Chinese companies and researchers are expanding the use of responses generated by American artificial intelligence models to develop their own systems. A Reuters investigation, published this Tuesday (17), analyzed more than 300 Chinese documents and identified the expansion of this practice from academic laboratories to industrial-scale operations.
Known as knowledge distillation, the technique allows certain capabilities to be transferred from an advanced model to another system. Instead of developing these skills entirely from scratch, researchers use responses, code and solutions generated by an existing AI as training data.
In November 2024, researchers at Shanghai Jiao Tong University documented an experiment that used approximately 5,000 examples produced by OpenAI's o1 model. With the help of GPT-4o mini, the team prepared 10,750 training examples in Chinese and English to improve Qwen2.5-72B, developed by Alibaba.
Although it makes it possible to develop specialized systems at lower cost, the method does not necessarily reproduce all the capabilities of the original model, nor does it eliminate the need for advanced computational infrastructure.
The practice has also become a source of diplomatic tension. According to Reuters, American authorities discussed with Chinese representatives in July possible intellectual property violations related to the collection of responses from American models.
Distillation is not inherently illegal, but it can contravene contractual terms. OpenAI, for example, restricts the use of responses from its systems to develop competing models, unless exceptions are provided for in its commercial contracts.
The investigation was published one day after Washington announced new restrictions on remote access to advanced chips installed in data centers in third-party countries, and on the eve of the Asia-Pacific Economic Cooperation (APEC) summit, scheduled for November 18 and 19, in Shenzhen.



