China's AI Gains Exceed Distillation Claims, Industry Leaders Say

A debate is intensifying over how China's artificial intelligence labs have achieved rapid advances that now rival American models on certain benchmarks. While U.S. officials and companies have attributed much of this progress to "distillation"—training models using outputs from more advanced systems—prominent AI figures are pushing back on that characterization.

Aidan Gomez, CEO of Cohere and co-author of the foundational 2017 "Attention Is All You Need" research paper, told CNBC that Chinese AI capabilities cannot be explained by distillation alone. He described Chinese models as "world class" and noted that the U.S. lead is "evaporating very quickly."

"A lot of these tropes were applied, and that was definitely true to an extent," Gomez said. "However, they have developed an exceptional capability independent of distillation, and it's proven by the fact that the latest models that are coming out, actually on some benchmarks, on some axis capabilities, beat the best American models. And you can't copy or distill to better."

The distillation controversy

Distillation has become a flashpoint in U.S.-China AI competition. Anthropic released a report this month claiming that Chinese companies including Alibaba, Moonshot, and DeepSeek attempted "illicit distillation" to train their models. The company's head of threat intelligence described "an entire illicit ecosystem" aimed at gaining access to Claude and other Western systems.

The U.S. Cybersecurity and Infrastructure Security Agency escalated the rhetoric further, stating that Chinese firms are conducting "industrial-scale knowledge distillation campaigns that form the core—not merely a supplement—of their AI development strategy."

Growing skepticism

Sriram Krishnan, a former senior White House policy advisor on artificial intelligence, questioned the significance of distillation concerns. He noted that services like ChatGPT and Claude themselves "came out of distilling human content" through training on internet data.

"So the idea of distilling has always been a core part of how computer science works," Krishnan told CNBC.

China's Ministry of Commerce rejected the allegations as "groundless and legally unsound."

Analysts point to U.S. export restrictions on advanced chips as a catalyst for Chinese innovation rather than an impediment. Neil Shah, a partner at Counterpoint Research, said chip restrictions forced Chinese developers to "build leaner, smarter architectures and do more with less compute."

"Dismissing that as mere imitation might make for convenient restrictive policy making, but it fundamentally misjudges the competition," Shah said.

Why it matters

The distillation narrative shapes U.S. policy decisions on export controls, investment restrictions, and technology partnerships. If Western policymakers are underestimating genuine Chinese technical capabilities—as Gomez and others suggest—those policies may prove ineffective or counterproductive. The debate also highlights how resource constraints can drive architectural innovation, a lesson relevant beyond U.S.-China competition as smaller labs worldwide seek to compete with frontier model developers.

These details were first reported by CNBC's Arjun Kharpal and Kai Nicol-Schwarz.