Report agrees with Anthropic CEO’s complaint to the US government on Chinese AI models

Chinese military-linked researchers are trying to transfer that expensive, proprietary reasoning from Western models into smaller systems they can control and deploy locally

Chinese military-linked researchers have used outputs from advanced American AI models made by OpenAI and Anthropic to train their own systems, according to a Reuters review of more than 80 Chinese academic papers and patents. The findings come after Anthropic CEO Dario Amodei raised concerns about China potentially using distillation to draw capabilities from powerful US frontier AI models. Reuters found that researchers linked to China’s military and security institutions have used a technique called “model distillation” to transfer selected capabilities and reasoning from US models into smaller systems that can be controlled and deployed locally, including for defence-related applications.

Chinese military researchers use US AI models for distillation

According to Reuters, the documents provide a look at how institutions linked to China’s military are using advanced US AI models while Washington tries to restrict Beijing’s access to advanced chips and other technologies. The research included material compiled by the Washington-based Jamestown Foundation, and found widespread use of model distillation among researchers linked to the People’s Liberation Army (PLA) and other military institutions.For those unaware, model distillation is a widely used AI technique in which the output of a powerful AI model is used to train a smaller model. This can allow the smaller system to gain selected capabilities without requiring the massive computing resources needed to build a frontier AI model from scratch.

OpenAI and Anthropic models reportedly used in Chinese research

A paper published last year by researchers from PLA Unit 96941, described by Reuters as a military intelligence and cyber-warfare unit in Beijing, detailed the use of OpenAI’s GPT-3.5 to process sensitive military source code.The researchers said third-party AI models were not suitable for directly handling classified information. They instead used GPT-3.5 to summarise software code and trained a domestic AI model using those summaries. The resulting system could then operate inside Chinese military networks, according to the report.Researchers at the North University of China also used Anthropic’s Claude 3 Haiku to create synthetic training data for a model designed for social media monitoring and content moderation, Reuters reported.Anthropic told Reuters that it does not offer commercial access to Claude in China or to companies controlled by Beijing. The company also warned that distilled models may not retain the safety safeguards built into the original system, potentially allowing some capabilities to move into models outside its control.

Distilled AI models tested for drones and military operations, claims report

According to the documents, the use of model distillation extended to other defence applications. A 2024 paper from the PLA’s National University of Defense Technology described using distillation to reduce the size of an image-processing AI model so that it could run aboard unmanned aerial vehicles. The approach allowed drones to analyse live video and assist with navigation and targeting even when communications were unavailable.Researchers at China’s Academy of Military Sciences similarly used distillation to operate a target-recognition model on tactical hardware during simulated maritime operations involving drones, ships and unmanned submarines, according to a study published earlier this year.China has promoted technologies including “model lightweighting” and edge computing as it faces US restrictions on access to advanced chips. Such technologies can help AI models operate on devices including drones and satellites with limited computing power.However, the dispute between the US and China is focused on the alleged unauthorised extraction of capabilities rather than model distillation itself. Sunny Cheung, a Jamestown fellow who analysed more than 60 papers, told Reuters that Chinese military researchers were attempting to capture the reasoning abilities of Western AI models.“Teaching a model the right answer is one thing but teaching it the reasoning behind the answer is much harder,” Cheung said. “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.

Distillation does not fully reproduce frontier AI capabilities

Despite its potential benefits, experts said model distillation has limits. Smaller models generally inherit selected capabilities from their larger “teacher” models rather than reproducing their full intelligence.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.”Chinese military researchers are also studying the security risks created by distillation. Researchers at the Army Engineering University published a paper in January examining “data-free distillation,” which can be used to reverse-engineer some capabilities of an AI model without accessing its underlying parameters.