Chinese AI Developers Train on US Model Outputs at Industrial Scale, Reuters Finds

A review of more than 300 papers, patents and filings traces distillation from US models into work at Alibaba, Meituan and Shanghai Jiao Tong University.

WHY IT MATTERS

An MIIT-led factory guide endorses distillation while US officials warn some of it could be IP theft, turning Chinese labs' use of US model outputs into a trade-talks issue.

Chinese AI developers have moved distillation, the practice of training a model on another model’s answers, from lab experiments into routine industrial use, and US models are among the main teachers. That is the finding of a Reuters investigation published on November 17, which reviewed more than 300 Chinese academic papers, patents, corporate disclosures, recruitment notices and government documents.

Reuters names work at Shanghai Jiao Tong University, Alibaba and Meituan, and ties the trend to chip export controls: a former Alibaba engineer, speaking anonymously, said one costly query to a strong model can feed repeated training runs when computing power is scarce. Shanghai Jiao Tong University, Alibaba, Meituan and the Ministry of Industry and Information Technology (MIIT) did not respond to Reuters’ requests for comment.

Distillation is legal and widely used outside China, and Reuters notes it cannot fully replicate a leading model or remove the need for advanced chips. The dispute is over scale and permission. OpenAI’s and Anthropic’s terms restrict using their output to build competing models, with some exceptions.

Papers show US models supplying training data for Qwen and agent systems

The earliest case Reuters cites is a November 2024 Shanghai Jiao Tong University study. The paper, posted on arXiv by researchers from the university, the GAIR lab, SII and New York University, describes selecting "approximately 5,000 high-quality samples" from OpenAI’s o1, translating them into Chinese with GPT-4o mini, and fine-tuning Qwen2.5-72B-Instruct on the result. Shanghai Jiao Tong University researchers built 10,750 Chinese-English training pairs from about 5,000 selected OpenAI o1 answers to further train Alibaba’s Qwen2.5-72B. The authors themselves warned about "allowing the convenience of distillation to divert us" from first-principles research, and said their distillation of o1 was for research purposes only.

Alibaba’s own researchers used GPT-4 and GPT-4o in April 2025 to produce training material in languages other than Chinese, then refined and scored the responses to train smaller Qwen models, Reuters found. In February 2026, a team from Shanghai Jiao Tong University and Meituan used GPT-4o to write plans for coordinating AI agents on programming problems, which helped train a smaller model to organize that kind of work.

Washington has been building its case for months. A joint report by the NSA, FBI and CISA in September accused DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI of large-scale distillation from US models, Caixin Global reported. A Chinese IP compliance specialist told Caixin that neither US nor Chinese law clearly defines distillation. EastFrontier earlier covered OpenAI’s account of a distillation campaign linked to Moonshot.

Chinese industrial policy endorses the technique without naming its source

Beijing does not tell companies to distill American models, Reuters says, but it encourages the method. The clearest example is the AI+Manufacturing policy (《“人工智能+制造”专项行动实施意见》) issued in January by eight agencies. Xinhua reported that MIIT, the Cyberspace Administration of China and the National Development and Reform Commission were among the issuers, and the targets run to 2027.

The policy’s full text, document number 工信部联科〔2025〕279号 and dated December 25, 2025, carries an annex for factories, the AI Application Guide for Manufacturing Enterprises (《制造业企业人工智能应用指南》). Its eleventh item encourages companies to coordinate cloud, edge and device computing, with the cloud side handling model training, fine-tuning, quantization and distillation. Reuters adds that Shanghai and Shenzhen have also promoted techniques that cut computing needs.

That policy framing collides with the US position. A US official told Reuters that Washington sees industrial-scale distillation as the main reason Chinese models have narrowed the gap with American ones. Two people familiar with Chinese developers’ operations disputed that, saying engineering, algorithms, data and model design matter as much.

The issue moves from terms of service to trade talks

US Commerce Department officials raised the practice with Chinese regulators at meetings in Chengdu in July, warning that some collection of model output could amount to intellectual property theft, according to three people familiar with the talks cited by Reuters. Chinese officials defended distillation as a standard technique and said they were open to acting where it amounted to IP theft, but made no concrete commitments, a US official said.

Chinese developers have themselves sought clarity. In recent months they asked Beijing’s industry regulator where ordinary use of foreign models ends and organized collection of output for training begins, one person familiar with the discussions told Reuters.

The timing adds pressure. Reuters published a day after Washington said it would further restrict China’s access to advanced chips, and ahead of the Asia-Pacific Economic Cooperation summit in Shenzhen on November 18 and 19, where Donald Trump and Xi Jinping are expected to meet. Any written guidance from Beijing defining acceptable use of foreign model output would be the first official line on a practice its own factory policy already promotes.