Chinese Military Uses US AI for Defense Training: A Peek into Rising Tech Capabilities

A Reuters investigation has revealed that Chinese defense experts have been using advanced AI technologies from prominent U.S. companies such as OpenAI and Anthropic to enhance China’s military AI applications. This finding is based on an analysis of over 80 documents from China, including academic research and patent filings.

These findings, previously undisclosed, show how China’s military and security sectors are exploiting top-tier American AI innovations to fast-track their own specialized technology development. This is happening despite U.S. attempts to limit China’s access to sophisticated technologies like advanced semiconductor chips.

The documents highlight a prevalent method called “model distillation,” where a large-scale AI model’s output is used to train more compact, specialized models. These smaller models require significantly less computational power, allowing them to be developed and used locally without the need for the vast resources required to build large-scale AI systems from the ground up.

According to research analyzed by the Jamestown Foundation in Washington and shared exclusively with Reuters, this technique is extensively used by entities associated with the People’s Liberation Army and other Chinese military bodies.

The documents indicate that Chinese defense entities view these U.S. AI platforms not only as sources of technological insights but also as means to bridge the technological divide with the United States.

This issue has become a critical point of contention ahead of impending U.S.-China discussions on AI governance and security. U.S. officials have accused Chinese groups of using model distillation to bypass export controls and violate intellectual property rights, though distillation as a practice is not itself the focus of these allegations.

China has dismissed these accusations, accusing the U.S. of seeking to monopolize AI technology. Concurrently, Chinese companies have defended their AI advancements as being based on independent innovations, not dependent on U.S. technologies.

Jamestown analyst Sunny Cheung, who reviewed over 60 studies, noted that Chinese military researchers are methodically adapting the complex reasoning capabilities of Western AI models for applications in surveillance, cyber warfare, and tactical decision-making.

“Merely teaching a model the correct answer is simpler than teaching it the underlying reasoning,” Cheung explained. He highlighted that the goal is to transfer sophisticated, proprietary reasoning capabilities into more manageable systems that can be operated and maintained within China.

Further investigations by Reuters identified an additional two dozen military-related studies.

One notable study was conducted by PLA Unit 96941, which specializes in military intelligence and cyberwarfare. This study described utilizing OpenAI’s GPT-3.5 model to summarize sensitive military code, which was then used to train a domestic AI model to function entirely within Chinese secure networks.

Requests for comment from the White House, Pentagon, China’s foreign ministry, the PLA, and OpenAI received no responses.

EXTENSIVE APPLICATIONS IN SURVEILLANCE AND MILITARY OPERATIONS

The method of distillation has been employed across various sectors in China, from content monitoring to military operations, as shown by the joint review by Reuters and Jamestown.

At North University of China, which has strong ties to the national armaments industry, researchers have utilized Anthropic’s Claude 3 Haiku model to generate synthetic data for training a text classification algorithm aimed at monitoring social media content.

Anthropic has stated that it does not allow commercial access to its Claude model in China or to entities controlled by Beijing, and it actively monitors for any policy violations.

The company also warned that models created through distillation might lose certain safety features of the original systems, which could lead to sensitive capabilities being transferred to models outside its control.

In a paper from 2024, researchers from the PLA’s National University of Defense Technology described using distillation to reduce the size of an image-processing model for use on unmanned aerial vehicles. This allows drones to process live video feeds and make real-time decisions about navigation and targeting, even in situations where communication is disrupted.

Similarly, the Academy of Military Sciences in China used distillation for target recognition models on tactical hardware during simulated maritime operations with drones, ships, and unmanned submarines, according to a study published earlier this year.

ADVANTAGES AND CHALLENGES OF MODEL DISTILLATION

China has adopted the practice of distillation to maintain competitiveness with the U.S. in cutting-edge AI technologies, especially as it faces limitations due to U.S. export controls on advanced computing components.

Chinese governments at various levels have been pushing for “model lightweighting” and edge computing, providing subsidies and research grants to develop technologies that allow AI models to function on devices with limited processing capabilities, such as drones and satellites.

However, experts warn that distillation has its own set of limitations.

As Chinese AI systems begin to narrow the technological gap with their American counterparts, military researchers are also scrutinizing distillation as a potential security threat.

In January, a study by the Army Engineering University addressed the risks of “data-free distillation,” a technique for reverse-engineering a model’s capabilities without accessing its core parameters. The researchers suggested defensive strategies to conceal the underlying logic exposed in a model’s outputs.

Distilled models inherit only specific capabilities and cannot fully replicate the comprehensive intelligence of large-scale AI systems.

Trevor Koverko, co-founder of AI data firm Sapien, remarked that distilled models are inherently less capable than the original models from which they are derived.

“It should be seen as a method to transfer selected skills into a more cost-effective, locally managed system, rather than as a means to achieve full independence from advanced AI technologies,” he explained.