Autonomous Weapons

The core claim is about a specific delegation. Modern militaries already use machines that fire without a person pulling a trigger, including missile defence systems that engage incoming projectiles too fast for human reaction. What campaigners object to is a weapon that decides which human being to attack. The UN Secretary-General and the President of the International Committee of the Red Cross issued a joint appeal on 5 October 2023 asking states to prohibit autonomous weapon systems whose effects cannot be predicted, to restrict all others by place, time, target type and scale of force, and to conclude negotiations on a binding instrument by 2026.^[1]

Three distinct objections sit underneath that appeal. The first is legal: international humanitarian law requires proportionality and distinction judgements that depend on context a classifier does not have, and a machine cannot be prosecuted, which leaves an accountability gap when something goes wrong. The second is moral and does not depend on performance at all. In the Ipsos survey of 28 countries conducted for Human Rights Watch between November 2020 and January 2021, 66% of those who opposed autonomous weapons gave as their reason that machines should not be allowed to kill, ahead of accountability at 53% and technical failure at 42%.^[2] The third is strategic: cheap autonomous systems proliferate, and weapons that need no operator scale differently from weapons that do.

A 61-page report published in April 2025 by Human Rights Watch and Harvard Law School's International Human Rights Clinic extended the argument beyond wartime, holding that autonomous weapon systems threaten rights to life, peaceful assembly, privacy and remedy across their whole life cycle, including the mass surveillance needed to train them.^[3] That framing matters because it moves the concern from a narrow arms control question to a policing and civil liberties one.

The clearest evidence comes from Ukraine, where autonomy has been adopted for a practical reason rather than a doctrinal one: Russian electronic warfare severs the radio link between operator and drone. Kateryna Bondar of the Center for Strategic and International Studies reported in March 2025 that small first person view drone missions succeed roughly 10-20% of the time without autonomy, and that onboard terminal guidance raises that to roughly 70-80%. She also recorded that Ukraine produced about 2 million drones in 2024, that 96.2% of the uncrewed aerial vehicles used by its armed forces that year were made domestically, and that Ukrainian forces began buying 10,000 drones with onboard AI in 2024.^[4]

Bondar's central finding cuts against the more dramatic readings. She wrote that autonomy in the sense the US military defines it, a system accomplishing goals independently in complex and unpredictable environments, is not yet present on the battlefield in Ukraine, and that engagement decisions remain in the human domain with operators able to override autonomous functions.^[4] The autonomy that has been fielded handles navigation and the final seconds of an approach after a human has designated the target.

On the diplomatic side the record is precise. The UN General Assembly adopted resolution 80/57, entitled Lethal autonomous weapons systems, at its 52nd plenary meeting on 1 December 2025. The text calls on parties to the Convention on Certain Conventional Weapons to work towards completing the set of elements for an instrument being developed within the expert group's mandate, with a view to future negotiations, and it notes the Secretary-General's calls to begin negotiating a legally binding instrument on a two-tier approach of prohibitions and regulations.^[9] The predecessor resolution, 79/62, was adopted on 2 December 2024 by 166 votes to 3 with 15 abstentions.^[10] Human Rights Watch counts more than 120 countries calling for a treaty.^[3] The 2026 sessions of the Group of Governmental Experts were scheduled for 2-6 March and 31 August-4 September, ahead of the review conference of the Convention on Certain Conventional Weapons.^[11]

The most likely failure is not a machine going rogue but a classifier being wrong in a way nobody notices until afterwards. A system trained to recognise a target class from overhead imagery will encounter conditions its training data did not cover, and unlike a human operator it will not hesitate when a scene is ambiguous. Compounding this, the same system may be replicated across thousands of units, so a single classification weakness becomes a systematic error rather than one soldier's mistake. That is the concern behind the emphasis in the Convention on Certain Conventional Weapons rolling text on predictability, reliability and operational limits.^[11]

The second path is proliferation. Paul O'Neill, Sam Cranny-Evans and Sarah Ashbridge wrote in a 2024 Royal United Services Institute occasional paper that military off-the-shelf systems have high degrees of autonomy and are proliferating rapidly to anyone with sufficient funds, including non-state actors, and that the software, hardware and expertise needed to build a minimally functioning lethal autonomous weapon are widely available and within reach of non-state actors that see a benefit in adopting one.^[12] Unlike nuclear or chemical programmes, there is no scarce input to control: the machine learning components are published, and the airframes are consumer goods.

A third path concerns thresholds rather than accidents. If a state can strike without risking its own personnel and without any individual having to authorise a particular killing, the political friction that restrains the use of force weakens. Alexander Kmentt, director of disarmament, arms control and non-proliferation at the Austrian foreign ministry, wrote in Arms Control Today in January 2025 that geopolitical tensions and eroded trust make many states reluctant to adopt regulations their competitors might ignore, and that Russia has used procedural tactics to delay discussions and reject expressions of urgency in the expert group's reports.^[13] Because the Convention on Certain Conventional Weapons decides by consensus, individual states can block progress, which Kmentt argues effectively guarantees standstill.^[13]

Kenneth Anderson and Matthew Waxman made the most durable counterargument in a 2013 Hoover Institution essay. A prohibition, they wrote, trades speculative risks from autonomous systems for the real but less visible risk of never developing automation that could make the use of force more precise and less harmful to nearby civilians. They also argued that autonomy as such does not make a weapon unlawful, since the legal tests apply to whether a weapon can be directed at a military objective, and they proposed the gradual evolution of codes of conduct grounded in existing law rather than a categorical ban.^[14]

Hitoshi Nasu of the Lieber Institute made a narrower point about the Kargu-2 episode in June 2021: the public argument over whether an autonomous system had killed anyone distracted from the question that actually governs legality, which is whether the system can be used within controlled parameters that distinguish military objectives from civilians.^[6] On that framing, a weapon released into a defined area against a defined target class for a defined period may be lawful, and a human-operated weapon used carelessly may not be. Michael Schmitt reached a similar conclusion about the Israeli systems, holding that Gospel and Lavender are databases and recommendation tools whose outputs analysts can reject, not weapons that choose targets.^[8]

Public opinion is also less solid than the headline numbers imply. Michael Horowitz ran two survey experiments on the US public, reported in Research and Politics in 2016, and found that opposition is contextual: fear of other countries or non-state actors developing these weapons made respondents significantly more supportive of developing them, and respondents became much more willing to use them when doing so would protect US forces. He concluded that claims about how the public views unfamiliar technologies call for modesty.^[15] Ondrej Rosendorf, Michal Smetana and Marek Vranka reported in Security Studies in 2023 that American attitudes are significantly contingent on beliefs about how error-prone these systems are relative to human-operated ones, and found only limited evidence that concern about human dignity predicts opposition.^[16] A 61% figure measured in the abstract is not a stable political constraint.

Finally, the empirical claim that autonomous weapons are here is weaker than campaign material suggests. Bondar's assessment that battlefield autonomy in the full sense does not yet exist in Ukraine, published by an institution with no stake in reassurance, is the most direct evidence available from the largest drone war in history.^[4] The gap between autonomous navigation and autonomous target selection is where the entire debate sits, and that gap has not closed in public view.

The United States has had a written policy since 2012, reissued in January 2023 as Department of Defense Directive 3000.09. It requires that autonomous and semi-autonomous weapon systems be designed to allow commanders and operators to exercise appropriate levels of human judgement over the use of force, and it sets out a senior review before development and again before fielding.^[17] The directive regulates rather than prohibits: it sets conditions on how such systems are designed, reviewed and used, and it is departmental policy rather than law.

At the multilateral level the work sits in the Group of Governmental Experts under the Convention on Certain Conventional Weapons, which has been meeting on this subject for roughly a decade. Its chair has been circulating a rolling text, updated most recently on 5 June 2026, covering the characterisation of these systems, the applicability of international humanitarian law, human judgement and control, prohibitions on systems whose effects cannot be anticipated or controlled, and obligations across the life cycle including legal reviews and testing.^[11] The group met for five days in March 2026 and was scheduled to meet again from 31 August, at the close of its mandate.^[11] Stop Killer Robots launched a paper titled Inflection Point at the CCW: Taking Action on a Decade of Work at a side event on 4 March 2026, arguing that areas of significant convergence on the elements of an instrument have already been developed.^[11][18]

The 2026 deadline set by the Secretary-General and the ICRC has passed without negotiations formally opening.^[1] The General Assembly resolution adopted in December 2025 asks states to work towards completing elements for an instrument with a view to future negotiations, which is a step short of a negotiating mandate.^[9] That is the plainest available measure of where the diplomatic effort stands: broad General Assembly support for addressing the issue, no mandate to negotiate, and a consensus rule in the only forum currently tasked with drafting rules.^[13] National measures such as Directive 3000.09 fill part of the gap for one state, but they are policy rather than law, and they can be revised by the department that wrote them.^[17]

Ask what specific function is being automated in any given claim. Navigation without satellite positioning, recognising a vehicle type, holding a lock on a moving object and choosing which object to attack are four different capabilities with four different legal and moral profiles. Reporting routinely uses one word for all of them. The Ukrainian systems that raised hit rates from 10-20% to 70-80% automate the third, and that is a large military change that does not by itself involve a machine choosing a victim.^[4]

The second question is whether the disagreement in front of you is factual or moral. Anderson and Waxman are not disputing that autonomous weapons could be built; they are arguing that the legal framework already covers them and that a ban would forfeit precision gains.^[14] Campaigners who take the position that machines should not decide to kill are not making a claim that better classifiers would refute. Both positions are honestly held, and evidence about classifier accuracy only settles one of them.

For tracking the issue over time, the diplomatic indicators are more legible than the technical ones. Whether the review conference produces a negotiating mandate, whether the expert group's mandate is renewed when it closes at the end of the 2026 sessions, and whether any major power shifts its position at the General Assembly are all observable events with dates attached.^[9][11] The technical indicators, by contrast, sit inside classified programmes.