Overview
- A Reuters review published Friday examined more than 80 Chinese papers and patents and found researchers tied to the People’s Liberation Army used model distillation — training smaller “student” models on outputs from OpenAI and Anthropic systems — to build locally deployable tools.
- One PLA Unit 96941 paper described using GPT‑3.5 to summarise sensitive military source code and then training a domestic model on those summaries so the final system could run inside Chinese military networks.
- U.S. companies and officials say large‑scale harvesting of closed model outputs can transfer valuable capabilities and weaken export controls and IP protections, while China rejects the accusations and Chinese firms deny improper use of Western models.
- Congressional committees have opened inquiries into U.S. firms’ use of Chinese models, with lawmakers requesting records from DoorDash and pressing for stronger procurement rules, export limits, and pre‑release security testing.
- Experts note distillation transfers selected behaviours and reasoning traces but not full frontier capabilities, yet distilled systems can run on drones and ships, may lack original safety guardrails, and could change battlefield and commercial AI economics.