Back to News
NewsChinaAiFederal Reserve

China Leverages US AI Models for Domestic Development

Chinese developers are using outputs from US AI models, particularly OpenAI, to enhance their own systems, raising IP concerns.

Mercation Admin
November 18, 2026 2 min read
China Leverages US AI Models for Domestic Development

Chinese developers are reportedly utilizing outputs from American AI models, particularly those from OpenAI, to enhance their own artificial intelligence systems, including Alibaba's Qwen series. This practice, known as distillation, involves transferring capabilities from more powerful large language models (LLMs) to more compact and cost-effective solutions.

Utilization of American AI Models

Distillation works by having a larger model act as a "teacher," generating responses, code examples, or solutions to complex problems, which are then used to train a smaller LLM that absorbs specific skills from the original system. In November 2024, researchers from Shanghai Jiao Tong University reported using around 5,000 responses generated by OpenAI's GPT-4o mini, creating 10,750 bilingual training examples for refining Qwen2.5-72B.

By April 2025, Alibaba's researchers detailed a more systematic approach, utilizing GPT-4 and GPT-4o to generate training materials in various languages and refining these outputs before feeding them into smaller Qwen models. Additionally, in February 2026, a collaboration between Shanghai University and Meituan employed GPT-4o to develop coordination plans for AI agents tackling programming tasks, which were then used to train a more compact model capable of managing similar processes.

Overcoming Chip Limitations

This distillation method allows for more efficient use of computational resources, enabling developers to obtain quality examples from a costly model only once and subsequently train their own systems based on those examples. This approach is particularly relevant for Chinese companies amid U.S. export restrictions on advanced AI chips, as it helps maintain some capabilities of larger models in cheaper systems, potentially reducing reliance on foreign computing power.

However, while distillation aids in leveraging existing technologies, it does not fully replicate the original model's capabilities and does not eliminate the need for advanced chips. Concerns have been raised within the Chinese industry regarding an overreliance on external models potentially stunting domestic research advancement.

Intellectual Property Concerns

The use of American model outputs for training competing systems has sparked tensions between the United States and China. U.S. officials argue that the industrial collection of responses from leading AI models may exceed standard distillation practices and infringe on intellectual property rights. During discussions with Chinese regulators in July, U.S. Department of Commerce representatives highlighted these risks.

In response, Chinese officials characterized distillation as a standard AI development practice and expressed readiness to address any intellectual property violations, though they did not commit to specific obligations. OpenAI's usage terms restrict the application of its models for developing competing systems, while Anthropic also prohibits training competing models without permission.

There are differing views on the impact of distillation on the advancement of Chinese AI. An American official familiar with the discussions claimed it is a key factor in narrowing China's technological gap. In contrast, some industry insiders in China believe that advancements in algorithms, engineering, data, and model architecture are equally crucial.

Chinese government documents also support the development of distillation as a means to lower the cost of AI adoption in industry, with relevant guidelines included in the AI+Manufacturing package published in January by eight governmental departments.

Analysts from Bank of America have concluded that neither the U.S. nor China can achieve technological superiority in AI without international partners.