US Accuses Chinese AI Firms of Industrial-Scale Model Distillation

The U.S. government and leading AI developers have accused several top Chinese firms of conducting "malicious" and "industrial-scale" distillation of American frontier models to bypass innovation costs. This escalating conflict transforms a common technical training method into a high-stakes national security dispute just as Presidents Donald Trump and Xi Jinping prepare for high-level diplomatic talks.

In a joint advisory, the FBI, NSA, and CISA identified six Chinese AI developers—DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI—as participants in a systematic effort to mine U.S. models since 2024. The accusations center on the use of fraudulent accounts to query American systems and feed the outputs into Chinese "student" models. Anthropic provided specific data on one such campaign, reporting that approximately 24,000 fraudulent accounts were used to facilitate over 16 million exchanges to extract capabilities from its Claude model, specifically targeting DeepSeek, Moonshot AI, and MiniMax. Furthermore, U.S. government officials, including Treasury Secretary Scott Bessent, have accused Moonshot AI of distilling Anthropic's Fable model to develop its own Kimi K3 system.

The Mechanics of Distillation

Model distillation is a standard industry technique where a smaller, more efficient model is trained using the outputs of a larger, more complex "teacher" model. While typically used internally by companies to reduce compute costs and latency, the U.S. argues that Chinese firms are using the process to "freeload" on American research and development. By systematically copying the reasoning and capabilities of frontier models, these firms can effectively leapfrog the expensive and time-consuming process of original model training while potentially bypassing safety guardrails embedded in the original systems.

Geopolitical and Economic Stakes

This dispute arrives amid a broader geopolitical race for AI supremacy and tightening U.S. chip export controls. Treasury Secretary Scott Bessent characterized the practice as theft, stating that the Chinese "distil our models and they can never get ahead of us." The conflict provides significant political ammunition for the U.S. to implement even stricter semiconductor restrictions, as the infrastructure required for large-scale distillation still demands substantial computing power. For the U.S., the issue is not merely corporate intellectual property but a matter of maintaining a strategic technological edge.

Diplomatic Friction

Beijing has dismissed the allegations as "groundless" and a product of "technological hegemony anxiety." Mao Ning, a spokesperson for the Chinese Ministry of Foreign Affairs, countered the claims by asserting that China's AI development is the result of "high-level scientific and technological self-reliance." As the Trump and Xi administrations move toward their planned summit, these accusations add a layer of friction to diplomatic efforts. Observers are now watching to see if AI intellectual property and distillation practices will become a formal bargaining chip or a primary point of contention in upcoming bilateral trade and security negotiations.