China takes a page from US AI playbook to build cheaper domestic rivals as chip curbs bite: Report

What happened
Chinese companies and researchers are increasingly using a method called distillation to train domestic AI models with outputs from leading US AI systems. This approach is being adopted as US restrictions on advanced chip exports to China tighten. Chinese authorities are encouraging the use of distillation to make AI more affordable and accessible for the country's manufacturing sector. The US government has raised concerns about potential intellectual property violations related to this practice ahead of a major summit between US and Chinese leaders.
Key details
- Distillation uses answers and outputs from US AI models like OpenAI's GPT-4 to train smaller Chinese models.
- A November 2024 paper from Shanghai Jiao Tong University described using about 5,000 OpenAI model answers to train Alibaba's Qwen2.5-72B model.
- US companies such as OpenAI and Anthropic restrict use of their model outputs for training competing systems.
- A national AI+Manufacturing policy package in China, published in January, supports distillation for industrial applications.
- US officials have accused Chinese companies of large-scale distillation and raised the issue with Chinese regulators in July 2025.
Why it matters
Chinese technology firms and researchers can now develop domestic AI models that are less reliant on US technology, even as chip export controls become stricter. US companies face a risk of their AI outputs being used to train competitors, potentially undermining the effectiveness of export restrictions. Chinese manufacturers and AI developers gain access to cheaper, smaller AI systems, while US and Chinese authorities must address disputes over intellectual property and the boundaries of acceptable use.
