Chinese Military Uses OpenAI Models to Train Defense AI

Chinese military researchers are using the outputs of leading US artificial intelligence models from OpenAI and Anthropic to develop domestic AI systems intended to strengthen the country's defense capabilities, according to a Reuters analysis of more than 80 Chinese scientific papers and patents.

The researchers are using a technique known as model distillation, in which the outputs of a powerful AI model are used to train smaller models that can run on significantly less powerful hardware. Reuters, working with the Jamestown Foundation, found widespread use of the technique among researchers affiliated with the People's Liberation Army (PLA) and other Chinese military institutions.

One study describes researchers from the PLA's Unit 96941, a military intelligence and cyber unit, using OpenAI's GPT-3.5 to process military source code. The resulting summaries were then used to train a domestic AI model designed to operate exclusively on military networks. Reuters was not able to independently verify whether the system has been deployed.

Other papers describe the use of model distillation to develop systems for monitoring social media, deploying AI models on unmanned aerial vehicles and improving target recognition during simulated naval operations.

Reuters noted that model distillation itself is widely used across the AI industry. The dispute instead centers on allegations that Chinese researchers obtained outputs from proprietary US AI models without authorization. US officials argue that some Chinese entities are circumventing export controls and may be infringing intellectual property rights. Beijing has rejected those allegations, accusing Washington of seeking technological dominance in artificial intelligence.

(Reuters, Lud)