The US Trump administration is considering a ban on Chinese AI models, citing intellectual property theft through a technique called 'distillation' where smaller student models mimic larger teacher models. The proposed restrictions would prohibit domestic enterprises from using Chinese models and potentially sanction Chinese companies. This move represents an expansion of the US-China tech war from hardware restrictions to software limitations, according to scholars interviewed by NOWNEWS.
Scholar Liu Pei-chen Analyzes US-China Tech War Expansion to Software
Liu Pei-chen, Director of the Industrial Economics Database at the Taiwan Institute of Economic Research, stated that the US government's consideration to upgrade the tech war from hardware blockades to software bans on open-source models and APIs will inevitably trigger deeper US-China technological confrontation and market chilling effects.
For Taiwanese enterprises, if the US comprehensively tightens the software supply chain, it will not only increase compliance operational pressure but may also force the global AI ecosystem to face the severe test of choosing sides and supply chain reorganization.
Liu noted that despite expanded restrictions, Taiwan has played a core hub role in global chip manufacturing under the framework of continued restrictions on hardware exports to China during the US-China trade war. In the long term, as long as Taiwan continues to consolidate the irreplaceable nature of advanced processes, Taiwan can still secure a foothold in the AI wave.
Scholar Dai Zhi-yan Explains Enforcement Challenges Using Water Infrastructure Metaphor
"The reservoir is turned off, but it's difficult to completely remove the water pipes," said Dai Zhi-yan, Associate Research Fellow at the Second Research Institute of the Chung-Hua Institution for Economic Research, in an interview with NOWNEWS. He explained that if the US government adopts cloud processing restrictions—prohibiting internet use—it remains difficult to regulate enterprises downloading and using models internally.
Dai cited the example that when DeepSeek launched last year and faced potential bans, some US enterprises continued using it. The prohibition only extends to uploading network data and online use, but through Open Source downloads to local terminals, data processing and use can still occur locally.
Security Concerns About Backdoor Risks in Downloaded Models
Dai warned that many enterprises believe internal data that doesn't go online won't face leakage or training source issues, but it's difficult to guarantee against hard-to-detect malicious "backdoor" risks. He believes that at this stage, everyone focuses on inference accuracy and computing power costs while neglecting security concerns.
FAQ
What technique does the US cite as reason for the proposed Chinese AI model ban?
The US Trump administration cites "distillation" as a key concern—a machine learning training technique where smaller, simpler "student models" mimic large, powerful "teacher models"—which the US considers a form of intellectual property theft.
Why do scholars believe a US ban on Chinese AI models would be difficult to enforce?
Scholar Dai Zhi-yan explained using a water infrastructure metaphor that while the US can "turn off the reservoir" by blocking cloud and internet use, it cannot "completely remove the water pipes"—meaning enterprises can still download open-source Chinese AI models and use them locally on internal systems, which the US government cannot effectively monitor or prevent.
