's use of AI in its own drones is not covered by the sources we opened, so this page says nothing about it.

Why it matters for the wider debate: Stop Killer Robots cites Ukraine (and Gaza) among "reports of weapons with varying autonomous capabilities being developed and used", and the UN–ICRC call cites "increasing reports of testing and use" as a reason for urgency.172 On the other side, CSIS reports that Ukrainian forces keep engagement decisions with humans and can override autonomous functions.9

Should autonomous weapons be banned?

This is a real disagreement between governments, not a settled question. The main positions:

For a new treaty (prohibitions plus rules). In October 2023 the UN Secretary-General, António Guterres, and the ICRC President, Mirjana Spoljaric, jointly called on states to negotiate a legally binding instrument and "conclude such negotiations by 2026". Their argument: "The autonomous targeting of humans by machines is a moral line that we must not cross," and weapons whose "effects cannot be predicted" should be prohibited.2 The ICRC's own proposal (2021) is a two-tier approach: ban unpredictable autonomous weapons and those "designed or used to apply force against persons"; regulate the rest with limits on targets, duration, area and situations of use, plus effective human supervision.1

Against a new binding treaty (for now). Some militarily powerful states argue existing law is enough. During the final Geneva session in September 2026, a senior diplomat told Reuters that the US position had "hardened", with Washington "favouring something more akin to guidelines than a legally binding international treaty" because, it says, international humanitarian law already provides a framework; diplomats said Russia, India and Turkey supported that position.26 The US Congress's research service puts it plainly: "The U.S. government does not currently support a ban on LAWS."27 At the March 2026 session, according to the observer report of the Women's International League for Peace and Freedom (WILPF), the US delegation refused the term "human control" and proposed "good faith human judgement and care" instead; many delegations called that insufficient.25 (We could not find the US delegation's own statement.) US policy (Directive 3000.09, 2023) does not ban autonomous weapons; it requires them to "allow commanders and operators to exercise appropriate levels of human judgment over the use of force" and adds senior review before development and fielding.8 In June 2026 a presidential memorandum (NSPM-11) ordered the Pentagon to issue an update to the directive within 90 days;28 as of 16 October 2026 the Defense Department's list of directives still shows the January 2023 version.29

Where the votes stand. At the UN General Assembly on 1 December 2025, resolution 80/57 on lethal autonomous weapons systems passed 164 to 6, with 7 abstentions.4 The six against were Belarus, Burundi, North Korea, Israel, Russia and the United States; the seven abstentions were Argentina, China, Iran, Nicaragua, Poland, Saudi Arabia and Türkiye.5 The resolution does not start treaty talks. It calls on states in the Geneva process to work "towards completing the set of elements for an instrument … with a view to future negotiations".4

What happened in Geneva in 2026. The UN expert group (formally, the Group of Governmental Experts under the Convention on Certain Conventional Weapons, CCW), which works by consensus, adopted its final report on 4 September 2026. It agreed a characterisation of lethal autonomous weapons, restated that the law of war applies to them, said "control and human judgement with regard to LAWS are needed", and said "accountability cannot be transferred to machines". It did not use the word "ban", and it left the legal form open ("without prejudging its nature").3 It recommends that states at the CCW's Seventh Review Conference, 16–20 November 2026, "consider and endorse" these elements and decide how the work continues.3

What is the Stop Killer Robots campaign? A coalition of non-governmental organisations that wants a pre-emptive ban on fully autonomous weapons. By the campaign's own timeline, seven NGOs agreed to form it in New York on 19 October 2012, and it launched publicly in London on 22 April 2013.16 It takes part in the Geneva talks as an observer.3 In September 2025 it reported that 42 states had said they were ready "to move ahead towards negotiations".17

So: is there a ban? No. Will there be one? The next step is the November 2026 Review Conference, which the expert group asks to decide "by consensus" how the work continues.3 This page will be updated after it.

Can AI launch nuclear weapons? Who keeps a human in the loop

Short answer: no source we opened says any country lets AI decide to launch a nuclear weapon. The five nuclear powers recognised by the Non-Proliferation Treaty (China, France, Russia, the UK and the US), plus India and Pakistan, have all said a human keeps that decision; we found no stated policy from Israel or North Korea. The Arms Control Association (ACA) sums up the situation as "a broad but loose consensus for ensuring a 'human in the loop'".12

What each state has said, in its own documents (ACA's fact sheet collects them):12

• -United States: the 2022 Nuclear Posture Review: "In all cases, the United States will maintain a human 'in the loop' for all actions critical to informing and executing decisions by the President to initiate and terminate nuclear weapon employment."18 The US national report to the 2026 NPT Review Conference (dated 24 April 2026) repeats it: "Retain – in all cases – a human 'in the loop'".19 Congress wrote a similar principle into the 2025 defence authorisation act: "positive human actions in execution of decisions by the President with respect to the employment of nuclear weapons".20
• -United Kingdom, France, United States (jointly, 2022): "we will maintain human control and involvement for all actions critical to informing and executing sovereign decisions concerning nuclear weapons employment."21
• -United States and China (November 2024): Presidents Biden and Xi "affirmed the need to maintain human control over the decision to use nuclear weapons" (White House readout).13
• -Russia: its 2026 NPT national report says it "ensures decisive human involvement and judgment in the field of nuclear weapons on an ongoing basis" (UN English translation).22
• -India (February 2026): Foreign Secretary Vikram Misri told the Conference on Disarmament that decisions on nuclear weapons will remain under human control (as reported by the state broadcaster All India Radio, which paraphrases him).23
• -Pakistan: says it supports the principle and "urges all nuclear-weapon states to endorse" it.24

What is not settled. ACA notes that "there are no agreed definitions, constraints, or guidelines on how AI tools might be utilized in nuclear weapons operations", and no active talks among the major nuclear powers on how to put the principle into practice.12 A human pressing the button is one thing; experts also worry about AI in the warning, intelligence and decision-support systems that inform that human, which the pledges above do not fully cover.12

The AI war-game study. In February 2026, King's College London published a study by Professor Kenneth Payne that put three AI models (GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash) into 21 simulated nuclear crises. King's reported that "All 21 crisis games featured nuclear signaling by at least one side" and that "none of the models ever chose accommodation or surrender".14 Headlines turned this into "AI chose nukes 95% of the time"; King's own summary uses the 95% figure in two different ways, so treat that number with care. This was a simulation of chatbots playing a game, not a test of any real weapon system, and the paper was posted on arXiv, a preprint server.14 What it shows is a reason for the human-in-the-loop pledges, not evidence that AI controls nuclear weapons.

If this section left you uneasy, our AI anxiety guide has practical steps.

What could go wrong, and what would help

Even states that oppose a ban list risks in the agreed 2026 UN text: "unintended engagements, risk of loss of control of the system, risk of proliferation, and risk of acquisition by terrorist groups", plus "unintended bias" in data and AI models and "automation bias" (people trusting the machine too much).3 The General Assembly adds the "risk of an emerging arms race" and of "lowering the threshold for and escalation of conflicts".4 The ICRC's concern is more basic: that these weapons substitute "human decisions about life and death with sensor, software and machine processes".1

What the same texts say helps: a responsible chain of human command, limits on targets, time and area, the ability to switch a system off, legal review before a weapon is fielded, testing, and training.13

How this connects to AI doom

Autonomous weapons are a present-day, human-directed risk: governments build and use them. That is different from the scenario our Skynet page covers, where an AI system pursues goals of its own. The two meet in one place: the question of whether humans stay in control of systems that act faster than they can follow. For the bigger question, see Will AI kill us all? and Is AI dangerous?.

FAQ

Are robots going to kill us?

Today's autonomous weapons are tools that states aim at chosen kinds of targets; no source we opened describes one acting on goals of its own. See Will AI kill us all?.

Are autonomous weapons a war crime?

Not in themselves. Using any weapon in a way that cannot respect the law of war is unlawful, and the 2026 UN text says states and individuals "remain at all times responsible" for decisions to use them.3

Who is responsible if one makes a mistake?

The 2026 UN text says states must ensure "individual accountability" because "accountability cannot be transferred to machines".3