Autonomous Weapons Policy: International Negotiations on Lethal AI Systems

Executive Summary

No specific treaty governs lethal autonomous weapons systems (LAWS), existing international humanitarian law remains the only legal safeguard. A decade of negotiations within the Convention on Certain Conventional Weapons (CCW) framework has produced eleven non-binding guiding principles but no binding instrument, as a handful of heavily militarised states including Russia, Israel, India, Australia, the Republic of Korea, and the United States have consistently blocked formal treaty negotiations. The CCW Seventh Review Conference, scheduled for November 2026, represents the moment when the GGE must submit its final report, making this the last realistic window before the pre-proliferation window closes. Both the technology and the diplomatic stakes are accelerating simultaneously: in June 2026, a senior figure in the Ukraine defense industry disclosed that, two years earlier, "fully autonomous drones with no human oversight" had killed soldiers on the battlefield for the first time, a disclosure that underscores how rapidly the regulation-technology gap is widening.

The Structural Veto And Why It Has Held

The 2025 versions of the rolling text include formulations on characterisation of LAWS, applicability of international humanitarian law to all autonomous systems, human judgement and control as essential for lawful use, prohibitions on inherently indiscriminate systems, and regulatory measures related to predictability, reliability, and operational limits. Convergence on these language elements is real. Yet convergence on language is not the same as willingness to accept binding obligations, and the distinction matters enormously.

The Usanas Foundation's analysis of the "Big Three" positions makes the strategic logic clear: while the majority of nations seek a preemptive ban to avoid a new arms race, the United States, Russia, and China have adopted positions that protect their ability to innovate and deploy these systems according to their own timelines. The Pentagon has requested a record $14.2 billion for AI and autonomous research for fiscal year 2026, representing a sunk-cost dynamic that makes acceptance of prohibitory language increasingly low confidence from Washington regardless of diplomatic goodwill.

China's position merits separate treatment because it differs from that of Russia and the United States. According to the Carnegie Endowment for International Peace, in the 2024 GGE session, China stated that in the absence of a clear definition and characterisation, a tiered approach to regulating LAWS based on its five technical characteristics may be useful. Beijing has declined to oppose the concept of regulation outright but shapes its engagement to avoid constraints on systems it considers defensive or whose autonomy it defines narrowly. China did not sign the US-led Political Declaration on Responsible Military Use of AI, though China's Foreign Ministry said Beijing would "remain open and constructive in working with other parties."

The interplay between technology competition and legal forum choice translates directly into diplomatic strategy. Reaching Critical Will has argued that the CCW's consensus requirement functions as a chokepoint that allows the blocking coalition to exhaust the pro-treaty majority through procedural delay while autonomous weapons programs mature on the ground. The CCW has not been immune to the overall deterioration of multilateralism, and the forum choice question is now explicitly contested: states are divided between those that favour continuing discussion solely within the CCW framework and those that prefer advancing discussions toward a legally binding treaty in other venues, such as the UN General Assembly.

What The Rolling Text Actually Contains, And What It Does Not

The current rolling text covers elements related to a working characterisation of LAWS, application of international humanitarian law to LAWS, prohibitions and restrictions, accountability and responsibility, and risk mitigation measures. The American Society of International Law's January 2025 review notes that the CCW rolling text from November 2024 outlines potential regulatory measures including: ensuring LAWS are predictable, reliable, traceable, and explainable; maintaining context-appropriate human oversight particularly during target identification and engagement; restricting target types, duration, geographical scope, and scale of operations; and enabling human operators to deactivate LAWS after activation.

What the text does not yet contain, the absence that matters most, is a definition that commands agreement across the blocking coalition. China characterised LAWS as weapons with complete autonomy that cannot be terminated once deployed, while Russia's 2022 characterisation viewed LAWS as prospective weapons that do not yet exist on the battlefield. A definition that treats LAWS as a future technology rather than a present one allows states to acknowledge the concept while escaping any obligation to restrict what they are already deploying. Experts at SIPRI have warned that this definitional asymmetry is not accidental; it is a studied diplomatic technique.

The broader systemic implication is that the accountability gap identified by the Lieber Institute at West Point will widen as systems become more capable: when a machine selects and attacks a target independently, current legal rules struggle to hold any specific person responsible for the resulting harm, and since IHL and International Criminal Law were built around human decisions, when a critical choice is made by an AI-enabled system instead of a person, it becomes extraordinarily difficult to attribute violations to commanders, operators, or programmers. This security and legal accountability dynamic spills into broader governance debates about AI regulation, since the absence of binding LAWS standards creates a permissive environment for commercial AI companies contracted by militaries, a cross-domain risk that the G7's June 2026 AI discussions at Evian-les-Bains, reported by Reuters, only partially addressed.

The Coalition Pushing For Negotiations, And Its Limits

At the September 2025 session of the GGE, 39 High Contracting Parties led by Brazil, along with three observer states, delivered a joint statement ready to move ahead towards negotiations on the basis of the rolling text. The signatories included Austria, France, Germany, Mexico, New Zealand, Pakistan, South Africa, Spain, Sweden, Switzerland, and Uruguay, among others. By November 2025, that alignment had grown: according to Stop Killer Robots, 42 states delivered a joint statement at the most recent GGE session calling for the group to begin negotiations of an instrument on the basis of its rolling text.

The African Group has been among the most consistent advocates. The African Group issued a statement in 2018 urging the urgent launch of negotiations on a legally binding instrument to regulate fully autonomous weapons, and condemned the idea of transferring life-or-death decisions to machines as "inhumane, abhorrent, repugnant, and against public conscience." Germany's coalition agreement from 2021 stated explicitly that it rejects lethal autonomous weapon systems that operate entirely without human control and actively advocates for their international prohibition.

Yet this coalition faces a coordination problem. The ICT4Peace Foundation assessment from October 2025 identified the critical fault line: if talks stall, existing international humanitarian law will remain the only safeguard, leaving many worried that technology will advance faster than the rules designed to contain it. Even states that vocally support negotiations have not yet been willing to abandon the CCW forum and pursue a treaty outside it, as was done with the Ottawa Convention on anti-personnel mines. That outside-the-forum strategy worked for landmines precisely because major mine-producing states eventually joined, a dynamic low confidence to replicate for LAWS given the scale of current US, Chinese, and Russian military AI investment.

Counterarguments

1. The CCW may be more productive than critics assess, and the outside-treaty option carries its own risks. The most sustained challenge to the core finding is that the rolling-text process has actually produced meaningful convergence that a rushed outside-the-CCW treaty would sacrifice. The RSIS assessment noted that pursuing normative discussions outside the GGE may prove ineffective, as major militaries developing autonomous weapons support the GGE as the appropriate forum, and that a parallel process risks fragmenting rather than accelerating governance. A treaty signed by 130 states but not ratified by the US, Russia, and China would resemble the Ottawa landmines model, normatively significant but operationally incomplete, since the states with the largest LAWS programs remain unconstrained.

2. The accountability-gap argument overstates the distinction between LAWS and existing semi-autonomous systems already governed by IHL. The Lieber Institute and others have argued that the accountability problem is qualitatively new. But the Parliamentary Assembly of the Council of Europe and several national militaries point out that Article 36 of Additional Protocol I already requires states to review new weapons for IHL compliance, obligating evaluation of predictability and reliability to function as intended, without errors or unintended consequences. If Article 36 reviews were robustly conducted, some of the accountability gap would narrow under existing law. The counter to the counter is that Article 36 reviews are self-assessed and non-transparent, but that is a compliance problem, not a legal gap.

3. The definition problem may be irresolvable within any multilateral forum, not just the CCW. A structural blind spot in pro-treaty analysis is that the definitional asymmetry identified across major powers is not merely a diplomatic tactic; it reflects genuine technical disagreement about what autonomous targeting means. In weapons development, the term "autonomous" is somewhat ambiguous and can vary hugely between different scholars, nations, and organisations, and there is no definition of lethal autonomous weapon systems that is generally agreed upon among different countries. A treaty without an agreed definition has limited operational meaning: parties could ratify while disagreeing fundamentally about which of their existing or planned systems it covers. The Nuclear Non-Proliferation Treaty's definitional clarity was a prerequisite for its partial success; LAWS governance lacks an equivalent baseline.

Securitisation Analysis

Securitising Actor: The UN Secretary-General, the ICRC, and the pro-treaty state coalition (led by Austria and Brazil) have functioned as the primary securitising actors. UN Secretary-General Guterres stressed "the need to act urgently to preserve human control over the use of force" and called for a "legally binding instrument to prohibit lethal autonomous weapons systems that function without human control or oversight and that cannot be used in compliance with international humanitarian law," characterising autonomous targeting of humans by machines as "a moral line that must not be crossed."

Referent Object: The core referent object is human dignity and the requirement that life-or-death decisions in warfare remain subject to human moral agency. A secondary referent object is global strategic stability, framed through the arms-race risk. The Usanas Foundation noted that the 2026 deadline is increasingly seen as the "finish line" for global diplomacy; if a treaty is not reached, the speed of innovation in military AI driven by the very powers currently blocking the UN's progress will moderate-to-high confidence make any future regulation obsolete.

Existential Threat Construction: The ICRC and Stop Killer Robots have framed LAWS as crossing a threshold that existing political processes cannot address at normal speed, invoking the language of irreversibility. The Campaign to Stop Killer Robots warns that once autonomous weapons systems are widely deployed and the idea that machines decide who lives and who dies is normalised, it will be far harder to impose restrictions. Stuart Russell, professor of computer science at UC Berkeley, drew a parallel in the Guardian in June 2026 to nuclear and radiological licensing, arguing that a licensing regime requiring minimum safety standards before deployment should apply to AI systems used in warfare.

Target Audience: The securitisation move is directed at CCW High Contracting Parties, UN General Assembly majorities, and domestic publics in states with significant civil society influence over defence procurement decisions.

Extraordinary Measures Proposed: The two-tiered approach advocated by the UN Secretary-General and ICRC, prohibiting systems that cannot comply with IHL while regulating all others, represents a qualitatively new instrument rather than an adjustment to existing CCW protocols.

Classification: POLITICIZED. The existential framing has been advanced by credible norm entrepreneurs with institutional backing, and voting patterns indicate overwhelming supermajority agreement that action is needed. But the blocking coalition's resistance has prevented the issue from reaching the securitised threshold where extraordinary measures are operational. The issue remains within the bounds of contested multilateral politics rather than emergency action.

Classification: POLITICIZED

Process Tracing Analysis

Cause and Outcome: The cause is the CCW's consensus rule combined with major-power blocking behaviour. The outcome under assessment is the absence of a binding LAWS instrument after more than a decade of negotiations.

Causal Mechanism Chain: Step 1, CCW consensus rule established in 1980 as a feature of the treaty framework. Step 2, 2016 Fifth Review Conference creates the GGE, opening formal LAWS discussions within the CCW. Step 3, The United States, Russia, and allied states participate actively in the GGE while consistently opposing any negotiating mandate, using the consensus requirement to steer outcomes toward guiding principles rather than binding rules. Step 4, Each successive GGE session produces convergence on language without a mandate to negotiate, with the mandate repeatedly extended rather than resolved. Step 5, The pro-treaty majority pursues parallel legitimacy through UNGA resolutions but cannot generate binding force without major-power participation. Step 6, Technology deployment continues during the negotiating stalemate, increasing the sunk-cost dimension of blocking-state positions.

Evidence Assessment: The blocking-state posture constitutes a hoop test that has been passed consistently across administrations and geopolitical environments, Russia's position under the 2022 Ukraine invasion context, the US position under both Biden and Trump administrations, and Israel's position during active conflict deployment all confirm pattern consistency. The UNGA voting margins (156-5 in November 2025 per Stop Killer Robots) constitute a smoking gun for supermajority political will, but that evidence is consistent with continued blocking since UNGA resolutions are non-binding. The absence of a single binding instrument after a decade of formal negotiations is doubly decisive evidence that the consensus veto mechanism is functioning as a structural barrier.

CAUSAL_MECHANISM_STRENGTH: STRONG

Indicators To Watch

The table below identifies observable signals that would indicate whether the LAWS governance trajectory is moving toward a binding instrument, continued stalemate, or a forum shift.

Decision Relevance

The November 2026 CCW Seventh Review Conference is the operational decision point for corporate strategists, risk managers, and policy researchers assessing legal and reputational exposure in the defence AI space.

Scenario A (~55%): Stalemate at CCW Review Conference, UNGA process gains momentum - The Seventh Review Conference produces a GGE report with elements of convergence but no consensus mandate to negotiate a binding instrument. Russia and the United States block agreement. The UNGA majority, building on the 156-state November 2025 resolution, commences a parallel negotiating process outside the CCW, similar to the Ottawa Treaty model. A binding instrument is adopted by a broad majority but not ratified by the United States, Russia, or China within a 3-5 year horizon. Recommended: Companies in the defence AI sector should establish legal review mechanisms now for Article 36 weapons-review compliance in all jurisdictions where they operate, anticipating that a non-CCW treaty will generate domestic implementing legislation in signatory states even absent major-power ratification. Reputational risk management for commercial AI firms contracted by militaries should treat the pro-treaty majority as a growing compliance constituency.

Scenario B (~30%): CCW Seventh Review Conference produces a negotiating mandate - A combination of Chinese position softening, UK-European pressure on the US, and the political impact of continued LAWS-related civilian casualties creates conditions for a consensus negotiating mandate at the November 2026 Review Conference. Negotiations begin in 2027 with a multi-year timeline. Recommended: Defence-sector investors and strategists should treat this as a valuation inflection point for autonomous weapons hardware programs. Companies with significant LAWS-adjacent revenue would face a 2-4 year window of regulatory uncertainty. Engagement with the GGE rolling text's five "boxes", characterisation, IHL applicability, prohibitions, accountability, and risk mitigation, should inform product development roadmaps now, since any eventual treaty is moderate-to-high confidence to reflect those categories.

Scenario C (~15%): Forum collapse, technology sets de facto norms - CCW Seventh Review Conference fails procedurally; UNGA parallel process stalls due to major-power resistance; no new international instrument is agreed within the decade. LAWS proliferate as costs decline, with operational experience in Ukraine and Gaza establishing battlefield norms that substitute for legal ones. Recommended: Policy researchers and security professionals should treat conflict-zone AI deployment data as the primary evidence base for norm formation. The Lieber Institute's strict-liability framework for LAWS, treating developers and deploying states as legally responsible for system behaviour, is the most moderate-to-high confidence avenue for legal accountability claims in the absence of a treaty, and should be factored into corporate legal exposure assessments for defence AI contractors operating in active conflict zones.

Analytical Limitations

• The GGE's second 2026 session is scheduled for 31 August to 4 September 2026, after the date of this assessment; its outcome will be determinative for the November Review Conference and cannot be predicted from current evidence. Any significant position shift by a blocking state at that session would require this assessment to be revised.
• China's position is the least transparent of the three major blocking or ambiguous powers. Available evidence from the Carnegie Endowment and CCW working papers indicates engagement with definitional frameworks, but China's actual flexibility on prohibitions versus regulations is unknown. If China shifted to active support for a negotiating mandate, the blocking coalition's consensus veto would weaken materially.
• The accountability-gap assessment rests primarily on open-source legal analysis from the Lieber Institute, ASIL, and Case Western Reserve Journal of International Law. Classified military legal reviews by deploying states, which would reveal how Article 36 weapons reviews are actually being conducted for AI-enabled systems, are unavailable. If such reviews are substantive and documented, the urgency of the legal gap assessment would be partially reduced.
• The battlefield deployment evidence from Ukraine and Gaza is fragmentary. The June 2026 disclosure to New Scientist regarding autonomous drone killings was a single-source industrial claim. Independent verification of the degree of human control, or its absence, in active LAWS deployments remains limited, which affects the causal argument linking battlefield evidence to negotiating urgency.
• Potential anchoring bias toward the CCW as the primary forum should be noted; the Ottawa and Oslo Convention precedents demonstrate that effective arms control sometimes bypasses consensus forums entirely, and this assessment may underweight the speed at which a non-CCW process could develop if major European states commit to it.