Here is the full study: A language-based analysis of China’s vision of AI warfare (pdf)
...
The study, Decoding the Dragon: A Language-Based Analysis of China’s Vision of AI Warfare, came out recently by NATO Defense College. The authors are prominent experts; Anselm Küsters is scientific director for the Nordic countries and head of the Department of Digitalisation and AI at the Centre for European Policy; Marius Strubenhoff works for the same think-tank as an independent expert on foreign and security policy.
Mr Küsters and Mr Strubenhoff applied large language models to a corpus of over 171,000 Chinese documents spanning 2000 to 2025. The result is a rare empirical window into Beijing’s military AI ambitions. It also carries direct implications for EU security policy — and, by extension, for the bloc’s legislation.
The study’s most consequential finding concerns priorities. “Near-term PLA [China’s People’s Liberation Army] investments appear geared towards achieving a battlespace transparency that supports both information dominance and precision strikes,” the authors write. The goal is to see everything, and to use that knowledge to hit targets with pinpoint accuracy before the other side can react.
...
Surveillance first, autonomy later
Surveillance-related terms dominated the corpus by a wide margin. Computer vision accounted for 49.4 per cent of the subfield distribution, followed by robotics at 12.5 per cent and natural language processing at 6.2 per cent. AI safety, by contrast, accounted for under two per cent.
This is more than a defensive posture. Under China’s dissipative warfare doctrine, surveillance is an offensive enabler. As the study explains, “the doctrine involves maintaining order within one’s own system while deliberately creating disorder within an adversary’s political, economic and military networks”.
This pattern suggests that ‘meaningful human control’ functioned primarily as diplomatic vocabulary during peak international scrutiny but has been progressively de-emphasised.
...
The practical implication is clear. “The dominance of surveillance applications suggests that near-term crisis instability will stem from information asymmetries rather than from rogue autonomous weapons,” the authors conclude. Europe’s immediate exposure is not a swarm of killer robots. It is the gradual, AI-accelerated erosion of its situational awareness, and the compression of its decision time.
Translated into human English: in a crisis, the real danger is not that a robot weapon will go rogue and start firing on its own. It is that China will simply know far more about what is happening on the battlefield than its adversary does. It may then use that advantage to act faster and strike first.
...
Europe’s response to this trajectory runs through a legal framework built largely for civilian purposes. Brussels’ flagship law on the subject, the EU AI Act, explicitly excludes AI systems used exclusively for military, defence, or national-security purposes from its scope. That carve-out leaves a governance vacuum precisely where the NATO study sees China moving fastest.
The Dual-Use Export Control Regulation provides some leverage over outward flows of sensitive AI chips and surveillance software, including a catch-all licence requirement for cyber-surveillance items linked to human-rights risks. But it is export-centric and reactive. It does not yet cover foundation models, reinforcement-learning environments, or the synthetic datasets on which PLA intelligentised warfare depends.
...
[EU'] GDPR and the Law Enforcement Directive govern biometric and personal data processing. They are strong on individual rights. They are not designed to counter adversarial exploitation of European data streams by a foreign military. The European Parliament has passed resolutions in 2018 and 2021 urging a ban on lethal autonomous weapons that lack meaningful human control. Those resolutions remain politically significant as well as legally non-binding, i.e., next to useless.
...
The study identifies three investment priorities. “NATO members should invest in adversarial robustness research to develop AI systems that are resilient to data poisoning, model extraction and attacks that can blind, misdirect or compromise military decision-making systems,” the authors write.
They also call for investment in counter-ISR technologies: adaptive camouflage, decoy generation, and emissions management. It requires human-machine teaming architectures; that could demonstrate how “Human-in-the-loop systems can match or surpass autonomous systems in decision speed and accuracy, countering the narrative that human control requires accepting operational inferiority.” Meaning, if we have lost you along the way, that people still matter.
...
Most ambitiously, a European strategic autonomy framework (a new treaty instrument establishing common operational command and a joint defence AI budget) would require unanimous member-state agreement. We know how well that usually goes.
Messrs Küsters and Strubenhoff are careful not to overstate China’s current capabilities. China, they note, is not quite on the verge of autonomous military dominance. But as the dissipative warfare doctrine makes plain, the trajectory is clear: it wants to get there fast. And Europe’s legislative gap is real; and its home front is already part of the battlefield.
Archived