From 5,000 AI answers to a dispute over technological power
Shanghai Jiao Tong University researchers described in a November 2024 paper how they selected about 5,000 answers generated by OpenAI's o1 model, used GPT-4o mini to translate them and assembled 10,750 bilingual training samples. The material was then used to further train Alibaba's Qwen2.5-72B model.
Contents
- From 5,000 AI answers to a dispute over technological power
- How one model learns from another
- The research shows more than a single experiment
- Why Beijing wants cheaper models for industry
- What happened in the Chengdu talks?
- China's official response widened the argument
- Where the legal and commercial friction lies
- Chip controls are expanding beyond physical shipments
- The next dates and the limits of the strategy
- Key Points
That process, known as model distillation, sits at the centre of a growing dispute between Washington and Beijing. Chinese companies and researchers are using answers from leading American artificial intelligence systems to help train domestic models, potentially reducing development costs as the United States tightens access to advanced chips.
A review of more than 300 Chinese academic papers, patents, corporate disclosures, recruitment notices and government documents describes a shift from research experiments to wider industrial use. Companies have developed software to automate the collection of model outputs, while specialist workers prepare the resulting training material.
Washington regards distillation on an industrial scale as a major reason Chinese models have narrowed the performance gap with American competitors. People familiar with Chinese developers dispute that explanation, pointing to advances in engineering, algorithms, data and model design.
China's Ministry of Commerce has publicly rejected the American accusations and threatened countermeasures against restrictions targeting Chinese AI businesses. The disagreement now connects three questions: who may learn from an AI system's answers, how smaller models can become cheaper to run, and how much control chip export restrictions give Washington over AI development abroad.
How one model learns from another
Distillation uses a more capable model as a teacher and another model as a student. The teacher produces answers or examples of how to perform a task. Developers collect and select those examples, then use them to train the student.
A simple comparison is a student learning from worked examples rather than discovering every solution alone. In AI development, the examples might contain computer code, answers in several languages or plans for solving complicated problems.
The student is often smaller and cheaper to operate. It may learn useful skills without matching the teacher's full range of abilities. Distillation is widely used internationally and is not inherently unlawful, although particular uses can raise contractual or intellectual property disputes.
A former Alibaba engineer described the economic appeal: once an American model has answered a difficult question, a Chinese team can retain that example and repeatedly use it to train a domestic system. The expensive answer becomes reusable training material.
The research shows more than a single experiment
The Shanghai Jiao Tong example provides a concrete view of that process. Researchers selected roughly 5,000 o1 answers and expanded the material through translation into 10,750 bilingual samples. That is about 2.15 training samples for each selected answer, although the published figures are approximate.
Alibaba researchers have also described using OpenAI's GPT-4 and GPT-4o to produce multilingual training material. They filtered and scored the generated content before using it to train smaller Qwen models. Selection matters because a student's training depends on the quality and relevance of the examples it receives.
Research described in February 2026 by Shanghai Jiao Tong University and Meituan used GPT-4o to generate plans for programming tasks involving several AI agents. An agent is a software system that carries out tasks; several agents can divide a programming job into separate responsibilities.
Across these examples, the output being collected is broader than short chatbot answers. It includes language resources, programming material and plans for organising complex work. Automation allows developers to repeat the process at a scale that would be difficult to manage manually.
Why Beijing wants cheaper models for industry
Chinese government documents encourage distillation to reduce AI costs and make the technology easier to deploy across manufacturing. The documents described in the review promote the technique without directing companies to distil American models specifically.
For a manufacturer, a model with a narrower set of useful abilities may be more practical than a large general purpose system. A smaller model can cost less to run, making repeated use more affordable. The economic objective is to preserve enough capability for a particular job while reducing the resources needed each time the model operates.
China's industrial ambitions and its chip constraints therefore meet at the same point: efficiency. Distillation can help developers get more useful work from the computing resources they can obtain.
Advanced computing remains necessary. Distillation cannot reproduce every capability of a leading model or eliminate the need for powerful chips. Its value lies in transferring selected skills at lower cost, rather than replacing every part of the development process.
What happened in the Chengdu talks?
Three people familiar with discussions in Chengdu said US Commerce Department officials raised concerns with Chinese regulators in July 2026 about developers collecting outputs from leading American AI models. The exchange had not previously been made public.
American officials warned that some conduct could amount to intellectual property theft. A US official familiar with the talks said Chinese representatives defended distillation as a standard development technique and expressed willingness to act where conduct constituted intellectual property theft, without making specific commitments.
Chinese developers have also sought guidance from industry regulators in recent months. They want a clearer boundary between ordinary use of foreign models and organised collection of generated content for training.
The US Commerce Department, China's Cyberspace Administration and China's Ministry of Industry and Information Technology did not respond to requests for comment on the reported activity and discussions. No specific enforcement commitment emerged from the Chengdu exchange.
China's official response widened the argument
In a July 27 statement, China's Ministry of Commerce rejected threatened American investigations and sanctions. It argued that Chinese and American frontier models had appeared close together in time and said some Chinese models already led in certain capabilities.
The ministry also said many American AI businesses had distilled Chinese models. It cited nearly 200 US startups urging Washington to preserve access to Chinese models released with open access, arguing that cutting access would damage American competitiveness. Those claims form part of Beijing's official defence of its industry.
The dispute escalated in September. According to the ministry's account, the US National Security Agency, Cybersecurity and Infrastructure Security Agency and Federal Bureau of Investigation issued a joint advisory on September 8 accusing Chinese AI companies of distillation on an industrial scale and recommending defensive measures to American businesses.
The ministry's September 9 response called distillation a routine, neutral technique through which models learn from one another. It accused Washington of applying double standards and said China would take countermeasures if the United States used the issue to suppress Chinese AI companies.
Beijing also said the two countries' leaders had agreed to government dialogue on AI and offered professional discussions. Its public position combines a defence of cross border learning with a warning against sanctions.
Where the legal and commercial friction lies
OpenAI generally restricts users from employing its model outputs to develop competing models. Anthropic also restricts training competing systems without permission. Those conditions make the method of obtaining and using training material central to the commercial dispute.
A developer may view generated answers as useful examples, while the company operating the teacher model may regard their organised collection as an unauthorised route to building a competitor. Access conditions, permission and the intended use of outputs are therefore practical concerns for developers deciding how to assemble training data.
Washington's allegations concern particular conduct and its scale. Beijing's response argues that American companies also benefit from learning from Chinese systems and that restrictions on access protect established businesses from competition.
The disagreement also extends to why Chinese models have improved. A US official described distillation on a large scale as the main reason for the narrowing gap. Two people familiar with Chinese developers instead identified it as one contributor alongside engineering, algorithms, data and model design. Neither account provides a numerical breakdown of those contributions.
Chip controls are expanding beyond physical shipments
The United States is also preparing tighter controls on access to computing power outside its borders. The reported interim framework would require licences for exporting, reexporting or transferring AI chips and servers to data centres worldwide when they might supply computing resources to companies in China or other restricted countries.
The rules would cover data centres or networks capable of providing more than 12 million total processing performance, or TPP, a US export control measure of computing capacity. That threshold is described as roughly equivalent to the maximum computing power of 750 Nvidia H100 or H200 chips.
Indirect access through intermediaries and brokers would also be covered. Existing data centre operators supplying services within the rules would have 90 days after the rules take effect to obtain licences. The restricted D:5 country group includes China, Russia, Iran and North Korea.
Malaysia, Singapore, Indonesia, Thailand and Brazil would face additional licensing requirements for advanced chip shipments, with an exception for chips destined for US companies. The framework would replace the January 2025 AI Diffusion Rule and exclude commercial recommendation systems such as those used in online shopping and social media.
These controls target access to the machines that train and run AI. Distillation addresses a different part of the same development process: obtaining useful examples from a model that has already been trained. Cheaper training methods can reduce some pressure from chip restrictions, while developers still need computing resources to build and operate their own systems.
The next dates and the limits of the strategy
Researchers have warned that excessive reliance on distillation can discourage original development and leave teams dependent on breakthroughs made by the companies operating leading models. A cheap route to acquiring existing skills can still leave a developer waiting for someone else to produce the next advance.
For Chinese businesses, the immediate decisions concern training costs, access permissions and regulatory guidance. For American providers, the concern is that access to their systems may help finance and develop future competitors. Both countries also have companies that benefit from access to models developed abroad.
The main events and scheduled diplomatic opportunity are:
- November 2024: Shanghai Jiao Tong researchers described turning selected o1 answers into 10,750 bilingual samples for Qwen2.5-72B.
- July 2026: US officials raised concerns about collection of model outputs during talks in Chengdu.
- July 27, 2026: China's Ministry of Commerce publicly opposed threatened investigations and sanctions.
- September 8 and 9, 2026: A US security advisory was followed by China's formal rejection of its accusations.
- November 18 and 19, 2026: Shenzhen is scheduled to host the APEC summit, where President Donald Trump and President Xi Jinping are expected to meet.
The anticipated leaders' meeting gives the dispute a diplomatic setting, although no specific agreement on distillation has been announced. The central policy challenge is already concrete: governments are trying to regulate both access to powerful computers and the reuse of knowledge generated by the systems running on them.
Key Points
- Chinese researchers and companies are using outputs from American AI models to train domestic systems, with automation expanding the process.
- One documented project produced 10,750 bilingual training samples from roughly 5,000 selected OpenAI o1 answers.
- Distillation can lower costs and transfer selected abilities, but advanced computing remains necessary.
- US officials have raised intellectual property concerns; China defends the technique and threatens countermeasures against restrictions.
- Proposed US chip controls would also target remote access to computing resources in overseas data centres.
- Trump and Xi are expected to meet during the November 18 and 19 APEC summit in Shenzhen.






