China Increasingly Using U.S. AI Models to Train Domestic Systems

Chinese entities are increasingly utilizing output from leading American artificial intelligence models to train domestic competitors, according to documentation including academic papers, patents, and corporate disclosures. This practice, known as model distillation, has moved beyond academic research into mainstream industrial application as China works to enhance its own AI capabilities while navigating tightening U.S. restrictions on advanced semiconductor imports.
A review of over 300 documents revealed that Chinese researchers are using American models to generate answers, write code, and solve complex problems. These outputs are then used by domestic companies to train smaller, more efficient local models. While distillation is a standard technical practice globally, its expansion in China is seen as a strategic effort to lower the costs of deploying AI across the country’s manufacturing sector.

Distillation involves using a large, powerful “teacher” model to generate high-quality examples, which are then used to train a smaller “student” system. A former Alibaba Group engineer noted that for Chinese developers facing hardware constraints, this method allows for a single expensive query to produce reusable training data. For example, researchers from Shanghai Jiao Tong University and Alibaba published findings detailing the use of OpenAI’s GPT models to refine and train smaller versions of Alibaba’s Qwen models. Similarly, a joint effort by Shanghai Jiao Tong and the food-delivery firm Meituan used American AI to help organize and coordinate programming tasks for smaller local systems.
U.S. officials have expressed concerns regarding this trend. During meetings in Chengdu in July, U.S. Commerce Department representatives suggested that some of these activities could amount to intellectual property theft. Chinese officials defended distillation as a standard development technique and indicated a willingness to address cases of IP theft, though no formal commitments were made, according to a U.S. official familiar with the talks. Separately, Washington officially accused Chinese firms in September of conducting industrial-scale distillation of American technology.
While some U.S. officials believe this practice is the primary reason Chinese AI models have reduced the performance gap with American counterparts, industry observers note that other factors—such as advancements in algorithms and engineering—also play significant roles. Meanwhile, Beijing has encouraged such techniques, with a January policy package from eight government agencies supporting distillation to facilitate the integration of AI into Chinese industrial sectors. Amid rising geopolitical tension ahead of the Asia-Pacific Economic Cooperation summit, some Chinese developers have sought guidance from regulators regarding the distinction between legitimate use of foreign models and the unauthorized gathering of data for training purposes.