Particle.news
Download on the App Store

China’s Military Used U.S. AI Outputs to Train Local Defence Models

U.S. officials say smaller Chinese systems learned selected reasoning from closed U.S. models and this has triggered investigations and tighter export controls.

Overview

  • Reuters reported Friday that Chinese military-linked researchers used outputs from OpenAI and Anthropic models to train domestic defence systems, including a PLA Unit 96941 paper that used GPT‑3.5 to summarize military code before building a local model.
  • Model distillation is the practice of using a large ‘teacher’ model to generate answers or step-by-step reasoning that are then used to train a smaller ‘student’ model that can run on limited hardware.
  • U.S. firms and officials contend some Chinese groups carried out large-scale, unauthorised extraction of proprietary outputs and warn distilled models can keep useful capabilities while losing built-in safety protections.
  • Commercial forces are increasing use of cheaper Chinese open-weight models, and U.S. lawmakers have opened probes into companies’ use of those systems while regulators have adjusted export rules to curb risk.
  • Experts say distilled models transfer selected skills but not full frontier intelligence, and the trend raises risks of safety erosion, intellectual property disputes and a more fragmented global AI landscape.