Reuters Finds China Is Training Cheaper AI on Answers Pulled From American Models

A Reuters review of more than 300 Chinese documents found that training smaller domestic AI systems on outputs from US models like OpenAI's o1 has gone industrial, with Beijing's blessing. Washington warns some of it could be IP theft, while Chinese officials call distillation a standard technique.

Nov 17, 2026
Reuters Finds China Is Training Cheaper AI on Answers Pulled From American Models
Office towers light up at dusk in Beijing's Zhongguancun tech district

China does not need to steal America's best AI models to learn from them. It just needs to ask them enough questions. So says a Reuters review of more than 300 Chinese documents, which found that training smaller domestic systems on the outputs of leading US models has moved out of the university lab and into mainstream industrial use.

The technique is called distillation. A powerful "teacher" model generates answers or demonstrates how to complete tasks, and developers use that material to train a smaller "student" system. Reuters says the papers, patents, corporate disclosures, recruitment notices and government records show researchers using US models to answer questions, write code and solve hard problems. Chinese companies have then built software to automate the process, and specialist workers prepare the material for training, according to the review.

Before the geopolitics kicks in, note that Reuters itself says that distillation is not illegal and is common in other parts of the world. It also says the method cannot reproduce a leading model's full capabilities or remove the need for advanced computing power. What it can do is keep selected skills alive in smaller systems that are cheaper to run. For a country facing chip curbs, that is the point.

The paper trail names names

The most concrete examples in the Reuters review come from published research. In a November 2024 paper, Shanghai Jiao Tong University researchers described selecting about 5,000 answers from OpenAI's o1 model, translating them with GPT-4o mini and assembling 10,750 bilingual training examples to further train Alibaba's Qwen2.5-72B. Those same researchers warned, Reuters reports, that leaning too hard on distillation could stifle original innovation and leave developers dependent on breakthroughs made elsewhere.

Alibaba's own researchers described a more systematic approach in an April 2025 paper, according to Reuters, using GPT-4 and GPT-4o to generate training material in languages other than Chinese, then refining and scoring the responses before training smaller Qwen models. A February 2026 paper from Shanghai Jiao Tong University and food delivery giant Meituan described using GPT-4o to generate plans for coordinating AI agents on programming problems, to help train a smaller model. Reuters says none of the three responded to requests for comment.

A former Alibaba engineer told Reuters, anonymously, why compute starved teams like it. An American model answers a difficult question once, and a Chinese team can keep the useful examples and train a local system around them.

"Chinese government documents do not instruct companies to distil American models but encourage distillation as part of Beijing's push to spread AI across the world's second-largest economy." Reuters review

A national AI+Manufacturing policy package published in January by eight Chinese government agencies includes guidance supporting model training, fine-tuning and distillation for industrial uses, Reuters found. Shanghai and Shenzhen have also encouraged techniques that cut computing requirements.

Washington sees it very differently. Reuters reports that US authorities publicly accused Chinese companies in September of aggressive, targeted distillation of American models on an industrial scale. At meetings in Chengdu in July, US Commerce Department officials warned Chinese regulators that some output collection could amount to intellectual property theft, according to three people familiar with the talks. Chinese officials defended distillation as a standard technique and said they were open to acting where conduct amounted to IP theft, but made no concrete commitments, a US official told Reuters. The Commerce Department and China's Cyberspace Administration did not respond to Reuters.

The commercial rules are clearer than the diplomatic ones. OpenAI's terms restrict using its output to build competing models, with exceptions, and Anthropic similarly bars training rival systems without permission, Reuters notes. Anthropic has already been building defenses against copycat training of its Claude models.

The biggest claim in the story is also the shakiest. A US official told Reuters that Washington believes industrial scale distillation is the main reason Chinese models have narrowed the gap with American rivals. Two people familiar with Chinese developers' operations disputed that, saying distillation is just one factor alongside engineering, algorithms, data and model design. The documents prove the shortcut exists. They do not prove it is the whole story, and Washington's next round of chip curbs should not pretend otherwise.