Autonomous Weapons and the Future Battlefield: How Much Human Control Will Remain

Executive Summary

The character of warfare is being reshaped by a generation of weapons that increasingly select and, in a growing number of cases, engage targets with diminishing real-time human input. This is not a speculative future. It is an observed feature of the wars in Ukraine and Gaza, of Chinese and American military modernization programs, and of a South Asian arms competition that has already put autonomy-adjacent systems into combat use. The defining strategic question is no longer whether autonomous weapons systems (AWS) will exist, but how much human judgment will remain embedded in the decision to kill, and who will bear responsibility when that judgment fails.

Three forces are converging to compress human control. First, battlefield necessity: electronic warfare has made radio-linked drones increasingly unreliable, pushing militaries toward onboard artificial intelligence that can complete an attack after communications are jammed. Second, industrial economics: mass-produced, software-defined autonomous systems promise to offset the cost and manpower disadvantages that constrain Western militaries relative to China’s industrial base and Russia’s expendable-force doctrine. Third, institutional inertia in governance: a decade of United Nations deliberation under the Convention on Certain Conventional Weapons (CCW) has produced extensive diagnosis but no binding restraint, and its governing mandate concludes in 2026 without a clear successor mechanism.

This assessment examines the military, political, economic, technological, and legal dimensions of that convergence, drawing on developments through the concluding 2026 session of the CCW Group of Governmental Experts on Lethal Autonomous Weapons Systems (GGE on LAWS), the American update to Department of Defense Directive 3000.09, the maturation of Chinese “intelligentized warfare” doctrine, and the accelerating European defense-technology buildout led by firms such as Helsing and Anduril. It concludes that the erosion of human control is proceeding unevenly and asymmetrically: fastest in the terminal guidance of strike munitions, slower in the authorization of lethal force against personnel, and least advanced — but most contested — in the delegation of target selection itself. Where this balance settles over the next five years will shape not only battlefield outcomes but the future of international humanitarian law as a constraint on the use of force.

Key Judgments

High confidence. Autonomy in the terminal (final-attack) phase of strike munitions — where a weapon completes an engagement after losing its communications link — is now an operational reality in the Russia-Ukraine war and is diffusing rapidly to other militaries, including Pakistan and India. This form of autonomy is the least legally and ethically contested and is unlikely to be curtailed by any near-term international agreement.
Moderate-to-high confidence. No major military power currently fields, or has publicly acknowledged fielding, a weapon system that autonomously selects previously unspecified human targets and engages them without a human decision point beforehand. The most controversial reported cases — Israel’s Lavender and Gospel/Habsora systems in Gaza — involve AI-generated target recommendations subject to nominal human sign-off, not fully self-directed lethal autonomy, though critics argue the speed and volume of that process has hollowed out meaningful review.
High confidence. The multilateral effort to negotiate a binding international treaty on lethal autonomous weapons systems will not conclude successfully by the end of 2026, despite the United Nations Secretary-General’s explicit call for such an outcome. The CCW Group of Governmental Experts’ three-year mandate concludes with its final report to the Seventh Review Conference in November 2026, but consensus-based structure and firm opposition from major military powers to a categorical ban make a binding instrument within this timeframe very unlikely.
Moderate confidence. The United States, under National Security Presidential Memorandum-11 (June 2026), is recalibrating its own domestic policy — the update to DoD Directive 3000.09 — toward faster adoption and reduced procedural friction rather than tighter constraint, while formally retaining language on human judgment over the use of force. This signals a widening gap between the restraint-oriented rhetoric of multilateral diplomacy and the acceleration-oriented reality of national policy among leading military AI powers.
Moderate confidence. China’s “intelligentized warfare” (智能化战争) doctrine treats autonomy as a means to compress decision cycles and achieve “cognitive dominance,” with the People’s Liberation Army pursuing autonomy across unmanned aerial, surface, and underwater domains ahead of its 2027 centenary modernization target. Command-level AI decision-support (used for planning rather than direct target engagement) is already in use, which is a meaningfully lower step on the autonomy ladder than commonly assumed by Western commentary.
Moderate confidence. Europe’s defense-technology sector, led by Germany’s Helsing and supported by a record $49.1 billion in 2025 global defense-tech venture funding, is emerging as a third pole of autonomous weapons development alongside the United States and China — though publicly reported combat performance data (such as leaked German government assessments of drone strike-success rates) suggests the technology remains less mature than commercial and political messaging implies.
Low-to-moderate confidence. Absent a binding treaty, the most probable governance outcome through 2030 is a patchwork of national policies, voluntary political declarations, and informal battlefield norms — an outcome that preserves general-purpose flexibility for state militaries but leaves the “accountability gap” identified by international humanitarian law scholars substantially unresolved.

Introduction

For more than a decade, the prospect of “killer robots” — weapons that could kill without a human pulling the trigger — occupied a speculative, almost science-fictional register in strategic debate. That era has ended. The wars in Ukraine, Gaza, Nagorno-Karabakh, and the South Asian subcontinent have each, in different ways, demonstrated that meaningful elements of autonomous target selection and engagement are already present on contemporary battlefields, even where full “human-out-of-the-loop” lethality against human targets remains rare and contested.

This assessment addresses a specific and consequential question: as artificial intelligence becomes embedded across the targeting cycle — from detection and identification through recommendation, authorization, and terminal engagement — how much of that cycle will continue to require a human decision, and how much will migrate to the machine? The answer varies by domain, by military culture, and by the type of target involved, and it is evolving in real time as the CCW’s diplomatic process reaches a decisive juncture in 2026.

The analysis proceeds in five parts. It first traces the historical arc from early automated defensive systems to today’s AI-guided strike drones. It then situates the current moment within the broader strategic competition among the United States, China, Russia, and a resurgent European defense-industrial base. A detailed, multi-dimensional analysis follows, covering the military, political, economic, technological, and legal-diplomatic aspects of the autonomy question. The assessment closes with scenario analysis, probability estimates, and a strategic outlook for policymakers, militaries, and international institutions.

Historical Background

Automation in weapons systems is not new. Naval close-in weapon systems such as the U.S. Phalanx, first fielded in the 1980s, have long operated in automatic modes that detect and engage incoming missiles faster than a human could react — a narrow, defensive form of autonomy that has attracted little controversy because it engages objects, not people, in a constrained and reversible engagement envelope. Israel’s Harpy loitering munition, developed in the 1980s and 1990s, extended this logic offensively: it could loiter over a wide area and autonomously attack radar emissions matching a pre-set signature, without a human confirming each individual target.

What changed after 2020 was the application of this logic to more ambiguous, human-populated environments, and the diffusion of the underlying technology far beyond a handful of advanced militaries. In March 2020, a Turkish-made STM Kargu-2 loitering munition was reportedly used against retreating forces aligned with Libyan General Khalifa Haftar. A United Nations Panel of Experts report described the munitions as having been programmed to attack without requiring a live data link between operator and munition — a “fire, forget and find” capability. Independent analysts have since disputed whether the system was actually operating autonomously at the time, and whether the attack caused casualties, but the episode crystallized global attention on the possibility that a lethal autonomous engagement against human targets had already occurred.

The 2020 Nagorno-Karabakh war between Armenia and Azerbaijan further demonstrated the battlefield value of loitering munitions and drone-artillery integration, though questions about the degree of autonomy involved remained similarly unresolved. It was, however, Russia’s full-scale invasion of Ukraine in February 2022 that transformed autonomy from a niche capability into a mass-produced battlefield necessity. Confronted with pervasive Russian electronic-warfare jamming that severed the radio links between operators and first-person-view (FPV) drones, Ukrainian developers — operating through the state-backed Brave1 defense-innovation platform — began fielding drones with onboard terminal-guidance autonomy: once an operator locks onto a target, the drone can complete its attack run using computer vision even if its communications link is jammed. What began as an emergency adaptation against electronic warfare has, by 2026, become a defining feature of drone design on both sides of the front line, with Russia fielding comparable AI-assisted and fiber-optic-guided systems of its own.

In parallel, the Israel-Hamas war that began in October 2023 introduced a different and more controversial form of AI integration: not terminal weapons guidance, but AI-assisted target identification and recommendation at scale. Investigative reporting on Israeli systems known as “Gospel” (Habsora) and “Lavender” — which the Israel Defense Forces has partially acknowledged while disputing critics’ characterizations — revealed how machine-learning models were used to generate very large target lists (buildings in the case of Gospel, individuals in the case of Lavender) for human review and approval, compressing a targeting process that once took days into a process measured in minutes per target. This episode reframed the human-control debate: the central risk was not necessarily a machine pulling the trigger, but a human review process so rapid and voluminous that it functioned, in practice, as a rubber stamp.

By 2026, these threads — jamming-resistant terminal autonomy in Ukraine, AI-scaled target generation in Gaza, and accelerating great-power investment in autonomous platforms — had converged into the central preoccupation of military planners, arms-control diplomats, and defense economists worldwide.

Strategic Context

Analysts and policymakers commonly describe the human role in autonomous weapons along a three-point spectrum: human-in-the-loop, in which a person must authorize each engagement; human-on-the-loop, in which a system can act autonomously but a human supervises and can intervene or abort; and human-out-of-the-loop, in which the system selects and engages targets with no real-time human involvement at all. Chinese military theorists, including technical experts affiliated with the Central Military Commission, have described this spectrum explicitly as an evolutionary path along which autonomous systems are expected to progress — from assisted to supervised to delegated control — as trust in and battlefield necessity for such systems grow.

Three structural drivers are pushing military organizations along this spectrum faster than international governance can adapt.

The jamming imperative. Radio-frequency communications between an operator and a remotely piloted weapon are now a battlefield vulnerability rather than an advantage. Both Russian and Ukrainian forces have invested heavily in electronic-warfare systems designed to sever these links, and both have responded by pushing autonomy into the terminal phase of the attack specifically to defeat jamming. This dynamic is self-reinforcing: as EW capability improves, so does the incentive to remove the human-in-the-loop dependency that EW exploits.

Mass-production economics. China’s industrial base and Russia’s willingness to sustain high attrition rates have placed Western militaries at a numbers disadvantage that autonomous, low-cost, “attritable” systems are meant to offset. The Pentagon’s Replicator Initiative — since folded into the Defense Autonomous Warfare Group — was explicitly conceived as a counter to Chinese military mass through cheap, expendable, autonomous platforms fielded at scale. The same logic underlies the “drone wall” concept under discussion by the European Union and the rapid rise of European defense-technology firms building explicitly for volume production.

Great-power doctrinal competition. For China, intelligentized warfare is framed as a means of achieving “decision superiority” — compressing the observe-orient-decide-act (OODA) loop faster than an adversary’s human-centric command structure can match, and disrupting an opponent’s cognitive and psychological posture as much as its physical forces. For the United States, the policy debate under the current administration has shifted toward emphasizing the risk of strategic disadvantage from excessive caution, rather than the risk of unconstrained autonomy — a shift formalized in NSPM-11’s direction to accelerate national-security AI adoption while nominally preserving chain-of-command control.

Against this backdrop, the CCW’s decade-long deliberative process — grounded in incremental consensus-building among more than 125 High Contracting Parties — is reaching the end of its current mandate in 2026 without having resolved even the most basic definitional question: what, precisely, counts as a “lethal autonomous weapons system” for purposes of international law.

Detailed Analysis

Military Dimension

Ukraine: the live-fire proving ground. The Russia-Ukraine war has functioned as the largest real-world test bed for autonomous and semi-autonomous weapons in history. Ukrainian officials describe AI tools now processing tens of thousands of frontline video feeds monthly to identify, geolocate, and prioritize targets. FPV strike drones fitted with autonomy modules — developed by companies operating under Ukraine’s Brave1 innovation platform — can lock onto a target and complete the terminal attack run without a continuous control link, defeating the electronic-warfare systems that have historically been Russia’s most effective counter-drone measure. Intermediate-range strike drones with a 30–200 kilometer range, using AI guidance to home in on targets with reduced vulnerability to jamming, struck Russian logistics targets at a rate of roughly 130 assessed strikes in a single month (May 2026) across southern and eastern Ukraine, a rate that has trended upward since the start of the year. Analysts caution that these systems remain “AI-enabled” rather than fully autonomous in the strict sense — a human still designates the target before the terminal phase begins — but the operational trend toward reduced human involvement during the engagement itself is unambiguous and, per Ukrainian and Western officials, deepening.

Gaza: AI-scaled target generation. Israel’s use of AI-assisted targeting systems presents a materially different model of human-control erosion: not autonomous weapons in the classical sense, but an AI-accelerated human decision process operating at a volume and speed that critics argue undermines the substance of human review even while its form is preserved. According to investigative reporting that the IDF has partly acknowledged and partly disputed, the Lavender system generated recommendations that identified an estimated 37,000 individuals as suspected low-ranking militants, with the Israeli military’s own assessment citing a roughly 90 percent accuracy rate — implying a false-positive rate of approximately one in ten flagged individuals by the military’s own internal standard. The Gospel/Habsora system, by contrast, generated recommendations for structures rather than individuals, compressing a targeting process that previously required days of multi-source analysis into a matter of hours. The IDF has characterized these tools as auxiliary aids that assist, rather than replace, human analysts, and disputes the specific casualty-threshold figures reported by investigative journalists. Independent of how that dispute is resolved, the episode has become the most frequently cited real-world case study in the international debate over whether rapid, AI-scaled human review can still satisfy the substantive judgment that international humanitarian law presumes.

China: intelligentized warfare and the 2027 horizon. The People’s Liberation Army’s pursuit of autonomy is doctrinally comprehensive rather than confined to individual weapons platforms. Xi Jinping’s 2022 Party Congress report and subsequent instructions — including direction in mid-2026 to accelerate military application of unmanned intelligent technology — have elevated “intelligentization” to the PLA’s guiding modernization concept, alongside its 2027 centenary target. China’s military-civil fusion strategy channels commercial AI capability from firms such as Baidu, Alibaba, Tencent, and DJI into military applications, and the PLA has fielded autonomous or semi-autonomous systems across the aerial (GJ-11 and WZ-7 unmanned combat platforms, loyal-wingman concepts pairing crewed J-20/J-35 fighters with unmanned “mothership” formations), naval (multiple classes of autonomous surface and underwater vehicles), and ground (quadrupedal “robot wolf” platforms) domains. Notably, recent PLA use of AI has focused on command-level strike planning — recommending force packages and timing to human commanders rather than directly selecting and engaging targets — indicating that, at least in its most visible applications to date, China’s most advanced military AI use sits at a lower rung of the autonomy ladder than its doctrinal rhetoric might suggest. PLA analysts have also examined autonomous systems as a specific solution to the anticipated difficulty of urban warfare in a Taiwan contingency, where dense terrain and communications degradation would complicate traditional command and control.

The United States: from Replicator to CCA. After the Pentagon’s Replicator Initiative fell well short of its 2023 goal of fielding thousands of autonomous systems by August 2025 — the Congressional Research Service found only “hundreds” had materialized — the program was formally dissolved in late 2025 and absorbed into the newly created Defense Autonomous Warfare Group (DAWG) under U.S. Special Operations Command. The most technologically ambitious American program remains the Collaborative Combat Aircraft (CCA) initiative, which awarded prototype contracts to Anduril (the YFQ-44A) and General Atomics for autonomous “loyal wingman” aircraft intended to operate alongside crewed fighters, with autonomy and AI-piloting software contracts issued to RTX and Shield AI (whose HIVEMIND system underpins the effort). A $100 million “Orchestrator Prize Challenge,” launched in early 2026, reflects a shift in Pentagon thinking away from simply procuring large numbers of drones and toward solving the harder problem of commanding heterogeneous autonomous fleets through natural-language, effects-based instructions rather than granular manual control — an implicit acknowledgment that command-and-control software, not the platforms themselves, has been the primary bottleneck to fielding autonomy at scale.

Europe’s autonomy buildout. European militaries, propelled by the war in Ukraine and doubts about the durability of the American security guarantee, have made autonomous systems a centerpiece of rearmament. Munich-based Helsing — valued at roughly $18 billion following a 2026 funding round — has become Europe’s most prominent defense-technology firm, supplying more than 4,000 HX-2 strike drones to Ukrainian forces alongside autonomous underwater and AI-assisted electronic-warfare products, and holding a role in the pan-European Future Combat Air System. Reported field performance has been mixed: an internal German Ministry of Defence briefing, cited in press reporting, indicated the HX-2 achieved its target in five of fourteen missions in the Donbas in early 2026 — a data point that underscores the gap between promotional messaging and combat reliability still typical of this technology generation. At the European Union level, the proposed “European Drone Defence Initiative,” colloquially termed a “drone wall,” aims for initial counter-drone capacity along the EU’s eastern frontier by the end of 2026 and full operational capability by the end of 2027, in direct response to repeated Russian drone incursions into EU and NATO airspace.

South Asia: a parallel, lower-cost trajectory. Pakistan and India illustrate how autonomy-adjacent capability is diffusing well beyond the three leading military-AI powers. At Pakistan’s inaugural Indus Robotics and Autonomous Systems Expo in mid-2026, state and private firms displayed an expanding portfolio of loitering munitions, interceptor drones, and autonomous surface vessels. India, for its part, employed the indigenously developed Nagastra-1 loitering munition during the May 2025 India-Pakistan crisis and has since articulated a “Sindoor 2.0” concept explicitly built around AI-enabled fusion centers and manned-unmanned teaming. Analysts of the rivalry note that both states used drones during the 2025 crisis not merely for surveillance but for coercive signaling — a qualitative escalation that, combined with the compressed decision timelines autonomous systems enable, raises distinct crisis-stability concerns in a nuclearized dyad where misattribution or miscalculation carries exceptionally high stakes.

Political Dimension

The politics of autonomous weapons divide along a familiar but hardening fault line: states that possess or are close to possessing advanced military AI capability generally favor voluntary, principles-based governance that preserves operational flexibility, while a much larger coalition of small and middle powers, backed by UN leadership and civil society, favors a binding prohibition-and-regulation treaty.

In the United States, the governing instrument remains DoD Directive 3000.09, “Autonomy in Weapon Systems,” first issued in 2012 and updated in January 2023 to require that autonomous and semi-autonomous weapon systems be designed to allow commanders and operators to exercise “appropriate levels of human judgment over the use of force” — language that notably does not require a human-in-the-loop for every engagement and does not categorically prohibit fully autonomous lethal systems, provided they pass a senior-level review process (waivable by the Deputy Secretary of Defense in cases of urgent military need). On June 5, 2026, President Trump signed National Security Presidential Memorandum-11, directing an update to this directive within 90 days — a deadline falling almost exactly at the time of this assessment’s publication — with instructions to review it annually thereafter “to account for the rapidly evolving capabilities of AI systems.” Analysts broadly read NSPM-11 as tilting toward faster adoption and reduced procedural friction, particularly for non-lethal or lower-risk autonomous applications, while formally preserving human-judgment language for lethal systems. This American policy evolution is occurring under an administration that in September 2025 also issued an executive order authorizing the Department of Defense to use “Department of War” as a secondary public-facing title — a symbolically significant, if legally non-binding, rebranding (full statutory renaming requires congressional action) that reflects a broader institutional posture favoring offensive readiness over defensively framed caution.

At the multilateral level, the CCW’s Group of Governmental Experts on LAWS — chaired by Ambassador Robert in den Bosch of the Netherlands — held its concluding 2026 session from August 31 to September 4, closing out a three-year mandate to “further consider and formulate, by consensus, a set of elements of an instrument, without prejudging its nature.” The group’s final report is due to the CCW’s Seventh Review Conference, scheduled for November 16–20, 2026, in Geneva, where High Contracting Parties must decide whether to move toward negotiating a legally binding instrument, extend the group’s discussion-only mandate, or allow the process to lapse altogether. This decision point coincides with, but sits apart from, a parallel and more assertive push at the UN General Assembly: Secretary-General António Guterres, in his “New Agenda for Peace,” has called for states to conclude a legally binding instrument by 2026 that would prohibit lethal autonomous weapons systems operating without human control and regulate all others — a two-tier structure mirrored in the treaty framework proposed by the Stop Killer Robots campaign and the Harvard Law School International Human Rights Clinic, built around the concept of “meaningful human control.” A 2024 UN General Assembly resolution on the subject passed by a wide 166-to-3 margin, and more than 120 states now formally support opening treaty negotiations. The United States has consistently opposed a categorical ban, arguing that properly governed AI can improve targeting precision and reduce collateral harm, and favors continued reliance on voluntary instruments such as the 2023 Political Declaration on Responsible Military Use of Artificial Intelligence and Autonomy. Russia, China, and India have likewise resisted binding prohibitions, though for somewhat different reasons rooted in their own development trajectories and strategic calculations regarding future capability.

Economic Dimension

The economics of autonomous weapons are reshaping defense-industrial structures on both sides of the Atlantic and reordering the traditional dominance of legacy prime contractors. Global defense-technology venture funding reached a record $49.1 billion in 2025, according to industry tracking, with mainstream venture capital increasingly abandoning earlier ethical reservations about defense investment in favor of a “supporting democratic values” framing. Helsing’s rise from a 2021 startup to an $18 billion valuation in 2026 exemplifies this shift, as does the emergence of comparably fast-growing firms such as Germany’s Quantum Systems and Portugal’s Tekever. In the United States, Anduril Industries has positioned itself as the central private-sector beneficiary of the shift from Replicator to the Collaborative Combat Aircraft and counter-drone programs, holding contracts spanning autonomous aircraft prototypes, a $642 million Marine Corps counter-drone system, and a ten-year, $1 billion Special Operations Command counter-drone contract.

The strategic logic underpinning this capital flow is explicitly economic as much as it is military: autonomous, software-defined, mass-producible systems are intended to restore a cost-exchange ratio favorable to the side fielding them, offsetting China’s manufacturing scale and Russia’s tolerance for attritable losses. Whether this promise is being realized is contested. The Pentagon’s own experience with Replicator — falling short of its unit-fielding targets by an order of magnitude within its initial timeline — and reports of inconsistent field performance from European systems such as Helsing’s HX-2 both suggest that the translation from venture-funded prototype to reliable, mass-producible battlefield capability remains harder, slower, and more expensive than procurement rhetoric has implied. For defense budgets across NATO and Indo-Pacific allies, the practical economic question through the remainder of this decade will be whether autonomous systems genuinely substitute for, or merely supplement, existing high-cost crewed platforms — a question with major implications for force structure, industrial policy, and alliance burden-sharing.

Technological Dimension

The core technological driver of reduced human control is the maturation of onboard computer-vision and machine-learning models capable of operating without a continuous communications link — precisely the capability required to defeat electronic-warfare jamming. Ukrainian developers describe a “ladder of autonomy” spanning last-mile terminal strike guidance, last-mile autonomous bombing, target detection and classification, autonomous navigation in GPS-denied environments, and fully autonomous takeoff and landing, with most fielded systems concentrated at the lower rungs of this ladder (terminal guidance and navigation) rather than the more contested upper rungs (independent target selection, as opposed to target completion).

A second technological frontier — and arguably the harder engineering problem — is swarm orchestration: enabling many heterogeneous autonomous platforms from different manufacturers to coordinate toward a single commander’s intent, rather than being individually piloted. The U.S. Replicator Initiative’s most consequential shortfall was not platform procurement but the absence of software capable of orchestrating swarms of diverse systems, a gap the Pentagon’s 2026 Orchestrator Prize Challenge is explicitly designed to close. Reliability and error rates remain a significant technological and ethical constraint: the Israeli military’s own cited 90 percent accuracy figure for the Lavender system implies a roughly 10 percent false-positive rate at a scale of tens of thousands of generated targets — a level of algorithmic error that would be operationally significant in any targeting context, and starkly consequential in a densely populated urban environment. Adversarial vulnerabilities — spoofing, sensor deception, and data poisoning — remain comparatively underexamined in public reporting relative to jamming resistance, but are widely regarded by technologists as an emerging risk as autonomous targeting systems proliferate to actors with fewer resources to harden them.

Legal / Diplomatic Dimension

International humanitarian law’s core obligations — distinction between combatants and civilians, proportionality in anticipated collateral harm, and precautions in attack, all rooted in the 1949 Geneva Conventions and 1977 Additional Protocol I, alongside the residual protection of the Martens Clause — were designed around the presumption of human judgment applied to a specific, known situation. Legal scholars have converged on the concept of an “accountability gap” or “responsibility gap” — a term originating with philosopher Robert Sparrow and subsequently developed by legal scholars including Christof Heyns and Human Rights Watch’s Bonnie Docherty — to describe the structural problem that arises when an autonomous system’s unlawful targeting decision cannot be clearly attributed to a specific human’s intent, knowledge, or negligence, potentially leaving no individual legally accountable under existing international criminal law frameworks that require proof of individual mens rea.

The policy response favored by the Stop Killer Robots coalition (more than 270 non-governmental organizations across 70-plus countries), the International Committee of the Red Cross, and a majority of UN member states centers on codifying “meaningful human control” as a binding legal requirement — comprising a general obligation to maintain such control, specific prohibitions on weapons that by their nature cannot be meaningfully controlled or that are designed to target humans directly, and positive obligations ensuring control is maintained for all other systems that select and engage targets. This proposal has shaped both the UN General Assembly’s 166-to-3 resolution and the substantive “rolling text” under negotiation at the CCW, but it has not overcome the CCW’s consensus-based procedural structure, which allows any single objecting state — and the United States, Russia, and others have signaled clear reservations about a categorical ban — to block a binding outcome. Some legal scholars have proposed alternative or complementary approaches, including “embedded legality,” which would require distinction and proportionality principles to be built directly into a system’s design architecture and subjected to auditable “black box”-style recording, rather than relying solely on ex post facto legal review.

Scenario Analysis

Scenario One: Binding Treaty Achieved (2026–2028). The CCW’s Seventh Review Conference in November 2026 or a subsequent UN General Assembly track produces a legally binding instrument prohibiting lethal autonomous weapons systems that operate without meaningful human control and target humans directly, while regulating other autonomous systems through design, testing, and accountability requirements. This scenario would require either a shift in U.S., Russian, Chinese, or Indian positions, or a decision by the treaty’s proponents to pursue negotiations outside the CCW’s consensus framework (as occurred historically with the Ottawa Landmine Convention and the Oslo Cluster Munitions Convention, both negotiated outside existing consensus-bound forums after they stalled). Precedent suggests this pathway is possible but requires sustained political will from a critical mass of committed states willing to proceed without universal participation.

Scenario Two: Fragmented National Governance (most likely, 2026–2030). No binding global treaty emerges. Instead, individual states and blocs — the United States through DoD Directive 3000.09 and NSPM-11, the European Union through emerging AI and defense-procurement frameworks, China through internal PLA doctrine and testing protocols, and others through ad hoc national policy — pursue divergent, non-binding standards. Voluntary instruments such as the U.S.-led Political Declaration on Responsible Military Use of AI persist as normative reference points but carry no enforcement mechanism. Battlefield practice, shaped heavily by the ongoing precedents set in Ukraine and Gaza, becomes the de facto standard-setter in the absence of codified law.

Scenario Three: Accelerated Proliferation and Arms-Race Dynamics. Failure to reach even a fragmented governance equilibrium leads to accelerating autonomy adoption driven by competitive pressure among the United States, China, Russia, and a widening circle of middle powers (including India, Pakistan, Turkey, and Israel) seeking to avoid falling behind. Human control continues to erode incrementally — first in terminal guidance, then in target-selection recommendation systems operating at increasing speed and scale — without any single dramatic breach of an explicit “no fully autonomous lethal targeting” norm, but with the cumulative effect of substantially diminished human judgment in practice. Proliferation to non-state and smaller state actors, facilitated by the falling cost of commercial AI and drone components, increases the risk of use by actors with little regard for international humanitarian law or the technical capacity to constrain algorithmic error.

Scenario Four: A Significant Incident Forces Re-Negotiation. A high-casualty incident clearly attributable to autonomous-system malfunction, misidentification, or algorithmic bias — analogous in political effect to past mass-casualty events that catalyzed arms-control breakthroughs — generates sufficient public and diplomatic pressure to reopen negotiations outside the CCW’s consensus deadlock, potentially compressing the treaty timeline of Scenario One. This scenario is inherently unpredictable in timing but plausible in mechanism, given the reliability and error-rate concerns already documented in fielded systems.

Probability Assessment

Scenario	Likelihood by end-2028	Key Uncertainty
Binding global treaty achieved	Low	Requires either CCW consensus reversal or a coalition willing to negotiate outside it
Fragmented national governance persists	High	Depends on continued absence of a triggering incident or diplomatic breakthrough
Accelerated proliferation / arms-race dynamics	Moderate to High	Contingent on continued great-power competition and battlefield performance incentives
Incident-driven forced renegotiation	Low to Moderate	Inherently unpredictable; probability rises cumulatively as more systems are fielded
A CCW-negotiated non-binding political declaration (intermediate outcome)	Moderate	Most plausible “compromise” outcome if binding treaty talks stall at the Seventh Review Conference

These estimates should be read as directional assessments rather than precise probabilities. They reflect the structural difficulty of consensus-based multilateral arms control, the demonstrated unwillingness of leading military-AI powers to accept binding constraints on current development trajectories, and the absence, to date, of an incident severe enough to force the issue onto emergency diplomatic footing — while acknowledging that such an incident, given the error rates already documented in fielded systems, cannot be ruled out.

Strategic Outlook

Over the next three to five years, the most consequential erosion of human control is likely to occur not through a single dramatic delegation of lethal authority to machines, but through the cumulative, incremental compression of the human role at each stage of the targeting cycle: faster AI-generated target recommendations, broader delegation of terminal-guidance completion, and expanding trust thresholds for supervisory (“human-on-the-loop”) rather than authorizing (“human-in-the-loop”) control. Ukraine will likely remain the most important real-world laboratory for this evolution, given the scale, duration, and public documentation of its drone war; Gaza will likely remain the most consequential case study for the erosion of substantive review beneath the formal preservation of human sign-off; and the Taiwan Strait will likely remain the central scenario against which Chinese and American planners stress-test their respective autonomy investments.

Institutionally, the CCW’s Seventh Review Conference in November 2026 represents a genuine inflection point, though not in the direction its strongest proponents seek. The most probable outcome is neither a binding treaty nor complete collapse of the process, but an intermediate result — a non-binding political declaration, an extended discussion mandate, or a bifurcated track in which a coalition of willing states begins exploring negotiation outside the CCW’s consensus structure, mirroring the precedent set by earlier humanitarian disarmament treaties. Policymakers, military planners, and diplomats should treat the resulting governance vacuum as durable rather than transitional: national policy, battlefield precedent, and industry practice — not international law — will function as the primary constraints on autonomous weapons development through at least the end of this decade.

For military and defense-industrial planners, three considerations warrant particular attention. First, reliability and error-rate transparency will increasingly determine both operational trust and legal exposure, as the gap between promotional claims and documented field performance (illustrated by the reported HX-2 mission-success data from Ukraine) becomes a matter of public and legislative scrutiny. Second, swarm-orchestration software, not individual platform procurement, is emerging as the genuine technological bottleneck and the most consequential area of near-term investment. Third, states operating in contested, nuclearized dyads — most acutely Pakistan and India — face a distinct set of crisis-stability risks as autonomy compresses decision timelines in environments where miscalculation carries exceptional consequences, a dimension that deserves more sustained analytical and diplomatic attention than it has so far received.

Conclusion

Autonomous weapons have moved from theoretical concern to operational fact without a corresponding evolution in the international legal architecture meant to govern the use of force. Human control is not disappearing uniformly; it is being selectively and unevenly withdrawn — retained, in most documented cases, at the point of authorizing an attack against a human target, but increasingly delegated during the terminal phase of that attack’s execution, and increasingly attenuated by the sheer speed and volume at which AI-generated recommendations now flow to human reviewers. The multilateral effort to arrest or channel this trend through a binding treaty is very unlikely to succeed on the UN Secretary-General’s 2026 timeline, leaving national policy, industrial competition, and battlefield precedent as the operative forces shaping how much human judgment survives on the future battlefield. Whether that outcome proves compatible with the foundational principles of international humanitarian law — distinction, proportionality, and accountability — remains, at the close of 2026, an open and urgent question rather than a settled one.

Frequently Asked Questions

What is the difference between an autonomous weapon and a remotely piloted drone? A remotely piloted drone is continuously or intermittently controlled by a human operator who selects and authorizes each target engagement. An autonomous weapon system, once activated, can select and engage a target without further human intervention — though in practice, most fielded systems occupy a middle ground, such as drones that require human target designation but complete the final attack run autonomously once communications are lost.

Has a fully autonomous weapon ever killed a human being without human authorization? The most frequently cited possible case is the March 2020 use of a Turkish-made Kargu-2 loitering munition in Libya, described in a UN Panel of Experts report as having been programmed to attack without an operator data link. However, independent analysts have disputed whether the system was genuinely operating autonomously at the time and whether the attack caused any casualties, so this remains a contested rather than confirmed precedent.

Why has the United Nations been unable to agree on a treaty banning autonomous weapons? Negotiations under the Convention on Certain Conventional Weapons require consensus among all participating states, and several major military powers — including the United States, Russia, and others — have declined to support a categorical ban, preferring voluntary guidelines that preserve development flexibility. A large majority of UN member states support binding restrictions, but consensus-based procedure allows a small number of objecting states to prevent a binding outcome.

What does “meaningful human control” mean in this debate? It is a legal and policy concept, promoted chiefly by the Stop Killer Robots campaign and allied legal scholars, holding that humans must retain sufficient information, understanding, and time to assess the legality and morality of a specific use of force before it occurs. Proponents argue it should be a binding requirement for all weapons systems that select and engage targets; opponents argue existing human-judgment standards in national policy already achieve a comparable effect without a new treaty.

How does the debate over autonomous weapons differ between offensive strike drones and AI-assisted targeting systems like those used in Gaza? Strike-drone autonomy, as seen in Ukraine, typically concerns whether a weapon can complete an already-designated attack without a live communications link — a narrower and less legally contested form of delegation. AI-assisted targeting systems like Israel’s Gospel and Lavender concern a different problem: whether the speed and volume of machine-generated target recommendations leave human reviewers with enough time and information to exercise genuine judgment before authorizing an attack, even though a human formally remains “in the loop.”

Sources

UN Office for Disarmament Affairs, “Convention on Certain Conventional Weapons – Group of Governmental Experts on Lethal Autonomous Weapons Systems (2026),” meetings.unoda.org
Digital Watch Observatory, “GGE on Lethal Autonomous Weapons Systems,” dig.watch/processes/gge-laws
Human Rights Watch, “UN: Start Talks on Treaty to Ban ‘Killer Robots'” (2025) and “Killer Robots: New UN Report Urges Treaty by 2026” (2024), hrw.org
Stop Killer Robots, “The Story So Far” and related campaign publications, stopkillerrobots.org
Harvard Law School International Human Rights Clinic, “Elements of a New Treaty on Fully Autonomous Weapons,” humanrightsclinic.law.harvard.edu
CSIS, “Defining Autonomy: Why Software, Not Drones, Will Decide the Next War” (2026), csis.org
The White House, “National Security Presidential Memorandum/NSPM-11” (June 5, 2026), whitehouse.gov
Wikipedia, “Department of Defense Directive 3000.09” and “Executive Order 14347,” en.wikipedia.org
Congressional Research Service, “DOD Replicator Initiative: Background and Issues for Congress,” congress.gov
Defense One, “The Pentagon’s $54 Billion Bet on Autonomous Warfare” (2026), defenseone.com
Forbes, “Inside Ukraine’s Race to Build Autonomous Strike Capabilities” (2026) and “Ukraine’s Killer AI Drones Are Back With a Vengeance” (2026), forbes.com
Bloomberg, “Ukraine’s AI Drones Have Given Kyiv a Fresh Edge on Battlefield” (2026), bloomberg.com
RSIS, “Intermediate-Range Strike Drones: Lessons from the War in Ukraine” (2026), rsis.edu.sg
+972 Magazine, “‘Lavender’: The AI Machine Directing Israel’s Bombing Spree in Gaza,” 972mag.com
Lieber Institute, West Point, “The Gospel, Lavender, and the Law of Armed Conflict” and “Legal Accountability for AI-Driven Autonomous Weapons,” lieber.westpoint.edu
The Diplomat, “China’s Military Is Now Using AI to Plan Strike Operations” (2026) and “Machines in the Alleyways: China’s Bet on Autonomous Urban Warfare” (2026), thediplomat.com
Foundation for Defense of Democracies, “China’s War Wolves: From Commercial Tech to Combat Power” (2026), fdd.org
CNBC, “Why Europe Is Suddenly Betting Big on Drones” (2026), cnbc.com
MIT Technology Review, “The Future of Autonomous Warfare Is Unfolding in Europe” (2026), technologyreview.com
Al Jazeera, “EU Discusses ‘Drone Wall’ to Protect Airspace from Russian Violations” (2025), aljazeera.com
NPR, “A UN Report Suggests Libya Saw the First Battlefield Killing by an Autonomous Drone” (2021), npr.org
Brookings Institution, “Applying Arms-Control Frameworks to Autonomous Weapons” (2021), brookings.edu
Quwa Defence News, “Pakistan Takes the Plunge in Military Drones and Autonomy Stacks” (2026), quwa.org
Irregular Warfare Initiative, “The Strategic Use of Drones in Pakistan-India Irregular Warfare” (2026), irregularwarfare.org
Defence Research and Studies, “AI and Autonomous Systems: Implications for India’s Defence Strategy” (2026), dras.in
CESCUBE, “Lethal Autonomous Weapons Systems and the Governance Deficit” (2026), cescube.com

This assessment reflects open-source information available as of early September 2026. Given the pace of technological and diplomatic developments in this domain — including the pending outcome of the CCW’s Seventh Review Conference in November 2026 and the finalized text of the updated DoD Directive 3000.09 — readers should treat probability assessments as time-bound and subject to revision.