Tuesday, Nov 17 | --:--
Back to home

Reuters: China Has Moved Distillation of US AI Models From Labs Into Industry

Reuters reported on November 17 that its review of more than 300 Chinese academic papers, patents, corporate disclosures, recruitment notices and government documents shows distillation, training smaller models on a stronger model's output, has spread from academic experiments into mainstream industrial use, with companies automating the collection of output from US models such as OpenAI's o1, GPT-4 and GPT-4o. Government documents encourage distillation for industry but do not tell companies to distil US models. Distillation is not illegal; a US official says it is the main reason Chinese models have closed the gap, which two people familiar with Chinese developers dispute.

4 min read Shenzhen View as Markdown
A handmade editorial collage on torn ivory paper in sage with a thin oxblood accent: a large unadorned clay vessel stands beside a low wooden bench of many small clay copies and a wooden mould with a dark oxblood stain in its well; an official OpenAI logo is glued on a torn-paper clipping above the bench, the full wordmark visible.

Washington's chip curbs restrict the hardware; they do nothing about the answers. A Reuters review of more than 300 Chinese academic papers, patents, corporate disclosures, recruitment notices and government documents, published on November 17, describes distillation of American models' output moving from academic experiments into mainstream industrial use in China. If that is right, the leak in the export-control regime is not a smuggled GPU but an API response, and it arrives a day after Commerce tightened the rules on overseas data centres.

What the documents show#

Distillation uses a strong "teacher" model to generate answers or demonstrate tasks, then trains a smaller "student" on them. Reuters says the records describe Chinese researchers using US models to answer questions, write code and solve hard problems, companies building software to automate the collection, and specialist workers preparing the material for training. It cites three papers:

  • November 2024: Shanghai Jiao Tong University researchers selected about 5,000 answers from OpenAI's o1, translated them with GPT-4o mini and built 10,750 bilingual examples to further train Alibaba's Qwen2.5-72B.
  • April 2025: Alibaba researchers used GPT-4 and GPT-4o to generate training material in languages other than Chinese, then scored and refined it before training smaller Qwen models.
  • February 2026: Shanghai Jiao Tong University and Meituan researchers used GPT-4o to generate plans for coordinating agents on programming problems, training a smaller model to organise such tasks.

Government documents do not instruct companies to distil American models, Reuters says, but encourage distillation to spread cheaper AI across manufacturing: a national AI+Manufacturing policy package from eight agencies in January supports training, fine-tuning and distillation for industrial uses. Shanghai Jiao Tong University, Alibaba and Meituan did not respond to Reuters.

The dispute#

Distillation is not illegal and is common elsewhere, Reuters notes, and it cannot reproduce a leading model's full capability. OpenAI's terms restrict using its output to build competing models, subject to exceptions, and Anthropic's restrict training competing systems without permission. OpenAI has previously attributed one extraction campaign to people tied to Moonshot AI, and Anthropic says its defences cut distillers' harvest by 90 to 99%.

A US official told Reuters that Washington believes industrial-scale distillation is the main reason Chinese models have narrowed the gap; two people familiar with Chinese developers said it is one factor alongside engineering, algorithms, data and model design. Reuters reports, citing three people, that Commerce officials raised the issue with Chinese regulators in Chengdu in July and warned some activity could amount to intellectual-property theft; Chinese officials defended distillation as standard practice and made no concrete commitments, a US official said. Treasury Secretary Scott Bessent wrote on X in July that "sanctions and Entity List designations will be on the table" for covert, industrial-scale distillation that crosses into IP theft.

Limits#

  • The examples Reuters cites are published research; they show the method in use, not how much of any commercial Chinese model was trained this way.
  • The claim that distillation explains most of the gap is a US official's view, disputed by people familiar with Chinese developers; neither side offered data in Reuters' report.
  • Whether a given use breaches a provider's terms depends on facts Reuters does not establish. China's Cyberspace Administration and its industry ministry did not respond.
  • Chip Dispatch's account is based on Reuters' report, so it is not independent confirmation; we could not open Reuters' own site.

Sources#

Prior Coverage

Earlier Times of AI reporting on this thread.

Scroll to continue reading