"Trained on OpenAI and Anthropic models"... China develops military AI through 'distillation'

Reuters and Jamestown Foundation review found multiple cases where China’s PLA-linked institutions used outputs from OpenAI and Anthropic as training data to build lightweight military AI models.

The models were applied to tasks such as drones, naval operations, and content classification, helping boost defense AI capabilities with limited computing resources.

The findings highlight how U.S. chip and AI export controls may still leave room for China to repurpose frontier models for military use, raising concerns about tech leakage and enforcement.

They also underscore the risk that distillation can strip away some safety guardrails, increasing the spread of sensitive capabilities and the chance of loss of control.