Chinese Firms Train AI on American Model Outputs
Reuters review of more than 300 Chinese academic papers and records shows distillation of U.S. frontier AI models by Chinese companies.
Updated on Nov. 17, 2026 in Artificial Intelligence

Live Poll
Do you trust Chinese-built AI tools with your personal data?
Chinese tech firms used responses from frontier AI labs' models to train their own AI systems, according to a Reuters review of more than 300 Chinese academic papers, government files and other records. The Department of Commerce, meanwhile, released a rule requiring licenses to supply restricted advanced AI chips to overseas data centers.
Why it matters
The Department of Commerce warned Chinese regulators that some distillation of American models may be considered theft of intellectual property. OpenAI and Anthropic both restrict the use of their models' outputs to train outside AI, making the practice a contested issue between U.S. labs and Chinese developers.
Alibaba released its Qwen4-Max model on Sunday, and it ranked higher than Anthropic's Claude Fable 5.1 on Arena.ai's Code Arena leaderboard. Reuters' review of 300-plus Chinese academic papers and records does not quantify how much of Chinese training data derives from U.S. model outputs.
The players
Alibaba
Alibaba is a Chinese technology conglomerate whose Qwen model family and Alibaba Cloud unit are central to the distillation research described in the reporting.
OpenAI
OpenAI is the San Francisco-based AI developer whose GPT-4, GPT-4o and o1 models were used as training sources, and which restricts use of its outputs to train outside AI.
Anthropic
Anthropic is an AI safety company that develops the Claude model family and also restricts use of its models' outputs for training outside AI.
U.S. Department of Commerce
The U.S. Department of Commerce is the federal agency that issued the remote access rule for advanced semiconductors and warned Chinese regulators about distillation.
The details
Researchers at Shanghai Jiao Tong University and Meituan described in a February paper using OpenAI's GPT-4o to train Alibaba's small Qwen model for coordinating AI agent coding. Alibaba Cloud researchers published a paper in April 2025 describing distillation of OpenAI's GPT-4 and GPT-4o models.
Timeline
November 2024: Shanghai Jiao Tong researchers published a paper describing training an Alibaba model on OpenAI's o1 outputs.
April 2025: Alibaba Cloud published the DistilQwen2.5 paper.
February 2026: Shanghai Jiao Tong and Meituan published the AgentConductor paper.
July 2026: Commerce warned Chinese regulators about distillation during meetings.
Monday, November 16, 2026: Commerce released the remote chip access rule.
The Tech Race
The Department of Commerce's 'Addressing Remote Access to Advanced Semiconductors' rule requires licenses to supply restricted advanced chips to overseas data centers. Chinese firms' use of U.S. model outputs for training extends that competition from hardware access into the training data itself.
The Commerce license requirement affects companies that supply restricted advanced chips to overseas data centers, which can shape the availability and cost of AI services. Developers using frontier model APIs should note that OpenAI and Anthropic restrict use of their outputs to train outside AI.
The takeaway
Developers should review the terms of service for any AI model before using its outputs to train other systems, since frontier labs restrict this practice. Readers tracking the chip rule can follow Commerce's license requirements as they apply to overseas data center operators.
Further reading
Explore more reporting in our Artificial Intelligence section for context on how model training and chip export rules are developing.
Live Poll
Do you trust Chinese-built AI tools with your personal data?



