Chinese entities are adopting increasingly sophisticated ways to use output from leading American AI models to train domestic rivals, Reuters reported on 17 November in an exclusive based on a review of more than 300 Chinese academic papers, patents, corporate disclosures, recruitment notices and government documents.
Reuters wrote that the practice, known as distillation, “has expanded beyond academic experiments into mainstream industrial use in China”. In one paper it cites, about 5,000 answers from OpenAI’s o1 became 10,750 bilingual training examples, about 2.15 per answer by our arithmetic. Reuters noted that distillation is not illegal.
Business Pill · Distillation
A one-minute explainer of model distillation: how a smaller model is trained to imitate a larger one. It teaches the general idea only and says nothing about any organisation in this story.
The key insight: As we read it, Reuters’ review moves the US-China AI contest from who can buy chips to who can collect answers: one paper it cites turned about 5,000 o1 answers into 10,750 training examples, about 2.15 per answer by our arithmetic.
What the Documents Show
According to Reuters, the records describe Chinese researchers using US models to produce answers to questions, build computer code and find solutions to difficult problems. Chinese companies then developed software to automate the process, and specialist workers prepared material for training smaller domestic models, the agency reported.
Reuters said the documents also show that Beijing is encouraging distillation to make AI cheaper and easier to deploy across China’s manufacturing sector. It added that Chinese government documents do not instruct companies to distil American models.
Reuters wrote that distillation “cannot reproduce a leading model’s full capabilities or eliminate the need for advanced computing power”, but can help developers keep selected skills in smaller systems that cost less to run. For how the technique works, see our distillation explainer.

Three Papers Reuters Cites
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, Reuters reported. The same researchers cautioned that heavy reliance on distillation could stifle original innovation, according to Reuters.
Alibaba researchers described a more systematic approach in an April 2025 paper, using GPT-4 and GPT-4o to generate training material in languages other than Chinese before training smaller Qwen models, Reuters said. A February 2026 paper by Shanghai Jiao Tong University and Meituan researchers used GPT-4o to generate plans for coordinating AI agents on programming problems.
Shanghai Jiao Tong University, Alibaba and Meituan did not respond to requests for comment, Reuters reported. A former Alibaba engineer, speaking anonymously, told Reuters the attraction for teams short of computing power is making one expensive query do repeated work.
The Washington and Beijing Positions
Reuters reported that at meetings in Chengdu in July, US Commerce Department officials raised concerns with Chinese regulators about developers collecting output from leading American models, warning that some activities could amount to intellectual-property theft, citing three people familiar with the discussions.
Chinese officials defended distillation as a standard technique and said they were open to acting where conduct constituted IP theft, but made no concrete commitments, a US official told Reuters. In September, US authorities publicly accused Chinese companies of targeted distillation of American models on an industrial scale, Reuters reported.
A US official told Reuters that Washington believes industrial-scale distillation is the main reason Chinese models narrowed the gap. Two people familiar with Chinese developers’ operations disputed that, calling it one factor alongside engineering, algorithms, data and model design. The US Commerce Department and China’s Cyberspace Administration did not respond to Reuters.
OpenAI’s terms restrict using its output to develop competing models, subject to exceptions, and Anthropic similarly restricts training competing systems without permission, Reuters noted. Both companies have described defences before, as our pieces on OpenAI’s distillation case and Anthropic’s distillation defences show.

Where It Sits Against the Chip Curbs
Reuters wrote that China’s broader use of distillation could undercut the effectiveness of US curbs on advanced chip sales, a day after Washington said it would further restrict China’s access to such chips, the subject of our piece on the remote-access rule. A separate export case is in our Megaspeed piece.
A national AI+Manufacturing policy package published by eight Chinese government agencies in January supports model training, fine-tuning and distillation for industrial use, Reuters reported. Chinese developers have asked the industry regulator where ordinary use of foreign models ends and organised collection of output begins, one person told Reuters; the ministry did not respond.
The dispute comes before the 18-19 November APEC summit in Shenzhen, where Donald Trump and Xi Jinping are expected to meet, about four months after the Chengdu meetings by our count. Demand for Chinese models is in our DeepSeek usage piece.
The Structural Read
As we read Reuters’ account, the three papers it cites run from a one-off bilingual dataset (November 2024) to a systematic pipeline with scoring (April 2025) to agent planning (February 2026), about 15 months by our count.
As we read the policy side, Reuters describes Beijing encouraging distillation for industry without instructing companies to distil American models, and developers asking the regulator where ordinary use ends; the report lands a day after Washington said it would further restrict China’s chip access.
Reuters, on its review of more than 300 Chinese records, 17 November 2026
“has expanded beyond academic experiments into mainstream industrial use in China”
Three Implications
FOR US MODEL PROVIDERS OpenAI and Anthropic restrict using their output to train competing systems, Reuters noted; both have described distillation defences before, per our earlier pieces.
FOR EXPORT POLICY Reuters wrote that wider distillation could undercut the effectiveness of US curbs on advanced chip sales to China.
FOR CHINESE DEVELOPERS Two people familiar with their operations told Reuters distillation is one factor alongside engineering, algorithms, data and model design.
The Business Engineer Lens
This story maps onto the Business Engineer framework The Four Intelligence Moats.
The framework says “The transformer is a commodity now; what changes between paradigms is where intelligence accumulates and who can capture it”, and describes a Corpus Moat “built by pretraining, eroding as the public web is exhausted and synthetic data closes gaps”.
As we read Reuters’ account, model outputs now sit beside chips in the export-control argument: a US official told Reuters that industrial-scale distillation is the main reason Chinese models narrowed the gap, a view two other people disputed. That reading is our analysis; the framework does not discuss distillation or China.
What Is Not Established
Reuters did not report how much US model output Chinese developers have collected in total, or name any company as having broken a provider’s terms. Its account of the Chengdu meetings rests on people familiar with them, and the US view that distillation explains the narrowing gap is disputed in its own report.
We did not read the 300-plus documents ourselves; the paper details here are as Reuters described them. OpenAI and Anthropic are not quoted in Reuters’ report beyond their published terms, and we did not contact the companies named.
The Bottom Line
A Reuters review of more than 300 Chinese records describes distillation of US model output moving from academic papers into industrial use, with Beijing encouraging the technique. In one cited paper, about 5,000 o1 answers became 10,750 training examples, about 2.15 per answer by our arithmetic.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. This piece rests on a Reuters exclusive by Eduardo Baptista and Laurie Chen (17 November), which we read in full; the papers, policy documents, meetings and non-responses are as Reuters described them, and its account of the Chengdu talks cites people familiar with them. We did not read the 300-plus documents ourselves or contact the companies named. The 2.15 ratio is our arithmetic. Reuters noted that distillation is not illegal; nothing here says any named company broke a provider’s terms.
Sources: Reuters: China taps US AI models to train homegrown rivals as chip curbs bite (17 Nov 2026) · Reuters on X (17 Nov 2026)









