A new investigation has raised fresh questions about how advanced American artificial intelligence models are being used beyond the United States. According to a Reuters review of Chinese academic papers and patents, researchers linked to China’s military have been using outputs from leading AI models developed by OpenAI and Anthropic to train smaller domestic systems. The report comes as artificial intelligence becomes an increasingly important area of strategic competition between Washington and Beijing.
The findings have intensified an already sensitive debate over AI model distillation, a technique widely used across the industry to transfer knowledge from larger models into smaller, more efficient ones. While the process itself is common in AI development, the use of outputs from proprietary frontier models has become a growing point of tension between the US and China.
Reuters review links AI research to military institutions
According to Reuters, its review covered more than 80 Chinese academic papers and patents, including research gathered by the Jamestown Foundation. The documents reportedly describe multiple cases in which researchers associated with the People’s Liberation Army (PLA) and other defence institutions used outputs from advanced US AI systems to improve domestic models.
The report says many of the projects relied on model distillation, a process in which responses generated by a powerful AI model are used to train a smaller model capable of performing similar tasks with significantly lower computing requirements. This approach allows developers to deploy AI on local hardware without the enormous infrastructure needed to build a frontier model from the ground up.
According to the documents reviewed by Reuters, Chinese researchers viewed leading American AI systems as both a valuable technical reference and a way to accelerate domestic AI development despite ongoing US export restrictions on advanced semiconductor technology.
Papers describe military and cybersecurity applications
The report highlights several research papers involving military-linked institutions. One paper, attributed to researchers from PLA Unit 96941, reportedly described using OpenAI’s GPT-3.5 to analyse military software code before training a domestic AI model using summaries generated from those outputs. According to Reuters, the researchers wanted a model that could eventually operate entirely inside Chinese military networks rather than relying on external systems.
Other studies cited in the investigation describe broader military applications. Researchers at the National University of Defense Technology reportedly explored shrinking an image-recognition model so it could operate directly on unmanned aerial vehicles, enabling drones to analyse live video and assist with navigation even when communications were unavailable.
Reuters also reported that scientists at the Academy of Military Sciences examined the use of distilled AI models for target recognition during simulated maritime operations involving drones, ships and unmanned underwater vehicles. Meanwhile, another paper linked to the North University of China described using Anthropic’s Claude 3 Haiku to generate synthetic training data for text classification systems designed for content moderation and social media monitoring.
Distillation debate grows as AI rivalry intensifies
The investigation comes at a time when AI development has become one of the defining areas of technological competition between the United States and China. American officials have previously argued that some Chinese organisations use model distillation to extract valuable capabilities from US-developed AI systems, raising concerns over intellectual property and national security. Chinese officials have rejected those accusations, arguing that Washington is attempting to maintain technological dominance while pointing out that similar AI development techniques are used across the industry.
Anthropic said it does not provide commercial access to Claude in mainland China or to organisations controlled by Beijing and that it actively monitors for policy violations. The company also noted that distilled models may lose some of the safety protections built into the original systems.
Experts also caution that distillation has important limitations. While the technique can transfer specific capabilities into smaller, more efficient AI systems, it does not eliminate the need for the enormous computing resources and research required to build frontier AI models from scratch.
Trevor Koverko, co-founder of AI data company Sapien, told Reuters, “It is best understood as transferring selected capabilities into a cheaper, locally controlled system, not achieving independence from frontier AI.”
Neither OpenAI, the White House, the Pentagon, China’s Foreign Ministry, nor the PLA responded to Reuters’ requests for comment, according to the report.
As governments race to develop increasingly capable artificial intelligence systems, model distillation is expected to remain a central issue in future discussions on AI governance, security and intellectual property. The latest report illustrates how the technology is becoming intertwined with geopolitical competition, making AI development not only a commercial race but also a matter of national strategy.
