Skip to main content
Live
Main content

Chinese labs train homegrown AI models on OpenAI outputs, documents show

More than 300 papers, patents and government records show distillation moving into mainstream use in China as US chip curbs tighten.

Jaeden Schafer
Editor in Chief · · 5 min read
OpenAI logo

Chinese research teams and companies are using answers generated by American AI models, including OpenAI's o1, GPT-4 and GPT-4o, to train domestic rivals, according to more than 300 Chinese academic papers, patents, corporate disclosures, recruitment notices and government documents reviewed in an investigation published November 16. The records show distillation moving from academic experiment into mainstream industrial practice in China as US chip curbs tighten.

Distillation uses a powerful "teacher" model to generate answers or demonstrate tasks, then trains a smaller "student" system on that output. The documents describe Chinese researchers using US models to answer questions, write code and solve difficult problems. Companies are building software to automate the process, and specialist workers prepare the material used to train smaller domestic models.

A November 2024 paper from Shanghai Jiao Tong University researchers selected about 5,000 answers from OpenAI's o1 model, translated them with GPT-4o mini and assembled 10,750 bilingual training examples to further train Alibaba's Qwen2.5-72B. The same researchers cautioned that excessive reliance on distillation could stifle original innovation and leave developers dependent on more advanced model providers.

Key facts

  • 01More than 300 Chinese papers, patents, disclosures, job notices and government documents describe training local models on US model output.
  • 02A November 2024 Shanghai Jiao Tong paper turned about 5,000 OpenAI o1 answers into 10,750 bilingual examples to train Alibaba's Qwen2.5-72B.
  • 03Alibaba researchers used GPT-4 and GPT-4o in an April 2025 paper to generate non-Chinese training data for smaller Qwen models.
  • 04China's January AI+Manufacturing package from eight agencies supports training, fine-tuning and distillation for industrial uses.
  • 05US Commerce officials raised IP-theft concerns with Chinese regulators in Chengdu in July; APEC meets in Shenzhen November 18-19.

Alibaba's own researchers did something similar in an April 2025 paper, using GPT-4 and GPT-4o to generate training material in languages other than Chinese, then refining and scoring the responses before training smaller Qwen models. A February 2026 paper from Shanghai Jiao Tong and food-delivery company Meituan used GPT-4o to produce plans for coordinating AI agents on programming problems, which helped train a smaller model to organize those tasks. Shanghai Jiao Tong, Alibaba and Meituan did not respond to requests for comment.

Distillation is legal and widely used outside China. It cannot reproduce a leading model's full capabilities or remove the need for advanced computing power, but it can preserve selected skills in smaller, cheaper systems. For Chinese developers, that offers a way to reduce reliance on American technology, and potentially to blunt the effect of US limits on advanced chip sales.

Beijing's policy documents do not instruct companies to distill American models, but they do encourage the technique. A national AI+Manufacturing policy package published in January by eight government agencies includes guidance supporting model training, fine-tuning and distillation for industrial applications. Shanghai and Shenzhen have also encouraged methods that cut computing requirements, in line with Beijing's push to deploy smaller, cheaper AI systems across the economy.

Chinese AI developers have asked Beijing's industry regulator in recent months where the line sits between ordinary use of foreign models and organized collection of their output for training. China's Ministry of Industry and Information Technology did not respond to a request for comment.

OpenAI's terms restrict using its output to develop competing models, subject to exceptions, and Anthropic similarly restricts training competing systems without permission. US authorities publicly accused Chinese companies in September of aggressive, targeted distillation of American AI models on an industrial scale.

Related · from this week
Chinese AI models Kimi K3 and Qwen3.8 close the gap on OpenAI and Anthropic
Jaeden Schafer · 5 min read →

The dispute has already reached regulators. At meetings in Chengdu in July, US Commerce Department officials warned Chinese regulators that some developer activity could amount to intellectual-property theft. Chinese officials defended distillation as a standard AI-development technique and said they were open to acting where conduct amounted to IP theft, but made no concrete commitments.

The two sides also disagree on how much distillation matters. The US view, as described in the investigation, is that industrial-scale distillation is the main reason Chinese models have narrowed the gap with American rivals. That assessment is disputed on the Chinese side, where distillation is described as one factor alongside engineering, algorithms, data and model design.

The findings land a day before the Asia-Pacific Economic Cooperation summit in Shenzhen on November 18-19, and a day after Washington said it would further restrict China's access to advanced chips from companies such as Nvidia. The US Commerce Department and China's Cyberspace Administration did not respond to requests for comment.

ShareXLinkedInEmail
AI Box

Every AI model. One chat.

The latest models from ChatGPT, Claude, Gemini, Sora, ElevenLabs — 80+ models in a single chat. Compare answers side by side. Pick the best one every time.

  • ChatGPT, Claude, Gemini, Grok, DeepSeek — in one chat
  • Generate images & video with Sora, Veo, Ideogram
  • Compare any two models side by side
  • From $8.99/mo · 80+ models, all included
Try AI Boxaibox.ai
Trusted by 3,000+ teams
Got a tip?

Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.

Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.

AI Box Daily briefingFree · Daily · No fluff

Stay ahead of everyone in AI.

The tightly edited AI news email engineers, founders, and investors actually open. One email. Every weekday. Five minutes to finish.

Loved by 10,000+ AI professionals
Free forever. Unsubscribe with one click.

The briefing read inside teams at

Keep reading

More from Security

Chinese AI models Kimi K3 and Qwen3.8 close the gap on OpenAI and Anthropic
Analysis

Chinese AI models Kimi K3 and Qwen3.8 close the gap on OpenAI and Anthropic

Six of OpenRouter's top 10 tools are already Chinese. The latest releases from Moonshot and Alibaba shouldn't be shocking anyone.

Jaeden Schafer5 min read
Cohere says Alibaba's Qwen3.8-Flash-Next tried to upload a file to a Beijing cloud bucket
Security

Cohere says Alibaba's Qwen3.8-Flash-Next tried to upload a file to a Beijing cloud bucket

During a sandboxed security test, the open-weight model probed its host and reached for a cloud metadata address despite planning no tool calls.

Jaeden Schafer5 min read
US names six Chinese AI firms in industrial-scale distillation campaign
Security

US names six Chinese AI firms in industrial-scale distillation campaign

The NSA, CISA and FBI accuse DeepSeek, Alibaba, Moonshot, MiniMax, StepFun and Z.AI of siphoning capabilities from Claude, GPT, Gemini and Grok.

Jaeden Schafer5 min read