Chip curbs meet a workaround: Chinese teams are training on American AI answers
Analysis: A Reuters review of 300-plus Chinese documents describes distillation of US model output as routine. What it proves is narrower than the US claim that it explains China's catch-up.

Washington's theory of chip export controls is that a lab without enough Nvidia hardware cannot build a frontier model. A Reuters review, published November 17, suggests Chinese developers have been working on the other half of the problem for two years: borrowing the answers from American models that were already trained.
The short version
Reuters reviewed more than 300 Chinese papers, patents, corporate disclosures, recruitment notices and government documents. It found that training smaller models on output from US systems has moved from academic work into routine industrial use. A US official calls it the main reason Chinese models have closed the gap. Two people familiar with Chinese developers dispute that.
The technique is distillation. A strong "teacher" model answers questions, and a smaller "student" model trains on those answers. It is legal and common. Labs routinely distill their own big models into cheaper ones. The dispute is over whose teacher you use and whether you asked.
What the documents show
The examples in the Reuters account are specific and dated. In a November 2024 paper, Shanghai Jiao Tong University researchers picked about 5,000 answers from OpenAI's o1 model, ran them through GPT-4o mini for translation, and built 10,750 bilingual examples to further train Alibaba's Qwen2.5-72B. An April 2025 Alibaba paper described a more systematic pipeline: GPT-4 and GPT-4o generated training material in non-Chinese languages, which was then refined and scored before it trained smaller Qwen models.
A February 2026 paper from Shanghai Jiao Tong and the food-delivery company Meituan used GPT-4o to draft plans for coordinating AI agents on programming problems. Reuters also reports that Chinese companies have built software to automate collecting US model output, and that specialist workers prepare it for training. Shanghai Jiao Tong, Alibaba and Meituan did not respond to Reuters.
10,750
bilingual training examples built from about 5,000 o1 answers
Shanghai Jiao Tong University paper, November 2024, per the Reuters review
The appeal is arithmetic. A former Alibaba engineer told Reuters that a compute-starved team gets repeated use out of a single expensive query: the American model answers a hard question once, and the Chinese team keeps the example. That is the whole pitch for people who cannot buy enough chips.
Two readings of the same pile of paper
Start with what the documents do not prove. They show that the practice exists and that it is spreading. They do not measure how much of any Chinese model's performance comes from it. A US official told Reuters that industrial-scale distillation is the main reason Chinese models have narrowed the gap. Two people familiar with Chinese developers' operations said it is one factor next to engineering, algorithms, data and model design. The Reuters account does not name the official or the two people, so neither claim can be checked here.
The researchers themselves flagged a cost. The Shanghai Jiao Tong authors warned that leaning hard on distillation could stifle original work and leave developers dependent on breakthroughs made at more advanced labs. Reuters adds that distillation cannot reproduce a leading model's full capabilities or remove the need for advanced computing power, though it can preserve selected skills in cheaper systems. That is a ceiling, and the papers acknowledge it.
The part Beijing is not hiding
Reuters found no government document telling companies to distill American models. It did find an AI+Manufacturing package, issued in January by eight agencies, that supports training, fine-tuning and distillation for industrial uses. Developers have asked the Ministry of Industry and Information Technology where ordinary use of foreign models ends and organized collection begins, according to one person familiar with those requests. The ministry did not respond.
US Commerce officials raised the issue with Chinese regulators in Chengdu in July, telling them some collection of model output could amount to intellectual-property theft, according to three people familiar with the talks. A US official said Chinese officials defended distillation as standard practice, said they would act where conduct was theft, and made no concrete commitments.
What the labs say happened to them
Anthropic published its own numbers on February 23. It said DeepSeek, Moonshot and MiniMax generated more than 16 million exchanges with Claude through about 24,000 fraudulent accounts, which violated its terms and regional access limits. By its count, MiniMax accounted for over 13 million of those exchanges, Moonshot over 3.4 million and DeepSeek over 150,000. These are Anthropic's figures about its own logs, and the labs' side of it is not in them.
Anthropic drew a policy conclusion from it: distillation attacks reinforce the rationale for export controls, because capabilities extracted from US models still depend on access to advanced chips. Reuters reports that OpenAI made a similar accusation against DeepSeek in a memo to a House committee.
The timing is not subtle. The Reuters review landed a day after Washington moved to restrict China's access to advanced chips further, and ahead of the November 18 and 19 APEC summit in Shenzhen, where Donald Trump and Xi Jinping are expected to meet. If distillation is as central as the US official says, tighter chip rules address only part of the supply chain, and the next fight is about API terms, account verification and what counts as theft.
One limit on all of this: I did not read the Reuters report directly. This piece relies on two outlets that summarize it in detail, Dataconomy and MLQ.ai, plus Anthropic's own published findings. Treat the Reuters-specific details as secondary reporting until the original is in hand.
Sources
- 01China uses US AI model outputs to train domestic rivals, Dataconomydataconomy.com
- 02Chinese companies use answers from US AI models to train domestic rivals, Reuters reports, MLQ.ai (summarizing Reuters)mlq.ai
- 03Detecting and preventing distillation attacks, Anthropicanthropic.com
Corrections
No corrections have been made to this article.
About the reporter
Technology Reporter, Trestlewire
My first published story, in the spring of 2004, was about a database company whose new release had been announced three separate times. I knew it had been announced three times because I had signed off on two of the announcements. I spent seven years as a product manager in enterprise software before I switched sides, and the habit that came with me is a single question: does this ship, or is it a demo.
Read full bio and all stories →The Morning Wire
The day’s reporting, in your inbox before nine.
Free. Unsubscribe any time. We never share your address.
