Editorial Team
4 min read
China Disputes U.S. Allegations of Industrial-Scale AI Model Distillation
U.S. cyber agencies allege that six China-based AI companies used large-scale querying to extract proprietary capabilities from American frontier models. Beijing calls the accusation groundless, leaving the technical evidence and the companies’ responses as the central unresolved issues.
- U.S. agencies allege that six China-based AI companies extracted proprietary capabilities from American frontier models through large-scale knowledge-distillation campaigns.
- China says the allegation is groundless and argues that distillation is a standard industry practice.
- The agencies’ immediate prescription is operational: providers should detect suspicious use, alter responses to suspected campaigns and share intelligence.
China has rejected U.S. allegations that China-based AI companies used industrial-scale knowledge distillation to extract proprietary capabilities from American frontier models. The U.S. claims come from a joint advisory by the Cybersecurity and Infrastructure Security Agency, the National Security Agency and the FBI; they are allegations, not a public adjudication.
The dispute is consequential because it is about both a commonplace machine-learning technique and the terms on which model outputs can be used. CISA, the U.S. agency that describes itself as the nation’s cyber-defense agency and coordinator for critical-infrastructure security, says a less capable model can learn from a stronger model’s outputs through distillation. It also says that method can be misused to obtain a competitor’s capabilities more quickly and cheaply than developing them independently.
CISA’s advisory lists six China-based AI companies, named U.S. models and alleged functional areas. Source: www.cisa.gov.
What the U.S. agencies allege
In its September 8 release, CISA said the three agencies believe DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI extracted billions of tokens through millions of exchanges or requests from variants of Anthropic’s Claude, OpenAI’s GPT, Google’s Gemini and xAI’s Grok since at least late 2024. The agencies said this occurred likely with Chinese government awareness.
The release alleges that the companies violated providers’ terms of service, rather than merely using a conventional training method. It says the campaigns shortened research and development and undermined fair competition. That framing matters: the asserted problem is allegedly unauthorized access to model outputs, not distillation in the abstract.
