Chinese military researchers have trained domestic AI systems to improve China's defence capabilities using outputs from top US AI models created by OpenAI and Anthropic, according to a Reuters analysis of more than 80 Chinese academic publications and patents.
Despite Washington's efforts to limit Beijing's access to cutting-edge chips and other strategic technologies, the previously unreported findings provide a rare window into how Chinese military and security-related institutions are using cutting-edge US AI models as a shortcut to developing specialised systems of their own.
The documents demonstrate the widespread application of a method called "model distillation," which uses the outputs of a strong AI system to train smaller, more specialised models that can be implemented locally without the massive processing capacity needed to develop frontier AI systems from the ground up.
ALSO READ: AI Distillation Dilemma: What Is It, And Why US Labs Accuse China of Harvesting Their Reasoning?
Researchers connected to the People's Liberation Army and other military organisations frequently employ distillation, according to Reuters' study, which included data produced by the Washington-based Jamestown Foundation and supplied privately to the news agency.
According to the papers, Chinese defence organisations view cutting-edge American AI models as a source of technological knowledge and a means of catching up to American competitors. Unauthorised extraction is the main point of contention rather than distillation, which is a common commercial procedure.
Ahead of US-China negotiations on AI governance and safety, the subject has become a key flashpoint. Some Chinese organisations are allegedly utilising distillation to extract capabilities from American AI models, which could undermine export regulations and violate intellectual property rights, according to US officials.
China has denied the charges, claiming that Washington is seeking AI "hegemonism" and that identical actions have been taken by American companies.
Additionally, Chinese developers have refuted allegations that their advancements in AI are based on foreign models. Last week, Moonshot, an AI startup, refuted claims made by the Trump administration that their Kimi K3 model was developed by distillation, claiming it was powered by exclusive advancements.
Chinese military scientists are methodically documenting the reasoning steps of Western models to adapt them for surveillance, cyberwarfare, and tactical decision-making, according to Sunny Cheung, a Jamestown scholar who examined more than 60 of the publications.
"Teaching a model the right answer is one thing, but teaching it the reasoning behind the answer is much harder," said Cheung. "These papers show Chinese military-linked researchers are trying to transfer that expensive, proprietary reasoning from Western models into smaller systems they can control and deploy locally," he added.
Reuters found two dozen more cases with military connections and validated the scholarly literature.
Researchers from PLA Unit 96941, a military intelligence and cyber-warfare unit in Beijing, revealed utilising OpenAI's GPT-3.5 to parse classified military source code in a study they published last year.
Third-party models, according to the researchers, are inappropriate for managing classified data. In order to get around that restriction, they summarised software code using GPT-3.5 and trained a domestic model to operate just within Chinese military networks using those summaries.
According to Reuters' and Jamestown's analysis of the files, Chinese experts have employed distillation for everything from military deployment to content monitoring.
Researchers at the North University of China, which has strong ties to the nation's weapons industry, created artificial training data for a text classification model for social media monitoring and content moderation using Anthropic's Claude 3 Haiku.
According to Anthropic, it employs monitoring mechanisms to identify regulatory infractions and does not grant commercial access to Beijing-controlled companies or to Claude in China.
The corporation further stated that the safety precautions of the original systems may be lost in distilled models, which might allow sensitive capabilities to be transferred to models outside of its control.
In a 2024 paper, the PLA's National University of Defense Technology explained how to use distillation to reduce the size of an image-processing model for use on unmanned aerial vehicles. This would enable drones to analyse live video and support real-time navigation and targeting decisions even in the event that communications are disrupted.
According to a paper released earlier this year, researchers at China's Academy of Military Sciences employed distillation to run a target-recognition model on tactical gear during simulated maritime operations, including drones, ships, and unmanned submarines.
With Washington's export restrictions on high-end CPUs limiting its access to advanced computing capabilities, China has adopted distillation as a means of competing with the United States in frontier AI.
In order to enable AI models to operate on drones, satellites, and other devices with minimal processing capacity, central and local governments have pushed "model lightweighting" and edge computing, allocating research funds and subsidies to these technologies.
However, experts warn that distillation has serious drawbacks.
Military experts are looking at distillation as a possible security problem as Chinese AI models catch up to their American counterparts.
The possibility of "data-free distillation," a technique for reverse-engineering a model's capabilities without direct access to its fundamental parameters, was discussed in a paper published in January by researchers at the Army Engineering University.
They suggested defence measures intended to conceal the concealed logical information revealed in a model's public outputs to address that vulnerability.
Additionally, distilled models are unable to fully mimic the wide-ranging intelligence of frontier systems and only inherit specific characteristics.
According to Trevor Koverko, co-founder of Sapien, an AI data business, distilled models are still not as good as their instructor models.
ALSO READ: EM Expert Adrian Mowat Decodes AI Trade Volatility And Road Ahead For Next 12 Months
Essential Business Intelligence, Sharp Market Insights, Practical Personal Finance Advice, Daily Fuel, Gold and Silver Prices and Latest Stories — On NDTV Profit.