Market
Reuters review of more than 300 Chinese documents says distillation of US AI model outputs has turned industrial
Reuters reviewed 300+ Chinese papers, patents and government documents. They show distillation is now industrial; Beijing encourages it, but not of US models.
A Reuters review of more than 300 Chinese academic papers, patents, corporate disclosures, recruitment notices and government documents says distillation, training smaller models on a stronger model’s answers, has moved from research into mainstream industrial use in China. Reuters also reports that Chinese government documents encourage distillation, but do not instruct companies to distil American models.
The essentials
- The finding: the records describe pipelines that use US model answers to build training data for domestic models, with software to automate it and workers preparing the material.
- Beijing’s role: a January AI+Manufacturing package from eight agencies supports training, fine-tuning and distillation for industrial use. It does not tell firms to use US models.
- An example: a November 2024 Shanghai Jiao Tong study turned about 5,000 OpenAI o1 answers into 10,750 bilingual examples for Alibaba’s Qwen2.5-72B.
- The politics: US officials accused Chinese firms in September of industrial-scale distillation, and the issue is a friction point before the November 18–19 APEC meeting.
Legal, but against the terms
Distillation itself is legal, and Chinese officials defended it as standard practice at July talks in Chengdu. OpenAI and Anthropic forbid rivals from training on their outputs. A US Treasury official warned that covert, industrial-scale attacks that amount to IP theft could bring sanctions and Entity List designations.
What the review does not show
It shows that distillation is widespread and encouraged, not that any named company broke a provider’s terms. We read the Reuters piece only as a syndicated excerpt and secondary summaries.
What is not established
- Which Chinese firms distilled which US models: not established in what we read.
- The volume of US model output used: not quantified.
- Whether Chinese authorities directed use of US models: Reuters says documents do not.
- Any sanctions: only warned of.
- OpenAI’s or Anthropic’s response to this report: not found.
Summary table
| Point | What is stated as of November 17, 2026 | Source |
|---|---|---|
| Documents reviewed | More than 300 | Freedom 96.9 (Reuters) |
| Government policy cited | January AI+Manufacturing package, eight agencies | Technology.org |
| Example | 5,000 o1 answers became 10,750 Qwen2.5-72B examples | Technology.org |
| US accusation | September, industrial-scale distillation | Dataconomy |
| Reuters report date | November 17, 2026 | FourWeekMBA |
FAQ
Is Beijing telling companies to copy US models?
Not according to Reuters. Documents encourage distillation in general but do not instruct companies to distil American models.
Is distillation illegal?
Distillation is legal in itself. OpenAI and Anthropic ban rivals from training on their output in their terms.
Does this name companies that broke the rules?
Not in what we read. It describes a widespread practice, with some research examples.
Method and limits
Checked on November 17, 2026. We read a syndicated Reuters excerpt, Dataconomy, Technology.org and FourWeekMBA through search excerpts. We did not read the full Reuters article.
Sources
- Freedom 96.9 (Reuters): Exclusive-China taps US AI models to train homegrown rivals as chip curbs bite (verified 2026-11-17)
- Dataconomy: China Uses US AI Model Outputs To Train Domestic Rivals (verified 2026-11-17)
- Technology.org: China Uses US AI Model Outputs to Train Domestic Rivals, Documents Show (verified 2026-11-17)
- FourWeekMBA: Chinese AI Labs Train on US Model Output, 300+ Files Show (verified 2026-11-17)
Spotted an error in this article? Report it · Our corrections policy