Home Community Insights China Turns American AI Models Into Training Tools for Domestic Rivals as US Chip Curbs Tighten

China Turns American AI Models Into Training Tools for Domestic Rivals as US Chip Curbs Tighten

China Turns American AI Models Into Training Tools for Domestic Rivals as US Chip Curbs Tighten

Chinese companies and researchers are expanding their use of leading American artificial intelligence models to train domestic alternatives, using a technique known as distillation to reduce development costs and work around constraints on advanced computing hardware.

A Reuters review of more than 300 Chinese academic papers, patents, corporate disclosures, recruitment notices and government documents found that the practice has moved beyond isolated research projects into wider industrial use. Chinese developers are using American models to generate answers, write code and solve complex problems, then turning that output into training material for smaller systems built at home.

The approach offers China a way to acquire selected capabilities developed by American AI companies without having to reproduce every stage of their research. It also creates a new challenge for Washington, whose restrictions on advanced semiconductor exports are intended to limit China’s ability to develop powerful AI systems.

Distillation does not allow a smaller model to reproduce the full capabilities of its more advanced teacher, nor does it remove the need for computing resources. But it can transfer specific skills into cheaper systems that require less computing power to operate. That makes the technique particularly useful for Chinese companies seeking to deploy AI across manufacturing and other industrial sectors.

The findings come as the United States prepares further restrictions on China’s access to advanced chips from companies such as Nvidia. They also add another point of tension ahead of the November 18-19 Asia-Pacific Economic Cooperation summit in Shenzhen, where US President Donald Trump and Chinese President Xi Jinping are expected to meet.

How Chinese Developers Use American AI Output

Distillation involves a powerful model, known as the teacher, generating answers or demonstrating how to perform particular tasks. Developers then use those examples to train a smaller student model, which learns to reproduce selected capabilities at lower operating costs.

For Chinese developers facing limitations on computing capacity, the method can make a single expensive query serve as a reusable training resource.

A former Alibaba engineer told Reuters that an American model could answer a difficult question once, allowing a Chinese team to retain the response and use it to train a domestic system rather than repeatedly consulting the original model.

Research papers reviewed by Reuters provide examples of how the technique is being applied.

In a November 2024 paper, researchers at Shanghai Jiao Tong University described selecting about 5,000 answers generated by OpenAI’s o1 model. They used GPT-4o mini to translate the material and assembled 10,750 bilingual training examples to further train Alibaba’s Qwen2.5-72B model.

The researchers also warned that excessive dependence on distillation could weaken original innovation, leaving Chinese developers reliant on advances made by more sophisticated foreign AI providers.

Alibaba researchers described a more organized process in an April 2025 paper. They used OpenAI’s GPT-4 and GPT-4o to generate training material in languages other than Chinese, then refined and scored the responses before using them to train smaller Qwen models.

A separate February 2026 paper by researchers from Shanghai Jiao Tong University and food-delivery company Meituan detailed the use of GPT-4o to generate plans for coordinating AI agents on programming problems. Those examples were then used to train a smaller model to organize similar tasks.

The projects show how foreign AI output can become a source of reusable data for domestic developers. Rather than relying exclusively on their own models to generate training examples, Chinese companies can draw on the responses of more advanced systems to accelerate the development of selected capabilities.

Shanghai Jiao Tong University, Alibaba and Meituan did not respond to requests for comment.

Washington’s Intellectual Property Concerns

The growing use of distillation has become a geopolitical issue because leading US AI companies restrict how their models’ outputs can be used.

OpenAI’s terms prohibit using its output to develop competing models, subject to exceptions. Anthropic also restricts the use of its systems to train competing products without permission.

In September, US authorities publicly accused Chinese companies of carrying out aggressive, targeted distillation of American AI models on an industrial scale.

The dispute had also surfaced in regulatory discussions in Chengdu in July. Three people familiar with the meetings told Reuters that US Commerce Department officials raised concerns with Chinese regulators about developers collecting output from leading American AI systems, warning that some activities could amount to intellectual-property theft.

A US official familiar with the exchange said Chinese officials defended distillation as a standard AI-development technique and indicated they were open to taking action where conduct amounted to intellectual-property theft. However, they made no concrete commitments.

US Treasury Secretary Scott Bessent warned in a July 23 post on X that Washington could impose sanctions and add companies to its Entity List if their activities crossed the line into intellectual-property theft.

“When (Chinese) firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table,” Bessent wrote.

A US official familiar with the July discussions said Washington believes industrial-scale distillation is the main reason Chinese AI models have narrowed the gap with American competitors. Two people familiar with Chinese developers’ operations disputed that explanation, saying progress also comes from improvements in engineering, algorithms, data and model design.

That disagreement points to a broader difficulty for US policymakers. Export controls can restrict access to advanced chips, but they cannot easily prevent foreign developers from learning from information and outputs generated by models already accessible to them.

Beijing’s Cheaper Route to Industrial AI

Chinese government documents reviewed by Reuters do not direct companies to distill American models. They do, however, encourage distillation as part of a wider effort to expand AI adoption across the economy.

A national AI+Manufacturing policy package published by eight government agencies in January supports model training, fine-tuning and distillation for industrial applications. Authorities in Shanghai and Shenzhen have also promoted techniques designed to reduce computing requirements.

The policy direction is consistent with Beijing’s ambition to make AI useful across its manufacturing base, where smaller and less expensive models could be deployed for specialized tasks without requiring the resources needed to operate frontier systems.

This could broaden the economic value of China’s domestic AI development even if its companies remain behind American rivals in building the most advanced models. A manufacturer may not need a system with the full capabilities of a leading general-purpose model if a cheaper alternative can handle a specific production, design or quality-control task reliably.

However, the expanding practice has also created uncertainty for Chinese developers. One person familiar with industry discussions told Reuters that companies have recently asked Beijing’s industry regulator for guidance on the boundaries between ordinary use of foreign models and organized collection of their outputs for training.

The dispute exposes two competing dimensions of the AI race. Washington is seeking to restrict China’s access to the hardware needed to build and operate advanced systems, while Chinese developers are finding ways to extract more value from capabilities already available through foreign models.

Distillation cannot replace advanced chips, eliminate the need for original research, or guarantee that smaller models will match their teachers. But it can make limited computing resources go further and reduce the cost of deploying AI at scale.

No posts to display

Post Comment

Please enter your comment!
Please enter your name here