Chinese AI developers are increasingly using answers generated by American artificial-intelligence models to train smaller domestic systems, according to a Reuters investigation published on November 17, 2026. The practice, called knowledge distillation, is drawing geopolitical scrutiny as Washington tightens restrictions on advanced chips and the two countries prepare for talks at the APEC summit in Shenzhen.
What the documents show
Reuters reviewed more than 300 Chinese academic papers, patents, company disclosures, recruitment advertisements and government documents. It found examples of researchers using powerful US models to generate solutions to coding, reasoning and language tasks, then turning those outputs into training data for less expensive models. The records also describe tools for automating data generation and workers preparing material for model training.
The investigation describes a broader trend, not proof that every Chinese AI model depends on American outputs. Some research is publicly disclosed and academic; other activity may be commercially sensitive. The distinction matters because distillation itself is a widely used machine-learning method and is not automatically unlawful.
How knowledge distillation works
In a typical arrangement, a large teacher model produces answers or examples, and a smaller student model learns from them. Rather than repeatedly paying to consult the teacher, developers can use a curated dataset to improve a student model's performance on selected tasks. The resulting system may be cheaper to run and easier to deploy in businesses with limited computing budgets.
But distillation is not a shortcut to reproducing every capability of a frontier model. Students can inherit weaknesses or errors, and researchers still need their own datasets, engineering, evaluation and computing resources. It is also different from copying a model's hidden weights or obtaining its source code.
Specific examples in the investigation
Reuters cited a November 2024 project by Shanghai Jiao Tong University researchers that selected approximately 5,000 answers from OpenAI's o1 model, translated material using GPT-4o mini and assembled 10,750 bilingual examples for additional training of Alibaba's Qwen2.5-72B. The figures describe the researchers' disclosed process; they are not a measurement of all Chinese distillation activity.
An April 2025 Alibaba research paper described generating multilingual training examples with GPT-4 and GPT-4o, refining responses and using them in smaller Qwen systems. Another paper, from February 2026, described Shanghai Jiao Tong University and Meituan researchers using GPT-4o to help create examples for coordinating coding agents.
These examples illustrate why model outputs can become an economically valuable training resource. A developer may not need to reproduce a frontier model's entire training run to acquire useful examples of planning, translation or reasoning.
Why the US and China disagree
Reuters reported that US Commerce Department officials raised distillation concerns with Chinese regulators at meetings in Chengdu in July. According to people familiar with the discussions, American officials warned that some large-scale collection could cross into intellectual-property theft. Chinese officials defended distillation as a normal technical practice while indicating that actual IP theft could warrant action. Reuters said no concrete commitment emerged.
US Treasury Secretary Scott Bessent separately warned in July that covert industrial-scale activity crossing the line into IP theft could lead to sanctions or trade restrictions. Those are government warnings, not judicial findings that the research projects cited by Reuters were crimes. Chinese regulators did not respond to Reuters' requests for comment on the November investigation.
OpenAI's published terms prohibit using its service outputs to build competing models, and its rules also restrict automated extraction. A breach of contractual terms, however, is not identical to a criminal offence. Legal assessment depends on access conditions, jurisdiction, conduct and evidence. The broader research technique can be legitimate when a developer owns or has permission to use the teacher model.
Chip restrictions add urgency
Washington announced further controls on access to advanced US computing technology this week. Limits on high-end accelerators increase the appeal of approaches that make efficient use of available hardware. Distillation can help deploy selected AI capabilities more cheaply, but cannot eliminate the need for advanced chips in ambitious training programmes.
Reuters also reported a disagreement over how much distillation explains China's narrowing AI performance gap. One US official attributed much of the progress to it, while people familiar with Chinese developers argued that algorithms, engineering, data and model design are also important. Neither explanation should be presented as a proven single cause.
Beijing's industrial AI strategy
Chinese government documents encourage distillation and related efficiency techniques for manufacturing and other industries, Reuters found. A January AI-and-manufacturing policy package from eight agencies supported model training, fine-tuning and distillation. Reuters said the documents did not direct firms to distil American models specifically.
The policy objective is practical: smaller systems can be deployed for factory workflows, coding assistance, logistics and other specialised tasks without always relying on the largest available model. That could widen adoption while keeping computing costs manageable.
What happens next
The dispute raises questions for model providers, researchers and regulators. Providers must decide how to detect and enforce prohibited extraction without disrupting legitimate customers. Developers need clarity on which uses of outputs are permitted. Governments must distinguish standard technical practice from documented misconduct, rather than treating all distillation as theft.
Bottom line: Reuters' document review shows widespread and increasingly organised use of US model outputs in Chinese AI research and development. It does not establish that all such activity is illegal or that distillation alone explains China's progress. The issue is becoming a significant part of the wider US-China contest over chips, intellectual property and AI deployment.




