The UN's Libya report never confirms an autonomous weapon killed anyone

Weapons that choose their own targets

In July 2015, robotics and AI researchers signed an open letter warning about a coming weapon: one that finds a target, decides it is hostile, and fires, with no person choosing that specific target.¹ The letter's own definition draws that line: systems that “select and engage targets without human intervention,” unlike a cruise missile or a remotely piloted drone, where a person still chooses what gets hit.¹ It called such a weapon “the third revolution in warfare, after gunpowder and nuclear arms,” and it has drawn 34,378 signatures since, a running count, not a fixed 2015 tally.¹

That is a different danger from AI helping someone build or find a weapon, covered in a lesson on AI and software flaws and a lesson on AI and biological attack planning. Here the question is not whether AI helps make a weapon. It is who, or what, decides to fire it.

Stuart Russell, a UC Berkeley computer scientist who helped present the letter,² had already made the case in blunter terms. Autonomous weapons could be built to eliminate anyone showing “threatening behavior,” cutting human judgment out of who counts as a target, and that decision is “far too important to be left to machines.”³ Removing a person from the decision also removes risk from the attacker, lowering the threshold for war. The technology, he warned, was close: “It's a very easy step to go from remotely piloted drones to fully autonomous weapons.”³

The term this debate organizes around, meaningful human control, was coined by Article 36, a UK disarmament group active at UN talks since 2013.⁴ Its 2016 briefing names four things control needs to be meaningful: predictable, reliable technology, accurate operator information, a real chance to intervene, and a working line of accountability.⁴ Lose any one, the group argues, and the human decision the laws of war assume is missing.

Human Rights Watch, a founding member of the Campaign to Stop Killer Robots, lays out what breaks without it.⁵ Accountability breaks first: there are “obstacles to holding individual operators criminally liable for the unpredictable actions of a machine they cannot understand.”⁵ The laws of war break next. Telling combatant from civilian is already harder, since “enemy combatants in contemporary conflicts may shed visible signs of military status.” Weighing a strike's military value against civilian harm needs “distinctively human judgement,” HRW argues, in situations no programmer anticipated.⁵ The International Committee of the Red Cross oversees the treaties this argument is fought over. It reaches a similar concern: a weapon that fires at a generalized “target profile” instead of a person the operator chose produces a “loss of human control and judgement in the use of force.”⁶ HRW's term for the same failure is “mechanical slaughter” that reduces people to “data points.”⁵

The letter's most alarming scenario is scale. An autonomous weapon, unlike a nuclear one, is cheap enough to reach “the black market and in the hands of terrorists, dictators wishing to better control their populace, warlords wishing to perpetrate ethnic cleansing.”¹ That puts mass, selective killing within a small group's reach. Two years later, some of the letter's own signatories, Russell among them, put that scenario on screen. Slaughterbots, a short film shown at a UN weapons conference in Geneva in November 2017, depicted swarms of cheap, palm-sized drones killing people by their faces.⁷ Russell told the audience “the technology described in the film already exists and that the window to act is closing fast.”⁷ Slaughterbots is a dramatization its makers commissioned, not a report of an incident that happened.

What Kargu-2, Switchblade, and Harpy actually do

Set against that argument, the harder question is what has actually been built. Ethan Lee of Harvard's Belfer Center reviewed the record in February 2025 and concluded: “To date, there is no publicly available evidence indicating that countries have used fully autonomous weapon systems in combat.”⁸ The systems most often named in this debate are not fully autonomous in use: “human operators are still required to select targets and issue commands.”⁸

The Kargu-2 is also the system behind the incident widely cited as a possible first autonomous kill: loitering munitions deployed in Libya's 2020 civil war. The UN Panel of Experts monitoring the conflict reported that retreating forces were “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,” with “a true ‘fire, forget and find’ capability.”⁹ That does not say the Kargu-2 did the killing rather than the piloted drones named beside it. The word “or” leaves the question open, and the Panel never states that a person died from a fully autonomous engagement, a reading Hitoshi Nasu, a University of Exeter law professor writing for West Point's Lieber Institute, reaches independently.^{9 10} Zak Kallenborn, an adjunct fellow at the Center for Strategic and International Studies, offers the closest thing to a positive claim. He hedges it twice: “If anyone was killed in an autonomous attack, it would likely represent an historic first known case...”¹¹ If, and would likely: the field's most-cited claim will not commit past those two words.

One system in that table does not need a hedge. Israel Aerospace Industries markets the Harpy loitering munition as “Fully Autonomous.” Launched without prior intelligence on a target's location, it loiters for hours, and its onboard radar seeker detects, identifies, and strikes an enemy radar on its own.¹² Israel negotiated Harpy's sale to China in the mid-1990s and had reportedly delivered about 100 of the drones there by 1999, two decades before the 2015 letter.¹³ The company's page never says whether an operator can veto a strike once the seeker locks on. A manufacturer marketing full autonomy for decades has not needed to answer that to sell the weapon.¹² Article 36, the group that defined meaningful human control, concedes the same point from the other side: “there are already limitations to levels of human control exerted in existing military systems during the application of force.”⁴ Some erosion of the human decision on an individual shot is not a hypothetical. It is already normal.

The line between deployed and speculative

The fight over a treaty, three years on

At the CCW's 2023 meeting of parties, states negotiating a treaty could not agree on the meeting's own rules of procedure: “no consensus on this matter was reached.”¹⁴ Their expert group's mandate avoids deciding anything yet: “formulate, by consensus, a set of elements of an instrument, without prejudging its nature.” It meets 10 days a year.¹⁴

Just Security sorts the states into three camps.¹⁵ A pro-treaty group includes Austria, Brazil, and Chile. A middle group is wary of forcing a vote that could fracture the process. A status-quo group, including the United States, Russia, Australia, and Israel, argues existing law already covers the problem.¹⁵ Under the CCW's consensus rule, that smaller group can block what the larger group wants.

A 2024 UN Secretary-General's report compiled views from 73 states and 33 civil society groups. It closed with Secretary-General Guterres calling on states to conclude “by 2026 a new international treaty” to “prohibit weapons systems that function without human control or oversight and that cannot be used in compliance with international humanitarian law.”¹⁶ On 25 August 2026, Guterres and ICRC President Mirjana Spoljaric renewed that call: “We are now dangerously close to crossing a moral red line: the autonomous targeting of humans by machines.”¹⁷ “A failure to act quickly and decisively,” they warned, “will cost lives.”¹⁷

That is a claim about a weapon the UN has never confirmed killed anyone. “Dangerously close” outruns what the Libya report supports. But the opposite confidence has its own gap. Russia's argument that autonomy already helps armies comply with the laws of war says nothing about Harpy, which Israel was exporting to China in the mid-1990s, with no described human check on a strike. Calling a ban premature and calling deployed autonomy well governed are not the same claim. Only the second one has decades of Harpy sales against it.