Lethal autonomous weapons have already been deployed. They have already killed. Ninety-six nations met in New York in May 2025 to discuss a treaty. The two nations most aggressively developing the technology — the United States and Russia — voted against it.
For most of military history, a weapon required a human hand to fire it. That is no longer guaranteed to be true.
On an unnamed day in 2020 — the exact date remains classified — a Turkish-made Kargu-2 drone operating in Libya's civil war hunted down and attacked logistics convoys and retreating fighters belonging to the Haftar Affiliated Forces. According to a United Nations Panel of Experts report, the drone was operating in a "fire, forget and find" mode. It was not waiting for a human to identify the target. It was not waiting for a human to authorise the strike. It selected, tracked, and attacked on its own.
It may have been the first documented instance of a lethal autonomous weapon killing a human being. The U.N. panel was careful with its language — the report implies rather than confirms a lethal outcome. But the broader finding was unambiguous: a weapon with no human in the loop had engaged human targets in a real conflict, with no legal framework governing whether that was permissible, let alone who bore responsibility.
That was five years ago. The technology has not slowed down. The governance has not arrived.
This is the story of lethal autonomous weapons systems — what they are, how they work, who is building them, why they may be more dangerous than any weapons system since nuclear weapons, and why the world is failing, at speed, to do anything about it.
The word "autonomous" hides a spectrum. Understanding where on that spectrum current weapons sit — and where they are heading — is the first step to understanding the danger.
In November 2024, after a decade of debate, the Convention on Certain Conventional Weapons (CCW) Group of Governmental Experts provisionally agreed on a definition. A lethal autonomous weapons system (LAWS) is an integrated combination of weapons and technological components that enable the system to identify and/or select, and engage a target, without intervention by a human user in the execution of these tasks.[1]
That phrase — "without intervention by a human user in the execution of these tasks" — is the crux of everything. It is also deeply contested, because the degree of human involvement varies enormously across existing systems, and militaries have strong incentives to define their own systems as remaining within human control.
The field generally describes five levels of autonomy, from fully human-controlled to fully machine-controlled:
The AI underpinning these systems draws from the same capabilities transforming every other field: computer vision to identify targets, sensor fusion to track them across environments, reinforcement learning to optimise tactics, and — increasingly — large language models to interpret mission context and make situational judgments. The same technology that taught AlphaGo to play Go at superhuman levels is being used to teach weapons to find, identify, and engage humans.
What makes this particularly alarming is the speed at which these systems operate. A human fighter pilot processes a targeting decision in seconds. An autonomous system can do so in milliseconds — far faster than any human supervisor could meaningfully assess, approve, or intervene. As autonomy increases, human oversight becomes not just difficult but physically impossible.
"The operational speed of autonomous weapons often comes at the expense of meaningful human control and can heighten the risk of conflict escalation through rapid, machine-driven interactions." — Stanford Freeman Spogli Institute, Lethal Autonomous Weapons: The Next Frontier in International Security
The major military powers are not waiting for international law to catch up. Development is accelerating on every front simultaneously.
The autonomous weapons arms race is not hypothetical. It is well underway, funded at scale, and increasingly visible in the weapons systems being showcased at military parades and announced in defence budgets.
In 2023, the Pentagon's Replicator initiative promised "multiple thousands" of attritable autonomous systems by August 2025 — primarily to counter China's numerical advantage in the Pacific. The programme allocated $1 billion across FY2024–2025. By its own deadline it had fielded hundreds, not thousands, with documented failures including drones failing to identify targets correctly and autonomous boats going adrift. Replicator 2.0 pivoted to counter-drone systems. The US voted against the 2025 UN LAWS resolution.[6]
At China's Victory Day parade in September 2025, uncrewed ground vehicles, underwater and aerial drones, and collaborative combat aircraft — autonomous jets flying alongside piloted aircraft — were presented as core components of China's future fighting force. The PLA is developing coordinated swarms capable of identifying, tracking, and executing simultaneous attacks. China has not publicly committed to human-in-the-loop requirements.[6]
Russia and Ukraine have both deployed AI-powered drones that operate with significant autonomy in GPS-denied environments. Russia has used drone swarms to overwhelm Ukrainian air defences, with on-board AI guidance allowing drones to navigate when jamming would disable conventional guidance systems. Russia, like the US, voted against the 2025 UN LAWS resolution.[5]
Israel's Harpy and Harop loitering munitions — autonomous "kamikaze drones" that hunt radar emissions — are among the most sophisticated deployed LAWS. The Harpy can autonomously detect, track, and destroy radar systems without human authorisation. Israel has been a primary supplier of autonomous weapons technology to multiple conflicts and has not endorsed a binding treaty.[3]
STM's Kargu-2 — the drone deployed in Libya — represents the most documented case of an autonomous weapon in combat. Turkey has continued developing and exporting the system. The Kargu-2 uses machine learning-based object classification to identify and attack targets. Turkey is one of the leading exporters of autonomous drone technology globally, including to conflict zones.[3]
Unlike nuclear weapons, which require rare materials and industrial infrastructure, autonomous weapons can be built from commercially available components. The World Economic Forum has noted that non-state actors — including terrorist organisations — can now create lethal autonomous weapons from civilian products. The barrier to entry is falling rapidly.[4]
Unlike previous weapons revolutions — gunpowder, aircraft, nuclear weapons — autonomous weapons require no exotic materials, no large manufacturing base, and no specialised fuel. A nation state's monopoly on this technology cannot be assumed. The same AI capabilities that power autonomous weapons are available to anyone with a cloud computing budget, an open-source model, and a drone.
Nuclear weapons governed themselves, in a terrible way — through mutually assured destruction. Autonomous weapons have no such self-limiting mechanism. That is what makes them uniquely dangerous.
The lowered threshold for use may be the most consequential difference. When a nation deploys autonomous weapons, its own soldiers are not at risk. The political cost of initiating conflict falls. Military action becomes more domestically acceptable, more politically viable — and therefore more likely.
There is also the problem of machine-speed escalation. In a conflict where both sides are using autonomous systems, an engagement can begin, escalate, and reach catastrophic consequences before any human decision-maker is even aware it has started. The compressed timeline is not theoretical — it is an engineering consequence of removing human processing from the loop.
Perhaps most alarming is the convergence of autonomous weapons with nuclear arsenals. Analysts from ICAN and War on the Rocks have documented how autonomous drones integrated into nuclear early-warning infrastructure could blur the distinction between a conventional strike and a nuclear first strike — compressing the window for decision-making to seconds, and dramatically increasing the risk of accidental nuclear escalation.[7]
"Applied machine learning and autonomous systems mean faster warfare and an even shorter period in which decision-makers will have to choose whether to launch nuclear weapons or not." — ICAN, Emerging Technologies and Nuclear Risk
The international community has been discussing lethal autonomous weapons since 2014. The nations most aggressively developing them have ensured those discussions produce nothing enforceable.
The Convention on Certain Conventional Weapons (CCW) begins informal expert meetings on LAWS. The CCW was the body that banned blinding lasers in 1995 — a rare success. But the CCW requires consensus, meaning a single state can veto any outcome.[1]
Year after year, proposals for a binding treaty are blocked. The United States, Russia, India, and Israel exploit the consensus rule to prevent formal negotiating mandates. Talks continue. Nothing is agreed. Meanwhile, development of autonomous systems accelerates in every major military.[1]
After a decade, the CCW Group of Governmental Experts finally agrees on a provisional definition of LAWS — what an autonomous weapon actually is. It took ten years to agree on a definition. No rules governing those weapons are agreed. The definition itself remains provisional.[1]
166 countries approve a UN General Assembly resolution to launch a new forum to expand efforts to legally regulate autonomous weapons. The resolution passes. The forum has no binding authority.[5]
Officials from 96 countries attend a General Assembly meeting in New York — the largest diplomatic gathering specifically focused on autonomous weapons. More than 120 countries publicly support negotiating a treaty that prohibits and regulates LAWS. The meeting concludes without a negotiating mandate.[2]
The UN General Assembly passes a historic resolution: 156 nations vote to negotiate a legally enforceable LAWS agreement by the Seventh Review Conference in 2026. Five nations vote against it — most prominently the United States and Russia. Under the UN system, a General Assembly resolution carries no binding force. The nations most aggressively building the weapons have ensured it stays that way.[2]
UN Secretary-General Guterres and the International Committee of the Red Cross called for a new international treaty on LAWS concluded by the end of 2026. With the United States and Russia opposed, and the CCW consensus rule intact, that deadline is widely regarded as unlikely to be met. The window is not closed. It is, however, closing.[2]
International humanitarian law — the laws of war — requires that attacks distinguish between combatants and civilians, are proportionate, and involve human judgment. When a machine makes the kill decision, there is no legal mechanism to assign responsibility for violations. The programmer? The military commander who deployed the system? The state? Current law has no answer. Every autonomous weapons strike in a conflict zone is potentially a war crime with no one to prosecute.
The Human Rights Watch organisation and Harvard Law School's International Human Rights Clinic have argued that fully autonomous weapons should be prohibited outright — not merely regulated — on the grounds that no machine can currently exercise the human judgment required by international humanitarian law.[2]
The counter-argument, made primarily by the United States, is that existing law is sufficient, that human operators retain meaningful control, and that a treaty would only constrain nations that follow rules — not those that don't. This argument was sufficient to block a binding agreement for over a decade.
The window has not closed. But the history of arms control shows that treaties become exponentially harder to negotiate once technology is deployed at scale, normalised in doctrine, and embedded in military budgets.
The nuclear Non-Proliferation Treaty was negotiated before most nations had nuclear weapons. The Chemical Weapons Convention was negotiated while chemical weapons still existed. The blinding lasers protocol banned a technology before it was deployed. These precedents show that binding law is possible — but it requires political will from the leading military powers. Public pressure on the United States and its allies to support binding negotiations is the single most important lever available to civil society.
Even short of a full ban, a binding requirement that all weapons systems maintain meaningful human control over targeting and engagement decisions would establish a legal floor. This would not ban all autonomous capabilities — defence systems like missile interception could continue — but would prohibit fully autonomous lethal engagement of humans. The ICRC and 120+ nations already support this position. Codifying it in international law is achievable.
A key regulatory proposal from the CCW Group is requiring LAWS to be predictable, reliable, traceable, and explainable — meaning that any engagement must be reconstructable after the fact, and that human commanders must be able to understand how the system made its decision. This would not prevent autonomous operation but would create accountability mechanisms and enable post-hoc legal review of strikes.
The Kargu-2 in Libya was an export. Technology designed and manufactured in one country was deployed in another country's civil war, with no international oversight. Arms export control regimes — analogous to nuclear export controls under the NPT — would slow the spread of the most dangerous autonomous capabilities to conflict zones and non-state actors. This is achievable through existing export control frameworks without waiting for a full LAWS treaty.
In 2018, thousands of Google employees signed a petition opposing Project Maven — a Pentagon AI contract for drone targeting. Google did not renew the contract. AI researchers and engineers have more leverage than any previous generation of weapons-adjacent technologists. Industry ethics standards, researcher refusals, and institutional policies requiring human oversight provisions in all military contracts represent a bottom-up constraint that does not require government action.
The most catastrophic near-term risk is not autonomous weapons killing people on a battlefield — it is autonomous systems misidentifying threats and triggering nuclear escalation. Establishing a clear, legally binding firewall between autonomous weapons systems and nuclear command, control, and communication infrastructure should be a priority of the nuclear-armed states — regardless of their positions on a broader LAWS treaty.
"Because humans created these weapons, humans can limit them. But every year without a binding framework is a year in which deployment normalises, budgets entrench, and the political will to constrain what is already in the field weakens." — Lisa Pedrosa, drawing on analysis from ASIL, Human Rights Watch, and the UN Office for Disarmament Affairs
All findings cited reflect peer-reviewed research, institutional publications, or UN documentation. Where possible, the most recent available literature (2024–2026) has been prioritised.
Overview of CCW Group of Governmental Experts proceedings, including the provisional November 2024 definition of LAWS.
disarmament.unoda.org
ASIL Insights, 2025. Covers the November 2025 UN General Assembly resolution (156 vs. 5), the May 2025 New York meeting of 96 nations, and the 2026 treaty deadline.
asil.org
Analysis of the UN Panel of Experts report documenting the Kargu-2 autonomous engagement in Libya's civil war, 2020.
thedebrief.org
2022. Documents the falling barrier to entry for non-state actors to create lethal autonomous weapons from commercially available components.
weforum.org
2022. Overview of LAWS in Ukraine and Libya, and the regulatory gap. Confirms Russia/Ukraine AI drone deployments.
foreignpolicy.com
Authoritative overview of the Pentagon's Replicator programme, goals, funding ($1B over FY2024–25), and setbacks. Documents programme reorganisation and pivot to counter-drone systems.
congress.gov
Documents the convergence of autonomous systems with nuclear arsenals and the compressed decision-making windows created by machine-speed warfare.
icanw.org
January 2025. Analysis of the geopolitical divisions blocking LAWS regulation, including the CCW consensus veto and the positions of major military powers.
armscontrol.org
May 2025. Documents the HRW/Harvard Law School position calling for a prohibition on fully autonomous weapons, and the May 2025 New York meeting.
hrw.org
Comprehensive academic overview of LAWS risk profiles, including escalation speed analysis and the distinction from nuclear deterrence dynamics.
fsi.stanford.edu
Buy me a coffee