US AI Firms Block Chinese Labs Using Model Distillation
Anthropic and OpenAI implemented security measures to stop Chinese AI labs from harvesting data to train their own models through distillation.
Chinese artificial intelligence labs are increasingly using distillation to train their models on the output of superior American systems. This process allows Chinese firms to rapidly acquire reasoning capabilities that U.S. developers achieve through expensive reinforcement learning, narrowing the competitive gap to an estimated three to six months.
Anthropic and OpenAI have responded by implementing detection measures and closing exploits to block this data harvesting. Anthropic reported that five Chinese firms harvested over 180 million messages from its models over five months, but new security measures have since reduced that harvesting by 90-99%.
Internal divisions exist within China regarding the strategy. ByteDance founder Zhang Yiming instructed his staff to avoid distillation, arguing that relying on American models prevents China from ever surpassing the U.S. frontier. In contrast, labs like DeepSeek have used the technique to significantly improve mathematics test scores.
U.S. officials and tech firms view the practice as a commercial threat and a national security risk. Some observers suggest that China's dependence on American progress may simplify U.S. AI regulation without risking a loss of competitive advantage.
