An employee works at a semiconductor manufacturing facility
An employee works at a semiconductor manufacturing facility in Dresden, Germany, on June 10, 2026. Chinese technology companies are increasingly using outputs from American AI models to train domestic systems as US restrictions on advanced AI chips continue. Getty Images

Chinese researchers and technology companies are increasingly using American artificial intelligence models to train domestic competitors, a practice that could undermine Washington's efforts to restrict China's access to advanced AI technology.

A Reuters investigation published Tuesday examined more than 300 Chinese academic papers, patents, corporate disclosures, recruitment notices and government documents. The findings show how a technique known as distillation has expanded from academic research into commercial AI development.

The practice allows developers to use answers and other outputs generated by powerful AI systems, including those developed by OpenAI, to train smaller models capable of performing similar tasks.

Chinese companies are also developing software to automate the process, while specialist workers prepare the resulting material for training domestic systems.

The findings come as the Trump administration tightens restrictions on China's access to advanced American AI chips, including technology supplied by Nvidia.

Distillation could help Chinese developers reduce some computing costs and improve domestic models despite restrictions on advanced semiconductors. However, the technique cannot fully reproduce the capabilities of leading AI systems or eliminate the need for powerful computing hardware.

The process involves using a more advanced AI system, often called a teacher model, to generate answers or demonstrate how to complete tasks. Those outputs are then used as training material for another model.

One example identified by Reuters involved researchers at Shanghai Jiao Tong University.

In a November 2024 paper, the researchers described selecting approximately 5,000 answers generated by OpenAI's o1 model and translating them using GPT-4o mini.

The researchers then assembled 10,750 bilingual training examples to further train Alibaba's Qwen2.5-72B model.

The approach allowed the researchers to reuse outputs from American AI systems to improve a domestic model without repeatedly querying the original systems for the same information.

The researchers also warned that excessive dependence on distillation could discourage original innovation, leaving developers reliant on advances made by other companies.

Other Chinese technology companies have explored similar approaches.

In an April 2025 paper, Alibaba researchers described using OpenAI's GPT-4 and GPT-4o to generate training material in languages other than Chinese. The material was subsequently refined and evaluated before being used to train smaller Qwen models.

A separate February 2026 paper by researchers from Shanghai Jiao Tong University and Chinese food delivery company Meituan described using GPT-4o to generate plans for coordinating AI agents working on programming tasks.

Those plans were then used to help train a smaller model to manage similar activities.

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

The growing use of distillation has become another source of disagreement between Washington and Beijing over access to advanced AI technology.

The technique itself is widely used in artificial intelligence development and is not inherently illegal. However, American technology companies have raised concerns about competitors collecting large quantities of model outputs to develop rival systems without authorization.

OpenAI restricts certain uses of its model outputs to develop competing AI systems, subject to exceptions. Anthropic also imposes restrictions on using its services to train competing models without permission.

US officials have warned that some large-scale efforts to extract model outputs could amount to intellectual property theft.

According to Reuters, representatives from the US Commerce Department raised those concerns with Chinese regulators during meetings in Chengdu in July.

Chinese officials defended distillation as a standard AI development technique and indicated they were open to addressing activities that constituted intellectual property theft. However, they did not make specific commitments during the discussions, according to a US official familiar with the exchange.

Treasury Secretary Scott Bessent also warned in July that Chinese companies engaging in covert, industrial-scale distillation activities involving intellectual property theft could face sanctions or placement on the US Entity List.

The dispute has intensified as Chinese developers seek to close the performance gap with American AI companies.

A US official familiar with the Chengdu discussions told Reuters that Washington believes industrial-scale distillation has been a major factor in the improvements achieved by Chinese AI models.

However, two people familiar with Chinese developers' operations disputed that assessment, arguing that advances in engineering, algorithms, training data and model design have also contributed.

Chinese government policies have encouraged the wider use of distillation as part of Beijing's efforts to make artificial intelligence cheaper and more accessible across industries.

A national AI manufacturing policy package published by eight Chinese government agencies in January supported techniques including model training, fine-tuning and distillation for industrial applications.

Regional authorities in Shanghai and Shenzhen have also encouraged approaches that reduce computing requirements.

However, the government documents reviewed by Reuters did not specifically direct Chinese companies to distill American AI models.

The distinction is important because distillation can be carried out using domestic technology and does not necessarily involve unauthorized access to another company's systems.

The latest findings come as Washington expands its efforts to limit China's access to the computing power needed to develop advanced AI systems.

On Monday, the Trump administration introduced additional restrictions targeting remote access to American AI chips and computing infrastructure, extending existing controls intended to limit Chinese access to advanced semiconductors.

Those measures could make it more difficult for Chinese developers to obtain computing resources used to train sophisticated AI models.

Distillation offers a way to improve certain capabilities using outputs from existing systems, but its effectiveness depends on the quality of the training material and the resources available to developers.

Reuters also reported that some Chinese AI companies have sought guidance from Beijing's industry regulator on where legitimate use of foreign models ends and organized collection of training outputs becomes problematic.