Chinese military researchers have been using the outputs of leading American artificial intelligence models to help train domestic defence systems, according to a review of more than 80 Chinese academic papers and patents that has raised fresh concerns about technology competition between Beijing and Washington.
The review found that institutions linked to the People’s Liberation Army (PLA) used responses generated by AI models developed by U.S. companies, including OpenAI and Anthropic, to train smaller domestic systems through a technique known as model distillation. The approach allows developers to replicate selected capabilities of advanced AI models without incurring the enormous computing costs required to build frontier systems from scratch.
Model distillation involves using a large “teacher” model to generate responses that are then used to train a smaller “student” model capable of running on less powerful hardware. Analysts say the method is particularly attractive for Chinese developers facing U.S. export restrictions on advanced semiconductor technology.
According to the review, Chinese researchers have applied the technique to a range of military and security projects, including social media monitoring, drone navigation and targeting, maritime surveillance, and software analysis.
One paper from the North University of China, which has links to the country’s defence industry, described using Anthropic’s Claude 3 Haiku to generate synthetic training data for a text classification system designed for online content monitoring.
Researchers at the PLA’s National University of Defense Technology reported compressing an image-processing model for deployment on unmanned aerial vehicles, allowing drones to analyse live video and support navigation and targeting even when communications are disrupted.
Another study from China’s Academy of Military Sciences described testing a target-recognition system on tactical hardware during simulated maritime operations involving drones, ships and unmanned submarines.
Meanwhile, researchers linked to PLA Unit 96941 reportedly used OpenAI’s GPT-3.5 to summarise military source code before training a domestic model on those summaries, enabling the resulting system to operate entirely within Chinese military networks.
Analysts say the key objective is not simply to copy answers but to capture the reasoning patterns demonstrated by advanced Western AI models.
Sunny Cheung, a fellow at the Jamestown Foundation who analysed many of the papers, said Chinese military researchers appear to be systematically transferring reasoning capabilities from Western AI systems into locally controlled models for surveillance, cyber operations and battlefield decision-making.
Experts caution, however, that distilled models do not match the overall capabilities of the original systems. Trevor Koverko, co-founder of AI data company Sapien, said the process transfers selected functions into smaller, more efficient models rather than creating equivalents to frontier AI systems.
The findings come amid growing tensions between the United States and China over artificial intelligence, semiconductor exports and national security. Washington has argued that extracting capabilities from American AI systems undermines export controls and raises intellectual property concerns, while Beijing has rejected those claims and accused the United States of attempting to maintain technological dominance.
Anthropic has said it does not provide commercial access to its Claude models in China and employs monitoring systems to detect policy violations. The company has also warned that distilled models may not retain the safety protections built into the original systems.
The growing use of AI in defence by both the United States and China has intensified debate over how advanced AI technologies should be governed, particularly as the two countries continue discussions on artificial intelligence safety, security and international regulation.
Leave a comment