Source-reported · Independent verification status is not recorded for this article; claims remain attributed to their sources.
China’s use of U.S. AI outputs to train domestic models tests Washington’s chip export strategy
Chinese researchers are increasingly training cheaper domestic AI systems on the outputs of U.S. models such as those from OpenAI, using a technique known as distillation that Washington worries could amount to intellectual property misuse while Beijing defends it as a standard development tool.
Chinese researchers are leaning on American artificial intelligence systems to improve their own models, a trend that could soften the impact of U.S. chip export restrictions aimed at slowing China’s AI progress.
According to the available reporting, Chinese developers are feeding outputs from U.S. AI models, including products from OpenAI and other American firms, into domestic systems. They are doing this through a method known in the field as "distillation," where a powerful model’s answers are used as training data for a smaller or less capable one.
Distillation is a familiar machine-learning technique. It allows engineers to copy much of a stronger model’s behavior into a lighter system that is cheaper to run and can work on less advanced hardware. In this case, it means Chinese labs don’t need full access to U.S. model internals to benefit from them; they only need the ability to query the models and store the outputs.
Washington has raised concerns that this practice may cross into intellectual property theft. The fear is that if Chinese models can closely match U.S. systems by consuming large volumes of their generated content, American firms will carry the cost of frontier research while others extract much of the value by imitating the behavior.
Beijing, in contrast, defends the approach as normal AI development. Chinese officials and researchers argue that distillation is a standard tool for building and refining models and that using outputs from other systems belongs within that accepted toolkit.
These opposing views collide with U.S. export controls on advanced chips. Current restrictions focus on limiting China’s access to high-end semiconductors and related hardware needed to train the largest AI models. Distillation gives Chinese actors a partial workaround: they can run training queries against U.S. models hosted abroad, then use the results to train smaller domestic models that operate on less powerful Chinese chips.
That doesn’t remove China’s reliance on advanced hardware altogether, but it does reduce how decisive U.S. control over cutting-edge chips may be in shaping China’s AI capabilities.
The practice also raises questions for U.S. companies about how to manage access to their systems. If large volumes of outputs can be harvested and repurposed, firms may need to adjust usage limits, pricing or technical protections to guard against large-scale distillation by competitors.
The next clear signals will likely come from policy and corporate responses: any moves by Washington to define when training on another company’s AI outputs constitutes intellectual property infringement, and any changes by major U.S. model providers to how their systems can be accessed from China or used to generate bulk training data for rival models.