Chinese military researchers tap US AI models to train defense systems

A comprehensive review of over 80 Chinese academic papers and patent filings has revealed that military and security-linked institutions in China are systematically utilizing outputs from premier U.S. artificial intelligence models to bolster domestic defense systems. The findings, first detailed by Reuters and supported by analysis from the Washington-based Jamestown Foundation, indicate that researchers associated with the People’s Liberation Army (PLA) have successfully leveraged models developed by OpenAI and Anthropic. This practice, known as "model distillation," allows Chinese scientists to extract the sophisticated reasoning capabilities of American AI to train smaller, specialized systems tailored for tactical military applications.

The revelation comes at a critical juncture in the technological rivalry between Washington and Beijing. Despite a series of escalating export controls intended to starve China of the high-end semiconductors and computing power necessary to build "frontier" AI models, Chinese defense researchers appear to have found a potent workaround. By using the outputs of existing U.S. models as a "teacher," they are developing "student" models that can run on less powerful, domestically available hardware while retaining many of the advanced decision-making characteristics of their Western counterparts.

The Mechanics of Model Distillation as a Strategic Shortcut

Model distillation is a widely recognized technique in the global AI industry, typically used to create efficient versions of large language models (LLMs) for use on smartphones or edge devices. In a military context, however, the implications are more profound. The process involves prompting a large, resource-intensive model—such as OpenAI’s GPT-4 or Anthropic’s Claude—with complex queries and using its high-quality responses to train a much smaller neural network.

For the Chinese military, this technique solves two primary challenges. First, it bypasses the need for the massive server farms and thousands of NVIDIA H100 GPUs that are currently subject to U.S. export restrictions. Second, it allows for the deployment of AI in "disconnected" environments, such as on a drone or a submarine, where constant access to a cloud-based supercomputer is impossible.

Sunny Cheung, a fellow at the Jamestown Foundation who analyzed dozens of these papers, noted that Chinese military scientists are not just looking for facts; they are capturing "reasoning steps." According to Cheung, teaching an AI model to arrive at the correct answer through logical deduction is significantly more difficult than simple data memorization. By capturing the reasoning chains of U.S. models, the PLA is effectively "downloading" years of American research and development at a fraction of the cost.

Chronology of the US-China AI Tech Conflict

The current situation is the result of a multi-year escalation in technological competition. To understand the context of these recent findings, a timeline of the strategic environment is essential:

• October 2022: The U.S. Department of Commerce implements sweeping export controls on advanced computing chips and semiconductor manufacturing equipment to China, specifically targeting AI development.
• November 2023: Following the global explosion of generative AI, the U.S. tightens these rules, closing loopholes that allowed Chinese firms to access restricted chips through overseas subsidiaries.
• May 2024: U.S. and Chinese officials meet in Geneva for the first high-level intergovernmental dialogue on AI risk and safety. The U.S. expresses concerns about the "misuse" of AI by China.
• Early 2026: Reports emerge that despite hardware sanctions, Chinese military-linked AI performance in tactical simulations has significantly improved, leading to investigations into software-based shortcuts.
• July 2026: The Reuters and Jamestown Foundation review confirms that model distillation using U.S. proprietary models is a standard operating procedure for PLA-linked research units.

Case Studies: Tactical Applications and Cyber Warfare

The academic papers reviewed provide specific examples of how these distilled models are being integrated into the PLA’s operational framework. One of the most striking instances involves PLA Unit 96941, a secretive Beijing-based group focused on military intelligence and cyber operations.

Researchers within this unit published findings on using OpenAI’s GPT-3.5 to process and summarize sensitive military source code. Recognizing that they could not upload classified data to a U.S.-hosted cloud service without risking a security breach, the researchers used the U.S. model to generate high-level logic summaries of non-classified but structurally similar code. They then used these summaries to train a domestic model that could operate entirely within an air-gapped military network, capable of identifying vulnerabilities in enemy software.

In another instance, the North University of China—an institution with deep ties to the national defense industry—utilized Anthropic’s Claude 3 Haiku model. The goal was to generate synthetic training data for a text classification system designed for social media monitoring. This system is intended to assist in content moderation and the identification of "hostile" information patterns, demonstrating how AI is being used to automate domestic security and psychological operations.

The National University of Defense Technology (NUDT) has also focused on "lightweighting" AI for physical combat. Their research describes shrinking image-processing models so they can be embedded in unmanned aerial vehicles (UAVs). This allows a drone to perform real-time target recognition and autonomous navigation even when its communication link to a base is jammed or severed, a capability critical for modern electronic warfare.

Official Responses and Geopolitical Friction

The exposure of these practices has sparked a heated debate between Washington and Beijing regarding intellectual property and international "hegemonism."

The U.S. government and the private companies involved have expressed varying degrees of concern. Anthropic stated that it does not provide commercial access to its models in China and maintains rigorous monitoring systems to detect policy violations. However, the company acknowledged a fundamental technical vulnerability: once a model’s outputs are public or accessible via an API, the company has little control over how those outputs are used to train secondary systems.

OpenAI has previously stated that it takes steps to prevent its technology from being used for military purposes, but the "distillation" method remains a difficult loophole to close. U.S. officials have argued that this practice potentially undermines the spirit of export controls and infringes on the intellectual property rights of American innovators.

Conversely, China’s Ministry of Foreign Affairs has dismissed these concerns, accusing the United States of attempting to maintain a technological monopoly. Beijing argues that its researchers are engaging in legitimate scientific inquiry and that American firms have historically benefited from open-source contributions and global data. Furthermore, Chinese AI startups, such as Moonshot, have moved to distance themselves from claims of foreign dependence, asserting that their latest models are the result of proprietary innovations rather than "theft" or distillation.

Broader Impact and Strategic Implications

The systemic use of U.S. AI by the Chinese military carries several long-term implications for global security and the balance of power in the Indo-Pacific region.

1. Erosion of the "Sanctions Gap": The primary goal of U.S. chip sanctions was to create a multi-year lead for the United States in AI capabilities. If China can effectively "distill" the intelligence of frontier models into smaller systems that run on older hardware, the strategic advantage provided by hardware restrictions is significantly diminished.

2. Autonomous Systems Proliferation: The focus on "model lightweighting" suggests that China is preparing for large-scale autonomous warfare. By optimizing AI to run on satellites, ships, and drones, the PLA is moving toward a "swarm" doctrine where large numbers of inexpensive, intelligent machines can overwhelm more sophisticated but fewer American assets.

3. The Security Risks of Distillation: Interestingly, Chinese researchers are also beginning to view distillation as a potential security threat to themselves. Some papers published by the Army Engineering University discuss "data-free distillation" as a method for Western intelligence to reverse-engineer Chinese models. This suggests a burgeoning "cat-and-mouse" game in the realm of AI security, where both sides are attempting to mask the logic of their systems while trying to extract the logic of their rivals.

4. The Limits of Transferable Intelligence: Despite the successes of distillation, technical experts caution that it is not a "silver bullet." Distilled models are inherently derivative; they can mimic the reasoning of the teacher but rarely exceed it. Trevor Koverko, co-founder of the AI data firm Sapien, notes that while distillation allows for cheaper, locally controlled systems, it does not represent true independence. China remains in a position where its most advanced military AI logic is effectively being "taught" by American algorithms.

Conclusion: The New Front in Intelligentized Warfare

The findings from the Reuters and Jamestown Foundation review underscore a fundamental shift in the nature of military competition. We have moved beyond the era of simply comparing the number of tanks or aircraft; the new frontier is "intelligentized warfare," a term frequently used in PLA doctrine to describe the integration of AI into every level of combat.

As the U.S. and China prepare for future rounds of AI governance talks, the issue of model distillation and "capability theft" is likely to become a primary point of contention. For Washington, the challenge is to find ways to protect the "reasoning" of its models without stifling the open-source collaboration that drives innovation. For Beijing, the goal is to continue the rapid absorption of global AI advancements to ensure that the PLA is not left behind in the most significant military-technological revolution of the 21st century.

The evidence suggests that the "silicon curtain" intended to divide the AI capabilities of the two superpowers is far more porous than previously thought. As long as the outputs of frontier models are accessible, the transfer of intelligence will continue, regardless of how many chips are blocked at the border.