Center for Security and Emerging Technology published material on how black-box distillation of models works in practice. The publication discusses the topic in the context of proposed legislation and ongoing political discussions about adversarial distillation.
The authors analyze how such distillation from closed American models could have occurred at various stages of the process, and they note the uncertainty around the methods used by Chinese laboratories. The material separately considers the potential impact of distillation on AI development in China.
The source is presented only by metadata and a brief synopsis, not by the full text or independent confirmation of the scenarios described. Therefore, the publication confirms the existence of analysis and its claimed directions, but does not prove a specific case of borrowing.
editorial commentary
Why it matters
A likely consequence is increased attention by lawmakers to the ways of transferring capabilities of closed models. The next observable signal will be the appearance of more concrete technical details or legislative formulation in the full text and related materials. Substantial uncertainty remains: the available synopsis does not allow establishing the fact of a specific distillation, its scale or impact.