Capability can cross a boundary without the model
Distillation lets a smaller system learn from examples or judgments produced by a stronger teacher model. That can make a narrow domestic model more useful even when the recipient cannot obtain the teacher's parameters or replicate the computing infrastructure behind it. Provider terms may prohibit military use, but terms are difficult to enforce when researchers can reach outputs through intermediaries or public services.
The policy implication is not that every model response reveals a weapon. It is that export control, access policy and national-security risk cannot be separated. Governments and providers need evidence about which behaviors transfer, how cheaply they transfer, what monitoring can detect, and whether defensive restrictions create measurable friction rather than a paper boundary.
Go to the source
Read the evidence behind this analysis. External links open in a new tab.
Reuters — Chinese military researchers tap U.S. AI models for defense systems
