China uses US AI models to train domestic rivals via distillation
Beijing is encouraging the use of American AI outputs to train smaller, cheaper domestic models to bypass chip restrictions.
The short version
- China uses 'distillation' to train smaller domestic models using outputs from leading American AI systems.
- The practice helps Chinese firms bypass some effects of US chip curbs by creating cheaper, more deployable AI.
- US officials warned Chinese regulators that industrial-scale distillation could amount to intellectual property theft.
- Beijing's 'AI+Manufacturing' policy package explicitly supports distillation for industrial applications.
- Researchers warn that over-reliance on distillation could stifle original innovation in Chinese AI development.
Why it matters
This move allows China to build sophisticated AI capabilities without needing the same level of high-end hardware as the US. It creates a technological shortcut that could make American export controls on chips less effective over time.
The mechanics of distillation
Distillation is a technique where a powerful 'teacher' model generates answers or demonstrates how to complete specific tasks. Developers then take those outputs and use them to train a smaller 'student' system. This allows the smaller model to learn complex skills without needing the massive computing power required to run the original large model.
For Chinese developers facing hardware constraints, the appeal lies in making one expensive query do repeated work. An American model can provide a correct answer to a difficult question once, and a Chinese team can then retain that example to train a local system. This method helps preserve selected skills in systems that are significantly cheaper to operate and easier to deploy at scale.
While the process cannot perfectly replicate the full capabilities of a leading model, it serves as a bridge for rapid development. It allows companies to build functional domestic tools by drawing on the intelligence of American rivals. This practice is not illegal and is used globally, but its scale in China has drawn significant international attention.
Industrial scale and use cases
A review of over 300 Chinese documents reveals that distillation has moved beyond academic experiments into mainstream industrial use. Chinese entities are adopting sophisticated ways to use outputs from leading American models to write computer code and solve difficult problems. Companies are even developing software to automate the collection and processing of these outputs.
Specific examples show researchers using OpenAI's o1 model to generate answers, which they then translated and assembled into thousands of training examples. These examples were used to further train Alibaba's Qwen2.5-72B model. Other projects have used GPT-4o to generate plans for coordinating AI agents to solve programming problems, helping train smaller models to organize similar tasks.
The goal is to create reusable data that improves domestic systems without requiring constant access to the original provider. By refining and scoring responses from American models, Chinese teams can create high-quality datasets in multiple languages. This systematic approach allows them to build a library of capabilities that can be integrated into various domestic applications.
Bypassing chip restrictions
The expansion of distillation is happening as China seeks to offset tightening US curbs on its access to advanced chips. Because distillation focuses on creating smaller, more efficient models, it helps China deploy AI across its powerful manufacturing sector. This strategy aims to make AI more accessible to various industries regardless of hardware limitations.
Washington has expressed concern that this practice could undermine the effectiveness of US restrictions on advanced chip sales. If Chinese developers can build high-performing AI by distilling American outputs, the need for the most advanced hardware may be slightly mitigated. US officials have warned that some of these activities could amount to intellectual property theft.
In response to these concerns, Chinese officials have defended distillation as a standard development technique. They stated they are open to taking action only where conduct constitutes actual intellectual property theft. However, they have not made any concrete commitments regarding new regulations or restrictions on the practice.
Policy and geopolitical friction
Beijing is actively encouraging these techniques through official policy packages. A national AI+Manufacturing policy published by eight government agencies includes guidance supporting model training, fine-tuning, and distillation. The goal is to spread AI across the world's second-largest economy by reducing the computing requirements for local firms.
This trend is creating friction between the world's two biggest economies. US authorities publicly accused Chinese companies of systematically distilling American models on an industrial scale in September. The tension is particularly notable ahead of high-level meetings between US and Chinese leaders in Shenzhen.
Chinese AI developers have recently asked regulators for guidance on the boundaries between ordinary use and organized collection of output. While the government encourages the practice, the line between standard research and systematic data harvesting remains a point of contention. The outcome of these discussions will determine how much further the practice can expand before facing legal hurdles.
What to watch
- Whether US authorities introduce new specific legal definitions for 'distillation' as a form of IP theft.
- How the gap between Chinese and American AI capabilities changes as more domestic models are trained on distilled data.
- If OpenAI or Anthropic update their terms of service to specifically target the industrial-scale distillation of their models.
Try it
Look up the 'AI+Manufacturing' policy package to see how China plans to integrate smaller AI models into industrial production.
Sources and links
First reported by The Business Standard. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.