The race to build the world’s most powerful AI models is getting increasingly competitive, and China has found an interesting way to keep up with its American rivals. Rather than relying entirely on expensive chips and building everything from scratch, Chinese developers are using answers generated by leading US AI models to improve their own systems.

According to a new Reuters investigation, which examined more than 300 academic papers, patents, company records, and government documents, this approach has become increasingly common across China’s AI industry. It’s another development in a market where Chinese AI companies already compete aggressively on pricing. The technique is called AI distillation, and while it isn’t new or inherently illegal, the scale at which Chinese companies are reportedly adopting it is raising concerns in Washington.

How does China turn American AI into its own advantage?

Think of distillation as a student learning from a more experienced teacher. A powerful AI model generates answers, writes code, or works through complicated problems. Developers then use those responses to train smaller models that can perform similar tasks without requiring nearly as much computing power. For Chinese companies facing restrictions on advanced American chips, that’s especially attractive. Although some companies, including ByteDance and Tencent, have secured shipments of Nvidia’s H200 AI chips, access to the most advanced American hardware remains restricted.

And this isn’t just theoretical. Researchers at Shanghai Jiao Tong University previously used responses from OpenAI’s o1 model to create thousands of training examples for Alibaba’s Qwen AI model. Other research projects involving Alibaba and Meituan have explored similar techniques using GPT-4 and GPT-4o. What’s interesting is that these efforts aren’t limited to universities anymore. Reuters found evidence of companies developing automated tools to collect AI-generated responses, alongside specialized workers preparing that information for training. The Chinese government is also encouraging the broader use of distillation to make AI systems cheaper and easier to deploy, particularly in manufacturing. However, distillation doesn’t magically turn a smaller model into an exact copy of a more powerful one. It can transfer certain capabilities, but developers still need considerable expertise, computing resources, and original research to build competitive systems.

Why Washington is getting worried

The growing popularity of distillation presents a complicated challenge for the US, which has spent years restricting China’s access to advanced AI chips. Tensions have already spilled into other areas, with the US recently banning DeepSeek from federal civilian agencies, including third-party software built on its models. If Chinese developers can improve their models using the outputs of American systems, those hardware restrictions might not be as effective as Washington hopes. There’s also the question of intellectual property. Companies including OpenAI and Anthropic restrict the use of their models’ outputs to train competing systems without appropriate permission.

US officials have already raised concerns about large-scale distillation, warning that activities crossing into intellectual-property theft could trigger sanctions. Chinese officials, meanwhile, have defended distillation as a legitimate development technique, although they have acknowledged that actual IP violations warrant action. The timing is particularly significant, with Presidents Donald Trump and Xi Jinping expected to meet at the November 18–19 APEC summit in Shenzhen. For now, distillation remains a legitimate and widely used AI training method, but how companies obtain and use that training material is becoming a major point of contention. And as Chinese AI models continue improving, the bigger question is whether restricting access to powerful chips will be enough to maintain America’s technological advantage.

Share.
Exit mobile version