China uses US AI models to train military systems

China US AI models model distillation military AI PLA

Chinese military researchers are employing artificial‑intelligence models developed by U.S. firms such as OpenAI and Anthropic to build domestic defence‑grade AI. A Reuters investigation, based on more than 80 Chinese academic papers and patents and analysis by the Jamestown Foundation, found that the People’s Liberation Army (PLA) uses a technique called model distillation – extracting the outputs of large “teacher” models and training smaller “student” models that can run on local hardware without the massive compute required for frontier models.

The practice has been documented in several PLA units. Unit 96941 used OpenAI’s GPT‑3.5 to process and summarise sensitive military source code, then trained a home‑grown model to operate entirely within China’s internal networks. A defence university applied the same method to compress an image‑recognition model for unmanned aerial vehicles, enabling real‑time navigation and target identification even when communications are lost. In August 2026 the PLA Air Force unveiled an AI‑assisted strike‑planning system that can coordinate large‑scale air campaigns involving hundreds of targets and more than 100 tactical units, while keeping human commanders in charge of final decisions.

U.S. officials say the approach may circumvent export‑control rules and infringe intellectual‑property rights. Beijing rejects the accusations, arguing that the United States is pursuing an “AI hegemony” and noting that American companies also use similar techniques.

Entities

Alibaba Group · Anthropic · Jamestown Foundation · OpenAI · People's Liberation Army · People's Liberation Army (PLA) · Senior Colonel Deng Jianping · Unit 96941

Claims

What the coverage asserts, and how many sources carry each claim.

[● 8 SOURCES] More than 80 Chinese academic papers and patents document the use of model distillation for defence purposes. emmepress.com · bccnews.com.tw · www.businessinsider.es · formiche.net · pplware.sapo.pt · +3 more
[● 8 SOURCES] Chinese military researchers have used U.S.-developed AI models (OpenAI GPT‑3.5 and Anthropic Claude) to train domestic AI systems. emmepress.com · bccnews.com.tw · www.businessinsider.es · formiche.net · pplware.sapo.pt · +3 more
[● 8 SOURCES] The technique used is model distillation, where outputs of large models train smaller models that run locally. emmepress.com · bccnews.com.tw · www.businessinsider.es · formiche.net · pplware.sapo.pt · +3 more
[● 8 SOURCES] PLA Unit 96941 used GPT‑3.5 to process and summarise military source code, then trained a domestic model for internal network use. emmepress.com · bccnews.com.tw · www.businessinsider.es · formiche.net · pplware.sapo.pt · +3 more
[● 3 SOURCES] U.S. officials allege Chinese entities are evading export controls and infringing intellectual‑property rights by using model distillation of U.S. AI models. emmepress.com · www.businessinsider.es · siliconangle.com
[● 8 SOURCES] A Chinese defence university used distillation to compress an image‑recognition model for deployment on drones, enabling real‑time navigation and target identification without communications. emmepress.com · bccnews.com.tw · www.businessinsider.es · formiche.net · pplware.sapo.pt · +3 more
[● 3 SOURCES] China denies the accusations, describing them as attempts at U.S. AI hegemony. emmepress.com · www.businessinsider.es · formiche.net
[○ 1 SOURCE] The PLA Air Force has unveiled an AI‑assisted strike‑planning system that can coordinate large‑scale air operations involving hundreds of targets and more than 100 tactical units. avioblog.it
[● 7 SOURCES] PLA Unit 96941 used GPT‑3.5 to process and summarise sensitive military source code, then trained a domestic model to operate within China’s internal networks. emmepress.com · bccnews.com.tw · www.businessinsider.es · pplware.sapo.pt · sapo.pt · +2 more
[● 7 SOURCES] China denies the accusations, saying the United States is pursuing AI hegemony. emmepress.com · bccnews.com.tw · www.businessinsider.es · pplware.sapo.pt · sapo.pt · +2 more
[● 2 SOURCES] China unveiled an AI‑assisted strike‑planning system for the PLAAF that can coordinate large‑scale air operations involving hundreds of targets and more than 100 tactical units. avioblog.it · noticiabrasil.net.br
[● 7 SOURCES] The technique used is model distillation, where outputs of large “teacher” models train smaller “student” models that can run locally. emmepress.com · bccnews.com.tw · www.businessinsider.es · pplware.sapo.pt · sapo.pt · +2 more