PLA labs distilled OpenAI and Anthropic outputs into drone and targeting AI

A Reuters review of more than 80 Chinese papers and patents found military-linked researchers training domestic defense AI on outputs from OpenAI and Anthropic models, a transfer route Washington's chip export controls do not cover.

Chinese military and security-linked researchers used outputs from OpenAI and Anthropic models to train smaller domestic systems for defense work, Reuters reported on July 31 after reviewing more than 80 academic papers and patents. The review drew on research compiled by the Washington-based Jamestown Foundation and shared exclusively with the agency.

The technique is model distillation: outputs from a large system become training data for a small one that runs locally, without frontier-scale compute. A 2024 paper from the PLA's National University of Defense Technology described shrinking an image-processing model to fly on unmanned aerial vehicles, analyzing live video for navigation and targeting when communications are cut, per Reuters. Researchers at the Academy of Military Sciences ran a distilled target-recognition model on tactical hardware in simulated maritime operations involving drones, ships and unmanned submarines. A paper from PLA Unit 96941, a Beijing intelligence and cyber-warfare unit, used GPT-3.5 to summarize sensitive military source code, then trained a domestic model on those summaries so it stayed inside Chinese military networks.