Key Highlights
- A growing number of AI industry voices are questioning whether "distillation" fully explains China's rapid progress in developing advanced AI models.
- Cohere's chief executive said Chinese AI labs have developed capabilities independent of distillation, noting some Chinese models now outperform top American systems on certain benchmarks.
- Anthropic has accused Chinese labs including Alibaba, Moonshot and DeepSeek of using "illicit distillation" that amounts to theft of proprietary model capabilities.
- Analysts note that US chip export restrictions may have inadvertently pushed Chinese developers toward more efficient, innovative model architectures.
A Contested Explanation for China's AI Gains
As Chinese AI labs continue releasing increasingly advanced models that compete directly with leading American systems, a growing debate has emerged over how much of that progress can genuinely be attributed to "distillation," a Training technique in which a less advanced model learns from the outputs of a more capable one. US companies and government officials have repeatedly accused Chinese labs of using the practice to gain ground, characterizing the approach as effectively amounting to theft of proprietary technology.
However, a growing chorus of industry voices, including some prominent AI executives, argue that distillation alone cannot fully account for China's rapid advancement, suggesting Washington's framing of the issue may be overly simplistic.
An Industry Leader Pushes Back
The chief executive of AI company Cohere, an authoritative figure in the field who co-authored a foundational 2017 research paper underpinning much of modern AI technology, described current Chinese AI models as "world class," adding that the technological lead US labs have historically held is "evaporating very quickly." He acknowledged that some degree of distillation-based copying has genuinely occurred among Chinese developers, but argued that Chinese labs have since developed exceptional capabilities independent of that technique.
He pointed to recent Chinese models that outperform the best American systems on certain benchmarks and specific capability dimensions as evidence that distillation alone cannot explain the progress, arguing that copying can narrow a technological gap but cannot produce genuine outperformance beyond what the copied source model itself achieves.
Anthropic's Continued Distillation Allegations
Anthropic, the maker of the Claude AI model, has repeatedly asserted that Chinese AI labs have engaged in distillation practices that amount to theft. The company's head of threat intelligence said this month that an entire illicit ecosystem exists specifically dedicated to gaining unauthorized access to Claude and other proprietary models. In a report released this month, Anthropic specifically named Alibaba, Moonshot and DeepSeek, three of China's leading AI labs, as having attempted to use illicit distillation techniques to train their own systems.
US government officials have echoed similar concerns. Earlier this month, a federal Cybersecurity agency stated that Chinese AI companies are conducting systematic extraction of proprietary capabilities from US AI companies' models through what it described as industrial-scale knowledge distillation campaigns forming the core, rather than merely a supplement, of China's broader AI development strategy.
Questioning the Scale of the Practice
A former senior White House policy advisor on artificial intelligence offered a more measured view, noting that services such as ChatGPT and Claude themselves emerged from a broad process of distilling human-generated content, since the underlying models were trained on vast amounts of internet data. He argued that even if Chinese labs are using distillation techniques, it remains genuinely unclear how significantly this practice is actually contributing to their overall capability advancement, echoing skepticism similarly expressed by Cohere's chief executive.
China's government has firmly rejected the distillation allegations, with a Ministry of Commerce spokesperson describing US claims of industrial-scale distillation as "groundless and legally unsound."
Chip Restrictions as an Unintended Catalyst
Some industry analysts suggest that US export restrictions on advanced semiconductors, intended to slow China's AI progress, may have inadvertently spurred genuine innovation among Chinese developers. A research analyst covering the sector said that chip restrictions forced Chinese developers to build leaner, more computationally efficient architectures, arguing that dismissing this adaptation as mere imitation fundamentally misjudges the nature of the competition China's AI industry now presents.
Outlook
Whether Western policymakers and AI labs adjust their assessment of China's AI capabilities will likely depend on how the performance gap between Chinese and American models continues to evolve in the coming months. Industry observers who have closely tracked China's technology sector suggest that recognizing a genuine shift from imitation toward independent innovation may take time, given how entrenched the "copycat" characterization has historically been in Western discourse about Chinese technology development.






Please wait processing your request...