In June 2026, New Scientist published an article that sparked widespread discussion: "Robots are about to overtake armed soldiers as the deciders of war." Behind this claim lies a technological trend rapidly accelerating but not yet fully understood by the public — autonomous weapons systems are evolving from "human remote control" to "human supervision," and in some domains have already approached the "human out of the loop" stage. Observatory believes this is not sci-fi apocalyptic prophecy, but an automation system technology inflection point that requires calm analysis.
Kill Chain Automation: Understanding the Evolution Through Four Stages
The military kill chain typically includes four stages: reconnaissance and target identification, target verification and authorization, strike execution, and battle damage assessment. The degree of automation at each stage varies considerably, as does the pace of progress.
Stage 1: Autonomous Reconnaissance and Target Identification (Achieved)
This was the earliest link to be automated. UAVs, UGVs, and USVs can conduct long-endurance patrol missions without human control, using onboard AI systems to automatically identify potential military targets. The Pentagon's Project Maven can automatically classify and tag targets from satellite and drone imagery with accuracy exceeding 94% in 2026. Israel's Harop "suicide drone" can autonomously patrol designated areas, identify and attack radar signal sources — with human only performing final confirmation before attack.
Stage 2: Target Verification and Authorization — Human-in-the-Loop (Transitioning)
This is currently the most controversial link and the core threshold for autonomy. Traditionally, target authorization must be made by a human commander — a requirement written into military rules of engagement and international humanitarian law. However, technological reality is changing this landscape:
In 2026, the US Department of Defense's Collaborative Combat Aircraft (CCA) program tested systems where AI can autonomously approve strikes against confirmed threats based on pre-set rules of engagement when communications are unavailable (due to jamming by adversaries) — effectively removing humans from the real-time decision loop while retaining only "mission-level" pre-authorization. More notably, combat experience in Ukraine has demonstrated that human-machine communication is highly unreliable in electronic warfare environments — human supervisors may be completely disconnected from the loop at critical moments.
Stage 3: Autonomous Strike — The UGV Tipping Point
The core case in the New Scientist article involves autonomous unmanned ground vehicles tested for front-line defense. These armored vehicles can perform patrol, sentry, and defensive engagement missions without any human operators. Unlike aerial drones, ground vehicles face more complex environments — urban terrain, civilian identification, friendly force distinction — placing higher demands on AI perception systems.
The US Army's Next Generation Combat Vehicle program includes optionally manned/unmanned armored platforms. In unmanned mode, vehicles can autonomously move within pre-set operational zones, classify approaching threats, and take defensive actions based on pre-set rules. In 2026 field exercises, these systems demonstrated reaction times superior to humans under visibility-limited conditions (smoke, night, urban).
Stage 4: Battle Damage Assessment — AI Closes the Loop (Near Achievement)
Traditionally reliant on human confirmation or subsequent reconnaissance, battle damage assessment can now be performed by AI systems within seconds of a strike by comparing pre-and post-strike sensor data, automatically evaluating strike effectiveness and determining whether a secondary strike is needed. This enables the kill chain to close within milliseconds — far exceeding human reaction speed.
The $13 Billion Pentagon AI Budget: Structural Implications
In 2026, the US Department of Defense's AI-related budget reached $13 billion — approximately four times the 2020 level. The allocation of these funds reveals the military's true priorities for autonomous systems:
| Program | Budget ($B) | Focus Area |
|---|---|---|
| Collaborative Combat Aircraft (CCA) | 2.8 | AI-driven unmanned combat aircraft swarms |
| Project Maven Upgrade | 1.5 | AI for all-domain target identification and tracking |
| Replicator Initiative | 1.2 | Mass deployment of low-cost autonomous systems |
| CYBERCOM AI Defense | 1.0 | AI-driven cybersecurity and electronic warfare |
| Next-Gen Combat Vehicle Autonomy | 0.8 | Unmanned ground combat platforms |
| Other Distributed Programs | 5.7 | Service-level AI integration and basic research |
The $13 billion figure is impressive in itself, but Observatory believes what truly matters is not the amount, but the leverage effect on human resource substitution. An F-35 fighter's full lifecycle cost exceeds $400 million — training a qualified pilot takes 5-8 years — while a CCA's unit cost target is $10-20 million, with production capacity not constrained by personnel training cycles. This means the Pentagon's AI investment is not a "nice-to-have" technology upgrade, but a structural transformation in combat cost accounting.
Technical Foundation: Perception, Decision, and Swarm Coordination
The technical maturity of autonomous weapons systems rests on the simultaneous advancement of three pillars:
Multi-modal perception fusion. Modern autonomous platforms typically carry EO/IR cameras, millimeter-wave radar, LiDAR, and acoustic sensors. AI systems fuse information from these different sensor modalities into a unified environmental representation, significantly improving target identification under adverse weather, night, and smoke conditions. By 2026, multi-modal fusion perception systems achieved a target false positive rate of 1.2 per thousand hours in military testing — approaching human operator levels.
Real-time path planning and obstacle avoidance. Autonomous navigation in military environments is far more complex than civilian scenarios — maps may not exist or be outdated, GPS may be jammed, roads may be destroyed. Modern autonomous systems use deep reinforcement learning trained on millions of hours in simulated environments, enabling them to autonomously select feasible paths in completely unfamiliar terrain. By 2026, US military-tested autonomous ground vehicles achieved 92% navigation success in GPS-denied environments (99% with GPS).
Swarm intelligence and task allocation. A single autonomous platform has limited combat capability, but the coordinated effect of dozens of platforms grows exponentially. The Replicator program's core hypothesis: large numbers of low-cost autonomous systems, when coordinated, can match or surpass a few expensive manned platforms. AI swarm control algorithms can automatically assign tasks to each platform — who handles reconnaissance, who jams, who strikes — dynamically adjusting based on real-time battlefield conditions. In 2026 tests, an AI swarm of 50 small drones successfully "defeated" a mixed formation of 4 manned fighters and 8 conventional drones in simulated exercises.
The Technical Dilemma of International Arms Control Frameworks
The development of autonomous weapons systems faces a fundamental regulatory dilemma: technology advances far faster than international legal frameworks can be updated.
Since 2013, the UN Convention on Certain Conventional Weapons (CCW) has discussed regulating Lethal Autonomous Weapons Systems (LAWS). After 13 years, negotiations have failed to reach consensus on legally binding provisions. The core disagreement centers on the definition of "Meaningful Human Control" — no nation can agree on the specific boundaries of this concept.
A thorny technical issue: when autonomous systems make decisions based on deep neural networks, their decision-making processes are inherently unexplainable. Even if human commanders retain a "veto" power, if the AI system cannot explain why it classified a target as a "threat," human "supervision" becomes a ceremonial confirmation. The US Department of Defense's 2023 directive on autonomous weapons systems explicitly requires "appropriate human supervision" but does not define when human supervision is "appropriate" — this ambiguity leaves a vast gray area for autonomy.
Observatory Analysis
Observatory believes the "tipping point" for autonomous weapons systems is not the result of a single technological breakthrough, but the convergence of multiple technology curves. The 2025-2027 window is critical because five conditions are simultaneously maturing:
First, AI perception accuracy has crossed the operational threshold of human-level performance (under controlled conditions). When a system's target identification accuracy exceeds 90% with lower false positive rates than human operators, the technical justification for keeping humans in the decision loop begins to be questioned.
Second, the prevalence of counter-communications technology (electronic warfare, GPS jamming) has sharply reduced the reliability of remote human supervision. If human supervisors cannot receive sufficient information to make informed decisions more than 30% of the time, then "human-in-the-loop" transforms from safeguard to risk.
Third, the cost curve of autonomous systems has dropped sharply, making large-scale deployment a realistic option. A thousand disposable drones cost far less than a single manned fighter — when quantity becomes a substitute for quality, the threshold for conflict is lowered, not raised.
Fourth, competitive asymmetry: if one side deploys autonomous weapons and the other does not, the former's combat effectiveness will overwhelmingly surpass the latter. This asymmetric pressure may force nations to accelerate autonomy without adequate preparation — a classic "security dilemma" manifested in the autonomous weapons domain.
Fifth, advances in generative AI have made "human-level" tactical decision-making possible. The combination of large language models and reinforcement learning allows autonomous systems not only to execute pre-programmed tactical actions but to generate novel, untrained tactical plans — in 2026 war games, certain AI-generated tactical plans even exceeded human military planners' expectations.
Forward Outlook
At the technical level, 2027-2028 milestones worth watching include: whether the US Air Force will formally approve autonomous CCA firing authority under communications-denied conditions, whether autonomous ground vehicles will conduct their first autonomous engagement in actual combat (rather than exercises), and whether the UN CCW negotiations can reach any consensus on a binding LAWS framework.
Over a longer horizon, the true impact of autonomous weapons systems may not lie in changing battlefield efficiency, but in altering the psychological threshold for conflict decisions. When the cost of launching a strike — whether economic, political, or human — is dramatically reduced by autonomous systems, the psychological threshold for using force is correspondingly lowered. This consequence deserves more careful consideration than any technological breakthrough.
Disclaimer: The information in this article is provided for reference only and does not constitute investment advice or business decision guidance. Data and time-sensitive information are current as of the publication date and may change with subsequent developments. Neither the author nor POC.HK assumes any liability for losses resulting from the use of this information.