Chinese developers increasingly train AI on U.S. models’ answers, Reuters finds
The investigation traces research projects and commercial collection tools, sharpening a dispute over where legitimate training ends and misuse begins.
Loading page…
The investigation traces research projects and commercial collection tools, sharpening a dispute over where legitimate training ends and misuse begins.
Listen to this story
Reuters found that Chinese AI developers are increasingly turning answers from U.S. models into training data, with its investigation drawing on more than 300 papers, patents, company records and government documents. The approach can help transfer specific capabilities when advanced chips are harder to obtain, but it does not create an exact copy or eliminate the need for computing, expertise and original research. The findings also raise questions about whether outputs were collected and used with providers’ permission, as commercial data-gathering efforts expand.
Shanghai Jiao Tong University researchers used OpenAI o1 responses to produce thousands of training examples for Alibaba’s Qwen.
Research projects involving Alibaba and Meituan also explored training with GPT-4 and GPT-4o outputs.
Reuters found companies developing automated tools to collect model responses, alongside workers who prepare them for training.
Restrictions on advanced chips limit what Chinese AI developers can buy, but American models’ answers offer another resource for improving their systems. A Reuters investigation found increasingly widespread use of those outputs to train Chinese rivals—a practice that can transfer capabilities without reproducing all the work behind the original model.
The investigation examined more than 300 academic papers, patents, company records and government documents, according to Digital Trends’ account of Reuters’ findings. Its examples span university research and commercial efforts to gather AI-generated training material. The finding concerns growing adoption across China’s AI industry, rather than a single model or company.
The technique is called distillation: developers use responses from a more capable AI system to train another model. Those responses might contain code, answers or worked solutions to difficult problems. A smaller model can learn to perform similar tasks while requiring less computing power, making the approach attractive when access to advanced hardware is constrained.
The investigation points to concrete research examples involving American models and Chinese-developed systems. These illustrate how generated responses can become a dataset for further training, rather than serving only as answers to the person who requested them:
Beyond academic projects, Reuters found evidence of companies developing automated tools to collect AI-generated responses. It also identified specialized workers preparing that information for training. That commercial activity is a distinct part of the investigation: the process involves both obtaining outputs and doing the work needed to turn them into usable training material.
Distillation is not new or inherently illegal. The dispute concerns how developers obtain the material and whether they have permission to use it. OpenAI and Anthropic restrict the use of their models’ outputs to train competing systems without appropriate permission. A legitimate training method can therefore still involve uses that a model provider prohibits.
U.S. officials have warned that distillation activity crossing into intellectual-property theft could trigger sanctions. Chinese officials have defended distillation as a legitimate development technique, while acknowledging that actual intellectual-property violations warrant action. Those positions distinguish the training method itself from the disputed circumstances in which companies obtain and use another provider’s outputs.
Seeking Alpha’s summary of the same Reuters investigation places the activity within China’s effort to broaden its AI strategy and offset tighter U.S. chip-access restrictions. The hardware constraints help explain the appeal of learning from existing systems, though distillation still leaves developers with substantial computing and research work.
China’s government is also encouraging broader use of distillation to make AI cheaper and easier to deploy, particularly in manufacturing. Its role is not limited to developing rival general-purpose models; officials are promoting it as a way to put AI into industrial use.
Loading discussion...
Join the conversation
Explain where you draw the line between fair competition and misuse.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.