Defence

No binding rule governs autonomous weapons

In October 2025 NATO's parliamentarians called for governance of autonomous weapons. In January 2026 a Pentagon memo appears to have dropped the requirement for human judgment over lethal force. No institution has the standing to reconcile the two.

P J Laszkowicz

Two documents, three months apart, define the current era of military artificial intelligence.

On 13 October 2025 the NATO Parliamentary Assembly adopted Resolution 503, "Preparing NATO for the Future of Uncrewed Warfare", at its seventy-first annual session in Ljubljana. The resolution acknowledged that battlefield tactics were "rapidly evolving with increased deployment of uncrewed systems", drew lessons from the transformation of warfare in Ukraine, and called on NATO governments to build governance frameworks for autonomous weapons that uphold international humanitarian law and the alliance's own principles for responsible AI.1

On 9 January 2026 the United States Secretary of Defense, Pete Hegseth, released a memorandum titled "Artificial Intelligence Strategy for the Department of War", setting out AI as the decisive advantage in competition between major powers and directing that standardised "any lawful use" wording be written into every departmental AI services contract within 180 days. Human Rights Watch reports that the same January memorandum appears to have removed the longstanding requirement, set out in Department of Defense Directive 3000.09, that operators of autonomous weapons systems be able to exercise "appropriate levels of human judgment over the use of force". The hedge belongs in that sentence: the directive itself was still being formally revised, and what exactly has changed is not yet fully public.2

The gap between these two documents, one calling for governance and the other loosening it, is not a contradiction better drafting or more careful diplomacy could resolve. It is a structural feature of the present international system, in which no binding legal framework governs the development and deployment of autonomous weapons, no institution has authority to enforce one, and the states with the most advanced military AI programmes have the least reason to accept limits.

What follows traces how that void opened, what it means in practice, and why the room for closing it is shrinking.

The NATO framework

NATO's approach to military AI rests on six Principles of Responsible Use, endorsed by every ally: lawfulness, responsibility and accountability, explainability and traceability, reliability, governability, and bias mitigation.3 They were adopted alongside the alliance's first AI strategy in 2021 and reaffirmed in the revised strategy of July 2024, and they remain the most detailed statement of AI ethics by any military alliance in the world.

They are also aspirations rather than rules. Lawfulness requires adherence to national and international law, including international humanitarian law, without defining what lawful use of an autonomous weapon looks like in practice. Governability requires that systems stay under human oversight, "enabling human actors to interact with, and where appropriate, override or deactivate AI systems, preventing unintended autonomous actions", without specifying how much human involvement is needed, at what point in an engagement it must happen, or what follows when the pace of operations makes meaningful oversight impractical.4

Resolution 503 recognised those limits. The accompanying report from the Science and Technology Committee documented the growing role of uncrewed systems, drawn largely from Ukraine's experience, and identified governance and deployment problems the alliance has not resolved.5 The resolution asked NATO governments to build regulatory frameworks, invest in counter-drone and counter-autonomy technologies, and keep deployment consistent with international law.

NATO's structure limits what a resolution can achieve. The alliance sets standards and coordinates capabilities; it does not regulate the weapons policies of its members. The United States, its dominant military power, is simultaneously the most advanced in military AI and the least constrained by the alliance's principles. The memoranda that followed three months after Ljubljana showed that the distance between alliance governance and national action can widen at any moment, and that no mechanism exists to stop it.

The Hegseth doctrine

The January 2026 memoranda did not arrive from nowhere. They completed a year-long reorientation of American military AI policy that began on the president's first day in office.

On 20 January 2025, Executive Order 14110, which had established mandatory adversarial testing, safety protocols, and inter-agency collaboration for high-risk models, was rescinded.6 Three days later Executive Order 14179 replaced it, ordering an action plan built around "sustaining and enhancing the United States' global AI dominance".7 In July 2025 the White House published "Winning the Race: America's AI Action Plan", setting out more than ninety federal actions organised around accelerating innovation and building infrastructure.8 In December 2025 a further order directed the Attorney General to challenge state AI legislation the administration considered onerous.9

The Hegseth memoranda carried that trajectory from civilian policy into the military. The stated ambition is an AI-first warfighting force, in which artificial intelligence is not a tool the military uses but the foundation of its advantage.10 The "any lawful use" contract wording serves two purposes at once: it removes restrictions that vendors might place on how the military uses their systems, and it establishes that where a use of force is legal under American law, the AI enabling it should carry no additional constraint.

The reported change to Directive 3000.09 is the most consequential element, and also the least well documented. The original 2012 directive, updated in 2023, requires autonomous weapons systems to be designed so as to "allow commanders and operators to exercise appropriate levels of human judgment over the use of force".11 That wording has generally been read as requiring meaningful human involvement in any decision to use lethal force, described in the field as keeping a human in the loop, or at minimum on the loop. If the January memorandum did remove it, the operative constraint becomes legality rather than human judgment: where an autonomous engagement is lawful, whether a person took part in the decision becomes an operational matter rather than a governance requirement.12

Human Rights Watch reads the shift as a slide toward fully autonomous killing, arguing that dropping the human judgment standard admits weapons that select and engage targets without meaningful human control, which in its view international humanitarian law cannot accommodate. The organisation connects the change to a specific sequence of events: on 27 February 2026 the Department rejected AI company Anthropic's proposed ethical limits on military use, designated the company a supply chain risk, and signed instead with a competitor that accepted the "any lawful use" requirement.

The international law question

Whether autonomous weapons can comply with international humanitarian law is not a theoretical question. It is the central legal issue in military AI governance, and after more than a decade of multilateral negotiation it remains unsettled.

International humanitarian law imposes three main obligations on the use of force. Distinction requires separating combatants from civilians. Proportionality requires that expected civilian harm not be excessive against the anticipated military advantage. Precaution requires taking all feasible steps to minimise civilian harm.13

Each demands judgment. Distinction means identifying who is a combatant where the line is unclear: fighters wearing no uniform, civilians carrying weapons for self-defence, a building that is a school one hour and a command post the next. Proportionality means weighing military advantage against civilian cost, which depends on context, on the tactical situation, and on the value placed on an objective relative to human life. Precaution means choosing, from the available options, the method of attack that causes least civilian harm, which requires assessing alternatives in real time.14

The International Committee of the Red Cross has argued consistently that these obligations require human judgment, that the assessments are complex and depend on the circumstances prevailing when the decision to attack is made, and that combatants must therefore retain a level of human control over how a weapon functions in carrying out an attack.15 Its October 2025 position paper states that the loss of human control and judgement in the use of force "raises serious concerns from humanitarian, legal and ethical perspectives", and recommends new legally binding rules, including a prohibition on systems whose effects cannot be sufficiently understood, predicted, and explained, and a rule against using autonomous weapons to target human beings at all.

The opposing view, implicit in the Hegseth memoranda and explicit among some military legal scholars, is that international humanitarian law constrains the use of force rather than the means of deciding. On this reading, if an autonomous system can be shown through testing to apply distinction and proportionality as reliably as a human operator, the absence of a human from the decision is not itself a violation.16 The law requires that the rules be followed, not that a human follow them.

The argument has run inside the Group of Governmental Experts on lethal autonomous weapons systems, convened under the Convention on Certain Conventional Weapons, since 2014. In more than a decade it has produced guiding principles, working papers, and a rolling text revised many times, most recently in May and August 2025.17 It has not produced a binding instrument. Its mandate states that it works "without prejudging" the nature of any eventual outcome, which means the negotiations do not assume a treaty at the end of them.18

At the September 2025 session, Brazil delivered a joint statement on behalf of forty-two states declaring them ready to move to negotiations on the basis of the Chair's rolling text, which they considered a sufficient basis for an instrument. Those forty-two represent roughly a third of the 128 states party to the Convention, and four more joined them in November 2025. The ICRC and the UN Secretary-General have jointly called for such negotiations to conclude by the end of 2026. The Group works by consensus, however, and the states with the most advanced military AI programmes, meaning the United States, China, Russia, and Israel, have consistently opposed binding constraints.19 The Convention's Seventh Review Conference in 2026 will decide whether the Group's mandate is renewed, widened into a negotiating mandate, or abandoned.

What autonomy means in practice

Debates about autonomous weapons often run at a level of abstraction that hides what is actually being fielded. The range of autonomy in current systems is wide, and the line between autonomous and merely automated is less clear in practice than in policy documents.

At one end sit decision-support systems: tools that process sensor data, identify possible targets, and put recommendations to a human operator who makes the engagement decision. These raise little controversy under existing law, because the human remains in the loop and the system assists rather than decides.

At the other end sit fully autonomous engagement systems, which select and strike targets without human intervention once activated. No state publicly admits to fielding them—though the boundary is blurring. Russia's V2U loitering munition, detailed by Ukrainian military intelligence in June 2025, was first seen in September 2024 and first used in combat in February 2025 in Ukraine's Sumy region. It navigates without satellite positioning by comparing live video against stored terrain imagery, searches for and selects targets using onboard AI, detects incoming interceptor drones and manoeuvres to avoid them, and can coordinate with others in a swarm. By mid-May 2025 the Ukrainian analyst Serhii Beskrestnov estimated Russia was flying thirty to fifty of them a day across several sectors of the front, training the system in combat as it went.

Between the extremes lies a spectrum. Loitering munitions, which linger over an area and strike when conditions are met, are a growing category. Some require an operator to pick the target and trigger the strike; others home autonomously on radar emissions or satellite signals. The market was worth about $530 million in 2024 and is projected to reach $815 million by 2029, growing at 9 per cent a year. The direction of travel is toward more autonomy in identifying targets, pushed by better automatic target recognition and by the tactical reality of contested electromagnetic environments where communications links may not survive.

Drone swarms, groups of autonomous or semi-autonomous systems operating together, are the next step. Chinese military writing describes them as the next evolution in uncrewed intelligent warfare, citing advantages in cost, scalability, and resilience over single platforms.20 Several nations are investing in swarms that would let dozens or hundreds of small systems coordinate with minimal human oversight.

Military AI is advancing along a continuum, and each increment of autonomy makes the previous governance framework slightly more obsolete. The Hegseth memoranda accelerated that, and the underlying trend toward faster decision cycles, more distributed operations, and less human involvement predates any single policy.

The European position

The European Union's approach is defined by a conspicuous gap. The AI Act, the most comprehensive AI legislation in the world, excludes military applications from its scope entirely.

Article 2(3) states that the regulation does not apply to systems "placed on the market, put into service, or used with or without modification exclusively for military, defence or national security purposes".21 The exclusion rests on Article 4(2) of the Treaty on European Union, which reserves national security to the member states, and on the argument that public international law rather than EU regulation is the right instrument for military AI.22

The exemption creates a paradox. Europe has built the most detailed AI regulation anywhere, addressing bias, transparency, human oversight, and risk management with unusual specificity, and none of it applies to the domain where the consequences of failure are gravest. A facial recognition system sold commercially faces strict requirements. The same technology identifying targets for a military strike faces none of them.

The European Parliament has acknowledged the tension. A 2025 research service briefing noted the blind spots the military exemption creates and questioned whether international humanitarian law alone suffices when the pace of technological development outstrips the capacity of legal frameworks to adapt.23 Analysts have pointed out that the gap is sharpest for dual-use systems, developed for civilian purposes and adapted for military ones, which sit inside the Act's scope until the moment they are repurposed.24

Individual states are going their own way. France's military AI strategy stresses keeping human control over the use of force while investing in autonomous decision support. Germany's approach is cautious, reflecting public sensitivity to the subject. The European Defence Agency runs initiatives on interoperability and common standards. No European institution has either the mandate or the capacity to set binding rules for military AI across the continent.25

Who decides?

No binding international framework governs autonomous weapons. The European Union has exempted military AI from its most comprehensive AI law. NATO has principles rather than rules. The Group of Governmental Experts has negotiated for over a decade without a treaty. The United States appears to have removed its own internal requirement. Russia and China take part in multilateral discussion while fielding increasingly autonomous systems.

The result is a void of authority rather than of discussion. The question of when, whether, and on what conditions a machine may be permitted to decide to kill a person is being answered in practice by the states and companies building the systems, unconstrained by any binding rule.

The room to answer it differently is shrinking. The Seventh Review Conference in 2026 may be the last realistic chance for multilateral action before the technology is too deeply embedded in military doctrine to regulate. The forty-two states seeking a binding instrument are a substantial coalition, and they do not include the states whose participation any effective regime would require.

The ICRC's call for clear prohibitions and restrictions by the end of 2026 reflects a view that governance has to precede widespread deployment, because once autonomous systems are woven into operations at scale, the political and institutional cost of imposing limits afterwards becomes prohibitive.26 Nuclear weapons are the instructive precedent: arms control there was negotiated decades after deployment, and only once the destructive potential had been demonstrated. Whether the international community learns from that sequence or repeats it is the open question.

Resolution 503 and the Hegseth memoranda are two answers to it. One says govern first and deploy within the governance. The other says deploy first and govern if it proves necessary. The distance between them is where the autonomous weapons era is being settled, not in treaty negotiations or parliamentary resolutions but in the decisions of military commanders, defence contractors, and political leaders building the future of warfare faster than any institution can regulate it.

Footnotes

  1. NATO Parliamentary Assembly, "Uncrewed Warfare—Report," Science and Technology Committee, 2025. ↩

  2. Human Rights Watch, "US Military's Dangerous Slide toward Fully Autonomous Killing," 3 March 2026. HRW's own wording is that the memorandum "apparently removed" the requirement. As of this writing the revised text of Directive 3000.09 had not been published, and reporting through 2026 indicates the formal revision was still in progress, so the change is credibly reported rather than confirmed. ↩

  3. NATO, "Summary of NATO's Revised Artificial Intelligence (AI) Strategy," 10 July 2024. ↩

  4. NATO, "Summary of NATO's Autonomy Implementation Plan," 13 October 2022. ↩

  5. NATO Parliamentary Assembly, "Uncrewed Warfare—Report," 2025. ↩

  6. White House, Executive Order 14110, "Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence," 30 October 2023; rescinded 20 January 2025. ↩

  7. White House, Executive Order 14179, "Removing Barriers to American Leadership in Artificial Intelligence," 23 January 2025. ↩

  8. White House, "Winning the Race: America's AI Action Plan," July 2025. ↩

  9. White House, Executive Order on State-Level AI Legislation, December 2025. ↩

  10. Foreign Policy, "The Pentagon and Anthropic's High-Stakes Game of Chicken," 26 February 2026. ↩

  11. Department of Defense Directive 3000.09, "Autonomy in Weapon Systems," January 2023 update. ↩

  12. CSIS, "DOD Is Updating Its Decade-Old Autonomous Weapons Policy, but Confusion Remains Widespread," 2025. ↩

  13. Additional Protocol I to the Geneva Conventions, Articles 48, 51, 57. ↩

  14. ICRC, "Autonomous Weapon Systems under International Humanitarian Law," 2025. ↩

  15. ICRC, "Position on Autonomous Weapon Systems," International Review of the Red Cross, 2025. ↩

  16. War on the Rocks, "Autonomous Weapon Systems: No Human-in-the-Loop Required, and Other Myths Dispelled," May 2025. ↩

  17. UN Office for Disarmament Affairs, "GGE on Lethal Autonomous Weapons Systems," 2025. ↩

  18. Ibid. ↩

  19. ASIL, "Lethal Autonomous Weapons Systems & International Law: Growing Momentum towards a New International Treaty," 2025. ↩

  20. CNA Corporation, "PRC Concepts for UAV Swarms in Future Warfare," October 2025. ↩

  21. EU AI Act, Article 2(3). ↩

  22. EU AI Act, Recital 24. ↩

  23. European Parliamentary Research Service, Briefing on Autonomous Weapons, 2025. ↩

  24. European Security Think Tank, "Blind Spots in AI Governance: Military AI and the EU's Regulatory Oversight Gap," 3 October 2025. ↩

  25. European Defence Agency, "AI in Defence," 2025. ↩

  26. ICRC/UN Secretary-General Joint Call on Autonomous Weapons, 2023. ↩