Pushback is growing within the U.S. AI industry against the view that the performance of Chinese AI models can be explained by “distillation” alone. On Sept. 18, CNBC reported that Cohere CEO Aidan Gomez (에이단 고메즈) called Chinese AI models “world-class” and said the advantage held by leading U.S. labs was shrinking quickly.

Distillation is a way of training one model by using the output, or answers, of a more advanced AI model. U.S. companies and Washington have criticised Chinese labs by calling it theft, saying they are boosting competitiveness through this approach.

But Gomez said China’s recent results were hard to explain with that alone. He said China had partly relied on copying and distillation, but had also built strong capabilities separate from distillation. Some of the latest Chinese models have surpassed top U.S. models on certain benchmarks and capabilities, he said, stressing that copying or distillation might narrow a gap but cannot pull ahead.

That view differs from claims by the U.S. government and some AI companies. Anthropic has repeatedly argued that Chinese AI labs’ use of distillation amounts to theft. This month, Jacob Klein (제이컵 클라인), Anthropic’s head of threat intelligence, said there was an illegal ecosystem seeking access to Claude and other models. In a separate report, Anthropic claimed Alibaba, Moonshot and DeepSeek tried “illegal distillation” to train their own models.

The U.S. Cybersecurity and Infrastructure Security Agency also said earlier this month that Chinese AI companies were systematically extracting functions and capabilities from U.S. AI company models through industrial-scale knowledge distillation, and that this was not an auxiliary means but the core of their development strategy.

Others inside the United States say distillation should be viewed more broadly. Sriram Krishnan (스리람 크리슈난), a former White House AI policy adviser, said on Sept. 17 that services such as ChatGPT and Claude were also created by training on human content on the internet, and that distillation was a core part of computer science. He said it was not clear how much distillation contributed to performance gains even if Chinese labs used it.

The Chinese government rejected the U.S. claims. A spokesperson for China’s commerce ministry said allegations of industrial-scale distillation raised by the United States were groundless and also not legally valid.

China’s AI advances are continuing despite restrictions on access to chips. China has expanded domestic semiconductor development as access to cutting-edge chips from companies such as Nvidia has been restricted. Neil Shah (닐 샤), a partner at Counterpoint Research, said chip rules forced Chinese developers to build more efficient architectures with less computing, and dismissing that as simple imitation would misread the competitive landscape. Gomez also answered yes when asked whether Western policymakers were underestimating China’s technological capabilities.