Chinese AI developers are increasingly using outputs generated by American large language models (LLMs), such as those from OpenAI, to train and improve domestic AI systems, including Alibaba’s Qwen series. This insight emerges from a Reuters investigation analyzing over 300 Chinese academic papers and corporate documents.
The technique employed is known as distillation, wherein a more powerful "teacher" model produces answers, code snippets, or solutions to complex problems. These outputs then serve as training data for smaller, more cost-effective LLMs that inherit certain capabilities from the original system.
In November 2024, researchers at Shanghai Jiao Tong University reported utilizing approximately 5,000 OpenAI o1 model responses. These were translated using GPT-4o mini to generate 10,750 bilingual training examples aimed at refining the Qwen2.5-72B model.
By April 2025, Alibaba researchers described a more structured methodology: they used GPT-4 and GPT-4o to produce multilingual training materials, which were subsequently evaluated and refined before being used to train smaller Qwen models.
Another case from February 2026 involved Shanghai University researchers collaborating with Chinese tech firm Meituan. They leveraged GPT-4o to create coordination plans for AI agents tasked with programming challenges. The resulting data was instrumental in training a compact model capable of orchestrating similar processes autonomously.
Reuters also reports that Chinese companies are developing automated tools to systematically collect responses from American AI models and prepare datasets for training purposes.
This distillation approach offers efficiency advantages by reducing the need to repeatedly access expensive, large-scale models. Instead, developers obtain high-quality examples once and use them to train their own systems, conserving computational resources.
The strategy is particularly significant for Chinese firms facing U.S. export restrictions on advanced AI chips. Distillation enables the retention of key capabilities from large models within more affordable, domestically deployable systems, potentially mitigating dependence on foreign hardware.
Nonetheless, the method does not fully replicate the original models’ capabilities and does not eliminate the necessity for modern computing chips. Industry representatives in China have expressed concerns that overreliance on external model outputs could hinder the growth of indigenous research and development competencies.
The practice of using American model outputs to train competing AI systems has become a contentious issue in U.S.-China relations. U.S. authorities contend that large-scale collection of responses from leading AI models may exceed accepted distillation norms and infringe on intellectual property rights.
During negotiations with Chinese regulators in July, representatives from the U.S. Department of Commerce warned of potential IP violations arising from such practices. Chinese officials, however, defended distillation as a standard AI development technique and indicated willingness to address intellectual property concerns, though Beijing has yet to make specific commitments.
OpenAI’s terms of use explicitly restrict the application of its model outputs for training competing systems, except under certain conditions. Similarly, Anthropic prohibits training rival models without explicit permission.
Opinions differ on the role of distillation in advancing Chinese AI capabilities. An American official familiar with bilateral discussions described it as a key factor in narrowing China’s technological gap. Conversely, sources within the Chinese AI industry emphasize the equal importance of breakthroughs in algorithms, engineering, data acquisition, and model architecture.
Chinese government documents also endorse distillation as a cost-effective means to accelerate AI adoption in industry. Notably, recommendations supporting this approach were included in the AI+Manufacturing policy package released in January 2025, prepared by eight governmental agencies.
Earlier analyses by Bank of America concluded that neither the U.S. nor China can achieve AI technological supremacy without collaboration with international partners, underscoring the complexity of the global AI ecosystem.
Chinese AI Developers Leverage Outputs from U.S. Models to Train Domestic Systems Amid Export Controls Chinese AI researchers and companies are employing outputs from American large language models (LLMs), including OpenAI’s systems, to train and refine domestic AI models through a process called distillation. This pract... Read the full IIPLA article: https://iipla.org/news/chinese-ai-developers-leverage-outputs-from-u-s-models-to-train-domestic-systems-amid-export-controls