America’s top artificial intelligence models are training their Chinese competitors, whether the leading labs like it or not.
asserted
labs → train → it
Chinese tech firms are using frontier AI labs’ model responses to train their own systems, according to a Reuters review of more than 300 Chinese academic papers, government files and other records.
uncertain
firms → use → papers
The practice of distillation is conducted worldwide to train AI models, but the process may allow China to compete in the AI race without accessing as many chips restricted through export controls.
uncertain
China → conduct → controls
One February research paper from Shanghai Jiao Tong University and Meituan researchers, “AgentConductor,” discussed its process for using OpenAI’s GPT-4o to train Alibaba’s small Qwen model to coordinate AI agent coding.
asserted
paper → discuss → coding
(RELATED: Alibaba’s New AI May Run Circles Around America’s Best)
uncertain
AI → run → America
“The U.S. allegations are unfounded and politicize a normal technological and commercial practice,” a spokesperson for the Chinese Embassy in Washington, D.C., told the Daily Caller News Foundation when asked about model distillation.
asserted
spokesperson → politicize → distillation
“The development of AI in China comes from greater self-reliance and strength in science and technology, and is fueled by open cooperation.”
asserted
development → come → cooperation
“We hope the U.S. side will stop leveling false allegations to smear Chinese companies,” the spokesperson added.
asserted
spokesperson → hope → companies
Similarly, researchers from Alibaba Cloud published a paper called “DistilQwen2.5” in April 2025, which outlined the distillation of OpenAI’s GPT-4 and GPT-4o models to collect non-Chinese training data for Alibaba’s Qwen models.
asserted
which → publish → models
OpenAI and Alibaba did not immediately respond to the DCNF’s requests for comment.
asserted
OpenAI → respond → comment
Qwen4-Max by @Alibaba_Qwen has landed as #6 in the Code Arena: WebDev!
asserted
Max → land → Arena
With 1750 pts (preliminary), it’s +74 pts over its predecessor, Qwen3.8-Max (0902) at 1676.
asserted
it → ’ → 1676
It’s also the top Chinese model in: - #6 HTML: 1748 pts - #8 Brand & Marketing: 1705 pts - #8 Reference-Based Design: 1770 pts Congrats to the @Alibaba_Qwen team on this release!
asserted
It → ’ → release
— Arena.ai (@arena) November 16, 2026
Another paper from Nov. 2024, “O1 Replication Journey – Part 2,” describes that researchers at Shanghai Jiao Tong University trained an Alibaba model on OpenAI’s o1 model outputs.
asserted
researchers → describe → outputs
The Department of Commerce warned Chinese regulators during July meetings that some distillation of American models may be considered theft of intellectual property, three anonymous sources told Reuters.
uncertain
sources → warn → Reuters
The department referred the DCNF to its rule released Monday to combat attempts to remotely gain access to advanced AI chips.
asserted
department → refer → chips
The rule, “Addressing Remote Access to Advanced Semiconductors,” requires companies to acquire a license from Commerce to supply restricted chips to data centers overseas.
asserted
rule → address → centers
Tight chip restrictions cannot ensure Chinese AI is weakened, Daniel Castro, president of the Information Technology and Innovation Foundation, told the DCNF Monday.
asserted
Castro → ensure → DCNF
“The U.S. government should be careful not to assume that restricting China’s access to technology automatically strengthens America’s competitive position,” Castro said.
asserted
Castro → assume → position
Alibaba released its most advanced AI model to date, Qwen4-Max, on Sunday.
asserted
Alibaba → release → Sunday
The reportedly closed-weight model ranked higher than Anthropic’s Claude Fable 5.1 on Arena.ai’s Code Arena leaderboard.
uncertain
model → close → leaderboard
Jason Hsu, a senior fellow at the Hudson Institute, said American businesses that incorporate Chinese-based AI face several risks.
asserted
that → say → risks
“American companies building on Chinese models could face intellectual property vulnerabilities, software supply-chain risks, and long-term technological dependence,” Hsu told the DCNF on Monday.
uncertain
Hsu → build → Monday
“The larger national security concern is that widespread adoption of Chinese AI models could give Beijing greater influence over global digital infrastructure and technological standards.”
OpenAI and Anthropic both restrict the use of their respective models’ outputs to train outside AI.
uncertain
OpenAI → give → AI