Chinese AI Firms Distill US Models at Industrial Scale
A Reuters review of more than 300 Chinese documents shows systematic use of US AI model output to train domestic rivals, even as Washington tightens chip export controls.
By Daniel Okafor
5 min read

What's News
- A Reuters review published November 17 examined more than 300 Chinese academic papers, patents, disclosures and government documents.
- Shanghai Jiao Tong University researchers used about 5,000 answers from OpenAI's o1 to build 10,750 bilingual training examples for Alibaba's Qwen2.5-72B.
- The US accused Chinese firms in September of 'aggressive, targeted distillation of American AI models on an industrial scale.'
- New Commerce Department rules require overseas data centers serving China-headquartered firms to obtain US licenses within 90 days.
- The US and nine allies signed the San Francisco Declaration on November 16 under the Pax Silica initiative.
Chinese AI companies are systematically using output from leading American models to train their own competing systems, according to a Reuters review published November 17 of more than 300 Chinese academic papers, patents, corporate disclosures, recruitment notices and government documents. The practice, known as distillation, has moved from academic experiments into mainstream industrial use—and it is testing the reach of US export controls at the exact moment Washington is tightening them.
The findings landed the same week the US Commerce Department restricted shipments of advanced chips to Malaysia, Singapore, Thailand, Indonesia and Brazil. The rule also imposed new licensing requirements on overseas data centers that supply significant computing power to China-headquartered firms. It was published days after Nvidia CEO Jensen Huang resigned amid reports of an affair with a suspected Chinese spy at Singapore-based cloud partner Megaspeed.
How does distillation actually work?
Distillation uses a powerful "teacher" model to generate answers or demonstrate task completion. Developers then use that output to train a smaller "student" system. The mechanism matters because Chinese developers face hard limits on computing power.
The economics are simple, according to a former Alibaba engineer who spoke to Reuters: "the attraction is in making one expensive query do repeated work." An American model answers a difficult question once. The Chinese team keeps the useful examples and trains a local system around them.
Distillation cannot reproduce a leading model's full capabilities or eliminate the need for advanced chips. But it can preserve selected skills in smaller systems that cost less to run—reducing, though not ending, reliance on American technology.
What did the documents show?
Reuters identified specific, dated cases in Chinese publications:
- November 2024: Shanghai Jiao Tong University researchers selected about 5,000 answers from OpenAI's o1 model, translated them using GPT-4o mini, and assembled 10,750 bilingual training examples to further train Alibaba's Qwen2.5-72B. The researchers cautioned that excessive reliance on distillation could stifle original innovation.
- April 2025: Alibaba researchers described using OpenAI's GPT-4 and GPT-4o to generate training material in languages other than Chinese, then refining and scoring responses before training smaller Qwen models.
- February 2026: Researchers from Shanghai Jiao Tong University and food-delivery company Meituan used GPT-4o to generate plans for coordinating AI agents on programming problems, helping train a smaller model to organize such tasks.
Alibaba, Meituan and Shanghai Jiao Tong University did not respond to Reuters requests for comment.
How are Washington and Beijing responding?
US authorities publicly accused Chinese companies in September of "aggressive, targeted distillation of American AI models on an industrial scale." In July, Commerce Department officials raised concerns with Chinese regulators in Chengdu, warning that some activities could amount to intellectual-property theft, according to three people familiar with the discussions.
Chinese officials defended distillation as a standard AI-development technique. They said they were open to acting where conduct constituted IP theft but made no concrete commitments, a US official said.
A US official familiar with the Chengdu talks said 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 assessment. They pointed to advances in engineering, algorithms, data and model design alongside distillation.
The terms of service leave little ambiguity. OpenAI restricts using its output to develop competing models, subject to exceptions. Anthropic similarly restricts training competing systems without permission.
Does Chinese policy encourage the practice?
Chinese government documents do not instruct companies to distil American models. But they encourage distillation as part of Beijing's push to spread AI across the economy. A national AI+Manufacturing policy package published by eight government agencies in January includes guidance supporting model training, fine-tuning and distillation for industrial applications. Shanghai and Shenzhen have also encouraged techniques that reduce computing requirements.
The policies reflect Beijing's ambition to deploy smaller, cheaper AI systems across industries. Chinese AI developers, concerned by the growing geopolitical sensitivities, have in recent months asked Beijing's industry regulator for guidance on the boundary between ordinary use of foreign models and organized collection of output for training, according to one person familiar with the discussions. The Ministry of Industry and Information Technology did not respond to a request for comment.
Can chip rules close the gap?
The Commerce Department's new rule requires overseas data centers providing significant compute volume to China-headquartered companies to obtain US export licenses within 90 days or face future restrictions. It also restricts chip shipments to data centers above a certain compute threshold. Washington frames the measures as closing a loophole that lets Chinese firms access advanced US AI chips through overseas data centers.
The diplomatic track is moving in parallel. The United States and nine allies—including Australia, India, Japan, South Korea and the United Kingdom—signed the San Francisco Declaration on November 16 to coordinate protection of AI technologies and critical infrastructure under the Pax Silica initiative. Germany, Sweden, the Netherlands and the European Commission are set to sign next week.
At the APEC summit in Shenzhen, where President Trump and President Xi are expected to meet November 18-19, APEC chair Chen Xu said members had "reached an important consensus around AI-related issues," including "the safe deployment and responsible application of AI."
Two questions remain open. Reuters' review covers documents through early 2026, and the current scale of distillation in production systems is not quantified. Whether the new chip restrictions will curb the practice—given that distillation needs far less compute than training from scratch—has yet to be assessed.
Source: epyhia.com
More from Daniel Okafor
Show full bio
Correspondent covering money, funding and market shifts around entrepreneurship, startups and venture funding at Business Bearings.
2,206 articles