China copies OpenAI’s AI: hundreds of documents reveal the trail of ‘AgentConductor’ on Alibaba
More than 300 Chinese documents reveal how Alibaba trained their models with responses from OpenAI. The technique is called distillation and the Chinese government defends it as ordinary commercial practice.
Distillation – training a small model with the responses of a larger, more expensive and more capable model – is practiced all over the world, including in Europe. The problem is not the technique. The problem is that it allows a Chinese laboratory to approach the technological frontier without buying the same number of chips that US export controls deny it. Washington’s competitive advantage was built on silicon; that of Beijing, on someone else’s text.
What exactly does the ‘AgentConductor’ paper trail say?
He paper The most revealing was published in February of this year and is authored by researchers from Shanghai Jiao Tong University and Meituan. It’s titled ‘AgentConductor’ and describes how they used GPT-4o to train a small version of qwen and teach you how to coordinate code agents. The American model acted as a teacher for the Chinese model that competes with it today.
It is not an isolated case. In April 2025, researchers at Alibaba Cloud published ‘DistilQwen2.5’, detailing the use of GPT-4 and GPT-4o to collect non-Chinese training data for their models. In November 2024, the same Shanghai team signed ‘O1 Replication Journey – Part 2’, which explains how they trained an Alibaba model with o1 outputs. Three jobs, three different generations and a single method.
The review that supports the investigation is not a leak: it is part of the analysis of more than 300 papers Chinese scholars, government records, and other records. The pattern repeats.
Washington does not fight this battle with tariffs or aircraft carriers. The pound with export licenses and with the fine print of the use contracts.
OpenAI and Anthropic They expressly prohibit in their terms that the outputs of their models are used to train foreign systems, a detail that went unnoticed in almost all European coverage. The Chinese Embassy in Washington sees it differently: ‘The US accusations are baseless and politicize a normal technological and commercial practice,’ a spokesperson said. The development of AI in China, he added, comes from greater scientific self-reliance and open cooperation.
Why Qwen4-Max worries more than any paper
On Sunday, November 15, Alibaba released Qwen4-Max, its most advanced and closed-weight model. In Code Arena, Arena.ai’s ranking focused on web programming, it came in sixth place with 1,750 preliminary points: 74 more than its predecessor, Qwen3.8-Max (1,676), and ahead of Anthropic’s Claude Fable 5.1. It is the best-placed Chinese model in HTML, brand marketing and design with references.
The next day, the Department of Commerce published the rule on remote access to semiconductors, which requires obtaining a license to supply restricted chips to data centers abroad and closes the telematic route. In July, according to three anonymous sources cited by Reuters, Commerce officials had already warned Chinese regulators that some of that distillation could be considered theft of intellectual property.
Not everyone in Washington shares the diagnosis. The US government should be careful not to assume that restricting China’s access to technology automatically strengthens America’s competitive position,’ said Daniel Castro, president of the Information Technology and Innovation Foundation. Jason Hsu of the Hudson Institute shifted the focus: ‘U.S. companies that build on Chinese models may face intellectual property vulnerabilities, software supply chain risks, and long-term technological dependency.’ The debate is no longer whether China copies, but how much the West depends on what China publishes.
Washington Logic
What Washington is trying to solve is not a moral problem, but one of return on capital. Training a frontier model costs hundreds of millions; If a competitor obtains a good part of the performance with the first’s responses, the American investment loses economic meaning. The coalition that supports this policy is transversal: national security hawks from the Republican Party and industrial policy Democrats voted together on 2022 chip restrictions.
The precedent is the semiconductor agreement with Japan in 1986, which forced Tokyo to stop selling below cost and open its market. The Japanese industry did not disappear then: it lost leadership a decade later, when Korea and Taiwan arrived. The controls brake; They don’t stop. And the list of entities that blocked Huawei in 2019 serves as a reminder: Beijing learned to build alternatives under pressure.
In Spain the blow does not come through tariffs. It arrives through the supplier’s. A good part of the open weight models that circulate in the repositories used by Spanish startups and development teams come from Chinese laboratories, and no one audits what American answers they contain. The European AI regulation requires large providers to publish a summary of their training data, but not to reconstruct the provenance of each output. Spanish companies that integrate a third-party model in customer service or credit risk analysis inherit that risk without seeing it. Germany and France have made model sovereignty an industrial priority; Madrid has less margin and its ability to influence depends on Brussels.
The next window is not a vote: it is the application phase of the November 16 rule. Each license that Commerce grants or denies to a foreign data center will set the pace. In parallel, the Department could expand the list of entities if new jobs with their own name appear. What we observe is that the result will not be decided in the courts, but in accounting: if distilling is cheaper than buying silicon, the Chinese advantage survives any rule.
Case File
- The case: More than 300 Chinese documents – papers, files and records – show how Alibaba trained its Qwen models with OpenAI outputs under names such as ‘AgentConductor’ and ‘DistilQwen2.5’.
- Key facts: Qwen4-Max entered Code Arena on November 15th in sixth place with 1,750 points. The Commerce Department published on November 16 the rule requiring a license to provide remote access to restricted chips.
- For Spain: Without a tariff involved, the risk is of dependency: Chinese models integrated into Spanish services without traceability of the data that trained them.
