Chinese companies and researchers are building increasingly automated workflows that turn American AI model outputs into training data for domestic rivals, according to a Reuters investigation published November 17, 2026. Its review of more than 300 papers, patents, corporate disclosures, recruitment notices and government documents describes businesses developing software to automate the process and using specialists to prepare training material.
The technique, known as model distillation, uses a stronger system’s outputs to train another model, often a smaller one.
A November 2024 paper by researchers affiliated with Shanghai Jiao Tong University and other institutions provides a concrete example. The team described selecting approximately 5,000 samples of distilled o1 output, using GPT-4o mini for Chinese translation and assembling 10,750 Chinese-English training pairs to fine-tune Qwen2.5-72B-Instruct.
Alibaba Cloud researchers described a more structured approach in their April 2025 DistilQwen2.5 paper. Qwen-max handled Chinese-language material, while GPT-4 and GPT-4o supplied material in other languages; additional AI agents expanded, refined and checked the examples before student-model training. Shanghai Jiao Tong University and Alibaba did not respond to Reuters’ requests for comment.
Beijing backs distillation for manufacturing
China’s Ministry of Industry and Information Technology published an AI+Manufacturing policy issued by eight government bodies on January 7, 2026. Its application guide encourages training, fine-tuning and distillation in the cloud while deploying lightweight models nearer industrial equipment, according to an English translation by Georgetown University’s Center for Security and Emerging Technology.
The guide identifies applications including production scheduling, visual inspection and predictive equipment maintenance. Reuters reported that the government documents encourage distillation generally, without instructing companies to target American models.
That industrial emphasis also appears in Alibaba’s research. The DistilQwen2.5 paper describes a database-query completion application and the integration of distillation capabilities into a cloud-native AI platform, extending the technique beyond a standalone research demonstration.
Washington targets computing access and model extraction
The investigation comes alongside another tightening of American computing controls. On November 16, 2026, the U.S. Commerce Department’s Bureau of Industry and Security announced an interim final rule addressing remote access to advanced American chips and servers through overseas data centres.
Model extraction is a separate concern. An April 23 White House memorandum alleged that foreign entities, principally based in China, were conducting unauthorised industrial-scale distillation campaigns, while also recognising legitimate distillation as an established part of the AI ecosystem.
The providers’ commercial restrictions draw their own boundaries. OpenAI’s services agreement prohibits using outputs to develop competing AI models, subject to specified exceptions, while Anthropic’s commercial terms restrict using its services to build competing products, including training competing models, without express approval.
Anthropic has also introduced technical defences. In an October 12 update, the company said early evidence suggested its new methods had reduced the amount of data harvested by distillers by around 90% to 99%; that estimate is Anthropic’s own assessment of collection activity, rather than a measurement of competitors’ resulting capabilities.
Transferred skills do not guarantee better facts
The published experiments show limits that can disappear in broader claims about copying AI capabilities. In the 2024 Shanghai Jiao Tong study, the authors reported mixed safety results and no significant factuality improvement after fine-tuning, despite improvements in other evaluations.
In one reported failure mode, generated reasoning simulated web searches and fabricated results even though the experimental model had no actual web access. Those findings concern that particular experiment, rather than establishing the quality of every distilled system.
The importance of distillation to China’s overall AI progress also remains disputed. A U.S. official told Reuters that Washington considered industrial-scale distillation the main reason Chinese models had narrowed their performance difference with American rivals, while two people familiar with Chinese developers’ operations pointed to engineering, algorithms, data and model design as additional factors. The documented training methods establish concrete uses of distillation, but do not quantify its contribution to China’s overall AI progress.

Tech Help Canada Staff researches, writes, and reviews practical content for business owners and professionals. Our coverage spans business, marketing, SEO, technology, and the tools and systems people use to grow and operate online. We focus on clear, useful information backed by research, hands-on experience, and editorial review. Learn more about our team and editorial standards. Need help with something? Contact Us







