China is increasingly using a technique known as AI model distillation to develop domestic artificial intelligence systems using outputs generated by leading American models. The approach allows Chinese researchers and technology companies to train smaller, less expensive AI systems by learning from the responses, code, and problem-solving demonstrations produced by powerful models developed in the United States.
A Reuters investigation published on November 17, 2026, found that the practice has expanded beyond academic research into commercial and industrial applications across China. Based on a review of more than 300 academic papers, patents, corporate disclosures, recruitment notices, and government documents, the investigation identified links to organizations including Alibaba, Meituan, and Shanghai Jiao Tong University. The findings raise questions about the effectiveness of US restrictions on advanced semiconductor exports, because distillation could help Chinese developers extract useful AI capabilities without independently training equally large models using restricted computing hardware.
What Is AI Model Distillation?
AI model distillation is a machine-learning technique in which a smaller model learns from the outputs of a more powerful model.
The larger system is commonly called the teacher model, while the smaller system is known as the student model.
Instead of training the student entirely from scratch using enormous datasets and expensive computing infrastructure, developers provide it with examples generated by the teacher. These examples can include answers to questions, computer code, mathematical solutions, and explanations of complex problems.
The smaller model learns to reproduce selected capabilities of the larger system.
How AI model distillation works
Powerful US AI model
Teacher model generates answers, code and demonstrations
Generated training examples
Training dataset
Examples are collected, filtered and prepared
Student-model training
Smaller domestic AI model
Learns selected capabilities with lower operating costs
The resulting model does not become an exact copy of the original. It does not automatically inherit the teacher model’s architecture, parameters, or complete capabilities. Instead, it learns selected behaviours that can make it useful for specific applications.
This distinction is important because distillation can transfer useful performance without transferring the full technology behind the original model.
How Chinese Companies Are Using American AI Models
Reuters reviewed more than 300 Chinese documents and found evidence that distillation is becoming a more organized part of China’s AI development ecosystem.
The investigation identified researchers and companies using outputs from American AI models to create training material for domestic systems.
Examples include:
- Alibaba: Research and development activities involving the use of advanced AI model outputs to support domestic model training.
- Meituan: Work involving AI-generated material for developing and improving its own AI systems.
- Shanghai Jiao Tong University: Academic research using outputs from powerful AI models to develop training datasets.
The documents describe methods involving automated software, specialist workers, and systems designed to generate large volumes of questions and responses.
Rather than manually collecting a small number of examples, developers can automate the process and create extensive datasets for training smaller models.
The investigation also found that Chinese researchers are increasingly interested in collecting detailed problem-solving demonstrations rather than only final answers. Such examples can help student models learn how to approach complex tasks.
Why Distillation Matters for China’s AI Industry
Training frontier AI models requires enormous computing resources, advanced processors, large datasets, and significant financial investment.
US restrictions have limited China’s direct access to some of the most powerful American AI chips, including advanced Nvidia processors.
Distillation offers an alternative way to develop useful AI capabilities with fewer resources.
A Chinese developer can use an existing powerful model to generate training examples and then use those examples to improve a smaller domestic system.
The resulting model may be cheaper to train, operate, and deploy across businesses.
This is particularly relevant for China’s manufacturing industry, where companies want AI systems that can support production planning, quality inspection, automation, logistics, and industrial decision-making.
The technique could therefore help China expand practical AI adoption even when access to the most advanced computing hardware remains restricted.
However, distillation does not eliminate the need for computing infrastructure. Developers still need resources to generate training data and train the smaller models, and the technique cannot reproduce every capability of a frontier system.
Real-World Examples of AI Distillation in China
The Reuters investigation identified several research projects illustrating how American AI outputs have been used to improve Chinese models.
One example involved researchers at Shanghai Jiao Tong University, who described selecting approximately 5,000 answers generated by OpenAI’s o1 model. They translated the material using GPT-4o mini and assembled 10,750 bilingual training examples to further train Alibaba’s Qwen2.5-72B model.
Other research from Alibaba researchers described using GPT-4 and GPT-4o to generate multilingual training material, which was subsequently refined, scored, and used to train smaller Qwen models.
In another February 2026 research paper, Shanghai Jiao Tong University and Meituan researchers described using GPT-4o to generate plans for coordinating AI agents working together on programming problems.
These examples demonstrate how outputs from advanced American models can become reusable training material for Chinese AI developers.
However, the researchers also acknowledged a potential disadvantage: excessive dependence on distillation could limit original innovation and make developers reliant on breakthroughs produced by more advanced AI companies.
Why OpenAI and Anthropic Are Concerned
The controversy is not simply about using one AI model to train another. Distillation is a widely accepted machine-learning technique used by researchers and companies around the world, including in the United States.
The dispute concerns how developers obtain the training material and whether they have permission to use proprietary AI outputs to build competing products.
Companies such as OpenAI and Anthropic have restrictions on using their model outputs to develop competing AI systems without authorization, subject to applicable exceptions.
Both companies have raised concerns about large-scale attempts to extract capabilities from their models.
Anthropic has previously accused Chinese AI companies, including DeepSeek, Moonshot and MiniMax, of conducting coordinated campaigns targeting its Claude models. OpenAI has also reported detecting attempts by Chinese actors to use its models for distillation.
The concerns include automated account networks, large-scale querying, and techniques intended to obtain valuable capabilities in areas such as software engineering and advanced reasoning.
These are allegations about specific activities, not evidence that all Chinese companies using distillation are engaging in unauthorized extraction.
China Is Encouraging Distillation for Industrial AI
China’s growing interest in distillation is also supported by its broader strategy to expand AI adoption across manufacturing and other industries.
A national AI-plus-manufacturing policy package published in January 2026 by eight Chinese government agencies includes support for model training, fine-tuning, and distillation for industrial applications.
Cities such as Shanghai and Shenzhen have also encouraged techniques that reduce computing requirements.
The objective is to make AI systems cheaper and easier to deploy across factories, logistics operations, industrial equipment, and other business environments.
Smaller models can be particularly useful in situations where companies do not need the most powerful general-purpose AI system. A specialized model trained for quality inspection or production planning may be sufficient for a specific industrial task.
Chinese AI developers have also reportedly sought guidance from Beijing’s industry regulator about the boundaries between ordinary use of foreign AI services and organized collection of model outputs for training.
China’s position is that distillation is a standard AI-development technique. Its officials have indicated openness to addressing conduct that constitutes intellectual-property theft, but have not made specific commitments to restrict the practice.
How Distillation Could Undermine US Chip Export Restrictions
Washington has imposed restrictions on China’s access to advanced American semiconductors to limit its ability to develop the most powerful AI systems.
These restrictions target advanced processors and the computing infrastructure needed to train frontier models.
However, distillation introduces another dimension to the technology competition.
If a Chinese company can obtain useful answers and demonstrations from a powerful American model, it can use those outputs to train a smaller domestic system. That model may require fewer resources to operate and can be adapted for specific commercial applications.
This does not eliminate China’s need for advanced chips. Training larger models, generating substantial datasets, and developing frontier-level systems still require significant computing infrastructure.
Nevertheless, distillation could allow Chinese developers to obtain more value from limited computing resources.
The US government is responding on multiple fronts. On November 16, 2026, Washington announced additional restrictions targeting indirect access to advanced AI computing through overseas data centers.
The new measures seek to prevent companies in restricted countries from accessing American AI chips remotely through infrastructure located in third countries.
This development shows that Washington’s concerns extend beyond the direct sale of semiconductor hardware to the wider computing infrastructure that supports AI development.
US-China AI Competition Is Expanding Beyond Semiconductors
The growing dispute over distillation adds another layer to the technology rivalry between the United States and China.
Until recently, much of the debate focused on access to advanced processors, semiconductor manufacturing equipment, and large-scale data centers.
Now, control over AI-generated training data and access to proprietary model capabilities are becoming equally important issues.
| Area | US position | China’s approach |
|---|---|---|
| Advanced AI chips | Restrict access to selected high-performance processors | Expand domestic semiconductor capabilities |
| AI model development | Protect proprietary models and commercial research | Develop domestic alternatives using available techniques |
| Model distillation | Challenge unauthorized extraction of proprietary capabilities | Support distillation as an AI-development technique |
| Industrial AI | Expand advanced AI applications | Deploy smaller, lower-cost models across manufacturing |
| Computing infrastructure | Restrict indirect access to controlled computing power | Seek alternative sources of computing capacity |
The disagreement is also becoming a diplomatic issue. Reuters reported that US and Chinese officials discussed distillation during meetings in Chengdu in July 2026, with Washington raising concerns about potential intellectual-property theft.
The issue is gaining additional importance ahead of the November 18–19 Asia-Pacific Economic Cooperation summit in Shenzhen, where US President Donald Trump and Chinese President Xi Jinping are expected to meet.
The Bigger Picture
AI distillation demonstrates that technological advantages are not determined exclusively by who owns the most powerful chips or trains the largest models. Access to existing AI capabilities can also help competitors develop cheaper alternatives.
For China, this offers an opportunity to expand practical AI adoption despite restrictions on advanced computing hardware. For American AI companies, it creates a challenge involving the protection of proprietary technology and the economics of providing access to powerful models.
However, distillation has limitations. A smaller model cannot automatically reproduce the full reasoning, reliability, knowledge, or general capabilities of its teacher. The technique can accelerate development, but it cannot replace original research, engineering, and access to computing resources.
Looking Ahead
The dispute over AI distillation is likely to become an increasingly important part of US-China technology relations. American companies will seek stronger protections against unauthorized extraction of proprietary model capabilities, while Chinese developers will continue exploring ways to make AI cheaper and more accessible. The central challenge will be distinguishing legitimate research and model optimization from industrial-scale attempts to exploit restricted technology.
Ultimately, the AI race may depend on more than semiconductor access. The ability to generate valuable training data, build efficient models, and deploy AI across industries could become equally important competitive advantages. China’s growing use of distillation shows how countries with limited access to frontier computing can still find alternative ways to participate in the global AI industry.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.



