The Doctrine That Made Human Judgment a Bottleneck
[Navigation links not applicable for narrative series]
In two thousand and twenty, the People's Liberation Army published a doctrine that rewrote the theory of how wars would be fought. Not a weapons programme. Not a procurement plan. A doctrine. The Academy of Military Science issued what it called the Science of Military Strategy, and its central argument was this: human judgment in the targeting loop was no longer a safeguard. It was a bottleneck.
A senior faculty member at China's National Defence University described the shift in terms that stripped away every euphemism. The operational factors, he wrote in journals cited by the Jamestown Foundation in two thousand and nineteen, were changing from information in the lead to machines leading in battle. Human judgment, in this framing, was a competitive vulnerability. The framing did not appear in a classified briefing. It appeared in academic journals, published openly, as a strategic signal, before most of the hardware existed to execute it. China announced its theory of autonomous war before it built the weapons to fight one.
That sequence matters. The doctrine was not waiting for the technology. The technology was being built to serve the doctrine.
Military-Civil Fusion and the architecture of obligation
In two thousand and seventeen, Xi Jinping personally chaired the Military-Civil Fusion Development Commission, the body whose mandate was to dissolve the boundary between civilian and military AI development so completely that the distinction lost operational meaning. Military-Civil Fusion was not a partnership programme or a government contractor scheme. It was structural law. Every Chinese AI company operating at scale became a participant in military AI by legal architecture rather than by contract.
SenseTime and iFlyTek, both listed on the United States Entity List for documented surveillance links, maintained connections to People's Liberation Army programmes. The state-owned defence electronics conglomerate CETC demonstrated coordinated autonomous drone swarms at public airshows. DJI, controlling roughly ninety per cent of the global consumer drone market by component share, manufactured platforms that appeared with regularity in conflict theatres from Ukraine to Sudan.
When DeepSeek R1 arrived in January two thousand and twenty-five, matching or exceeding frontier American model benchmarks under chip-constrained conditions, the Nvidia share price lost roughly six hundred billion dollars in a single trading day as investors recalculated the assumptions underlying American AI dominance. The founder of DeepSeek's parent company was direct: money is never the problem; chip bans are the problem. By December two thousand and twenty-five, a Pentagon assessment found that Chinese large language models had narrowed the performance gap with American frontier models to a material degree.
There was no walkout moment because the legal architecture made collective refusal structurally incoherent. Every Chinese AI company was already in the kill chain by design.
The speed problem
This is the episode that drives everything else in the autonomous weapons competition, and it is not primarily about capability. It is about tempo.
Wargaming exercises conducted by CNAS, RAND, and CSIS over the past decade have produced a consistent finding. In Taiwan scenarios where the People's Liberation Army employs autonomous systems and the United States does not match that autonomy, the PLA's speed advantage is often decisive. The finding is not that autonomous weapons are inherently superior. It is that human decision-making, operating through command chains designed for a slower tempo of conflict, cannot keep pace with an adversary whose systems do not require human authorisation at the tactical level.
The logic that follows is the most dangerous structural pressure in modern military competition. If the PLA's autonomous systems can identify and engage targets faster than a human-authorised response can reach the same targets, the operational incentive to remove human authorisation from your own systems becomes nearly irresistible. If China automates, the United States must automate to compete. If the United States automates, Taiwan must automate to survive. Each step removes another human from the decision chain. Each removal eliminates another point at which judgment, restraint, or legal accountability might be exercised.
This is not a hypothetical cascade. It is the documented finding of three major Western policy research institutions, applied to the most consequential military flashpoint on earth.
The August 2022 exercises
On four August two thousand and twenty-two, the People's Liberation Army Eastern Theatre Command launched the largest military exercises around Taiwan since the Third Taiwan Strait Crisis of nineteen ninety-six.
Eleven Dongfeng ballistic missiles arced over the island. Seven exclusion zones overlapping Taiwan's busiest shipping lanes sealed the strait. Forty-seven fighter aircraft crossed the median line in a single day, more than the total for the preceding two decades combined. Four PLA drones flew above the Kinmen Islands; Taiwanese forces fired warning flares to drive them away.
The exercises were nominally a response to a congressional visit. Their structure communicated something more precise than diplomatic displeasure. The zones demonstrated the capacity to blockade the island entirely. The missiles demonstrated the capacity to strike it. The drones, small and slow above Kinmen, demonstrated that the next confrontation would not be conducted by manned aircraft alone.
In December two thousand and twenty-five, during exercises codenamed Justice Mission twenty twenty-five, the Eastern Theatre Command deployed drones that filmed the island's tallest building from overhead and released the footage on Chinese social media. A piece of psychological theatre that also served as a proof of concept: unmanned systems could reach the capital. The exercises had become rehearsals. The rehearsals increasingly featured machines.
The hardware positioned
The capability being built for this scenario is not experimental. It is fielded.
Satellite imagery analysed by the Mitchell Institute for Aerospace Studies identified more than two hundred converted supersonic fighters, designated J-6W, stationed at six air bases in Fujian and Guangdong provinces facing the strait. These platforms, derivatives of a Soviet-era design, maintain speeds exceeding Mach one, a combat radius of several hundred kilometres, and a payload capacity of up to five hundred kilograms. The PLA's strategy is attrition through mass: launching these platforms in barrages of hundreds to overwhelm air defence batteries, forcing defenders to expend million-dollar interceptors against five-figure targets. The cost-exchange ratio is the weapon, not the platform.
The Type 076 amphibious assault ship, launched in December two thousand and twenty-four, is designed to project autonomous air power from its flight deck. The Jiutian drone carrier, which completed its maiden flight in December two thousand and twenty-five, carries a six-tonne payload across seven thousand kilometres, sufficient to launch and recover combat drones from well beyond any island's defensive perimeter.
The architecture being built is not a contingency plan. It is a fielded capability. For boards and technology leaders with supply chain or operational exposure in this theatre, the question is not whether this scenario might occur. It is whether your governance framework was designed for the world being built, or the world that existed when it was written.
The governance contradiction in Geneva
China participated actively in the Convention on Certain Conventional Weapons Group of Governmental Experts on lethal autonomous weapons systems. Sent senior diplomats. Submitted working papers. Made constructive-sounding statements from the floor. In March two thousand and twenty-four, a Chinese ambassador told delegates that China had always stood for the negotiation of a legally binding instrument and was ready to engage in further discussions.
Analysts who examined the working papers found a different architecture. China proposed five cumulative technical characteristics for weapons it would consider unacceptable: the system must be lethal, fully autonomous, impossible to terminate, indiscriminate, and capable of uncontrolled evolution. All five simultaneously. The Lieber Institute at West Point, in a two thousand and twenty-six analysis, concluded that China had drawn a remarkably narrow line that effectively excluded a wide range of emerging autonomous capabilities, many of which Beijing was actively developing. The Jiutian drone carrier would not meet all five criteria. The Wing Loong II operating under automated rules of engagement would not. The CETC swarms demonstrated publicly at the Guangzhou airshow would not.
The threshold was designed to exclude the systems China was actually building.
Every statement China made in Geneva was technically consistent with its military AI programme. Human control, in the People's Liberation Army doctrine, meant Party control: the Central Military Commission authorised campaigns. Within those authorised parameters, machines could lead in battle without further human review. The CCW statements never contradicted the Science of Military Strategy. They were written so as not to.
The nuclear intersection and the Petrov problem
In two thousand and twenty, the Stockholm International Peace Research Institute concluded that the interaction between autonomous conventional systems and nuclear forces created a new set of escalation risks that current arms control frameworks were not designed to address. Six years later, no framework has been designed to address them.
In November two thousand and twenty-four, at an APEC summit, two heads of state agreed in principle that artificial intelligence should not autonomously initiate nuclear launch decisions. The agreement contained no verification mechanism. It offered no definition of autonomous. It included no enforcement provision.
The Taiwan scenario is where the four dimensions of the accountability gap converge simultaneously. Capability advancing on all sides; no accountability framework to assign it; governance producing declarations without mechanisms; and the question of whether democratic publics have authorised the autonomous systems being built in their name, never meaningfully raised in any legislature this series has examined.
In September nineteen eighty-three, a Soviet officer named Stanislav Petrov correctly identified a satellite early warning of five incoming nuclear missiles as a false alarm generated by sunlight reflecting off clouds. He had minutes to decide. He decided correctly because he applied contextual reasoning the automated system could not, and because he chose caution over the data. He died in two thousand and seventeen.
Now consider what replaces him. AI systems integrated into nuclear command and control architectures can process sensor data and generate response recommendations in seconds. In a Taiwan crisis, hundreds of autonomous drones would be operating simultaneously across the strait, launched from platforms that also carry nuclear-capable missiles. The distinction between a drone swarm and a nuclear first strike may depend on sensor readings interpreted by algorithms that cannot explain their reasoning, processed at speeds that leave no time for the judgment Stanislav Petrov exercised.
China's accountability gap is not a gap. It is the doctrine. And the doctrine, exported at Wing Loong prices to seventeen countries and counting, is the world's emerging baseline.
The Responsible AI in the Military Domain process, which produced declarations from forty-seven signatory states following the 2023 summit in The Hague and the 2024 Seoul follow-on conference, represents the most substantive multilateral effort to define norms for military AI outside the Convention on Certain Conventional Weapons. REAIM is not a treaty. It is a political commitment. Its declarations affirm that states should maintain human control over nuclear weapons decisions, require meaningful human oversight for lethal force, and commit to responsible development and use principles. Neither China nor Russia are signatories. The J-6W programme, the Jiutian carrier, and the Science of Military Strategy two thousand and twenty were built outside any framework REAIM can reach. For boards and procurement teams sourcing technology with dual-use potential, the governance gap is not a distant policy question. REAIM's forty-seven signatories govern the buyers and the component suppliers. The non-signatories govern the doctrine.
What does accountability look like when the doctrine is the gap?
The views expressed in this article are entirely my own, informed by more than 30 years of professional experience in architecture, security, and technology leadership in New Zealand. They do not represent the views of my employer, any government agency, or the New Zealand government. My commentary on legislation and policy is analytical, drawing on publicly available sources and my professional expertise in architecture, security, and AI governance. I follow the Public Service Commissioner's Code of Conduct for the Public Sector and social media guidance.
Andreas Hamberger is a Wellington-based enterprise architect and governance practitioner with more than 30 years of experience across New Zealand's public and private sectors. He holds TOGAF, IAPP, and AMInstD credentials, is an Associate Member of the Institute of Directors New Zealand, and is the founder of Te Pono Limited. Lethal by Design: Military AI and the Crisis of Human Control examines the accountability gap in autonomous and AI-assisted weapons, and the choices that built it.
I use AI tools, including Sudowrite, Claude, Perplexity AI, DeepSeek AI, ChatGPT, Grok, Copilot, Openart and Gemini, as deliberate production tools, not ghostwriters. This is consistent with my position: AI amplifies human judgement; it does not replace it. The frameworks, arguments, and editorial decisions in this series are original work. AI accelerated the process. The thinking is mine.
[1] Lethal by Design: Military AI and the Crisis of Human Control, Chapter 18. Te Pono Limited, 2026.
[2] Lethal by Design: Military AI and the Crisis of Human Control, Chapter 19. Te Pono Limited, 2026.
[3] Jamestown Foundation. "The Algorithm Game: China's Use of AI in Military Decision-Making." April 2019. https://jamestown.org
[4] Boulanin, Vincent et al. Stockholm International Peace Research Institute. "Artificial Intelligence, Strategic Stability and Nuclear Risk." June 2020. https://sipri.org
[5] Mitchell Institute for Aerospace Studies. "China Airpower Tracker: J-6W Conversion Programme." February-March 2026.
[6] Center for a New American Security. "Hellscape: How Ukraine's Drone Tactics Translate to Taiwan." Pettyjohn and Campbell. February 2026.
[7] Lieber Institute, West Point. "China's CCW Proposal and the Scope of Permitted Autonomous Weapons." 2026. https://lieber.westpoint.edu
[8] Center for Strategic and International Studies. "ChinaPower: Taiwan Strait Military Exercises August 2022." 2022. https://chinapower.csis.org
[9] Arms Control Association. "Biden-Xi APEC Agreement on AI and Nuclear Weapons." December 2024. https://armscontrol.org
[10] REAIM (Responsible AI in the Military Domain). "Seoul Call for Action." September 2024. https://reaim2024.kr
[11] Allen, Gregory C. Center for Strategic and International Studies. "China's Military AI Doctrine." Testimony, February 2026. https://csis.org
[12] The War Zone. "Jiutian: China's Long-Range Combat Drone Carrier Completed Maiden Flight." December 2025. https://twz.com