Cohere’s Aidan Gomez calls Chinese AI models ‘world-class’ and says the US lead is shrinking fast

The story that Chinese AI progress is mostly borrowed homework has an official version in Washington and, as of Friday, a prominent sceptic inside the US industry itself. Cohere chief executive Aidan Gomez called Chinese AI models "world-class", said the advantage held by the leading American labs is shrinking quickly, and argued that distillation alone cannot explain the results.

Cohere Watch analysis

What happened

Speaking in an interview reported by CNBC on Friday, Gomez said China had partly relied on copying and distillation but had also built strong capabilities separate from both. Some of the latest Chinese models, he said, have surpassed top US models on certain benchmarks and capabilities. His logic was simple: copying or distillation might narrow a gap, but it cannot pull one ahead. Asked whether Western policymakers were underestimating China's technical capabilities, he answered yes.

That puts him against the strongest version of the theft narrative. Anthropic has repeatedly argued that distillation by Chinese labs amounts to theft, and per the report its head of threat intelligence, Jacob Klein, has described an illegal ecosystem seeking access to Claude and other models. Anthropic has claimed Alibaba, Moonshot AI and DeepSeek attempted illegal distillation to train their own models. CISA went further earlier this month, saying Chinese AI companies were extracting functions and capabilities from US models through industrial-scale knowledge distillation, and calling that the core of their development strategy rather than an auxiliary technique. Those are allegations and agency characterisations, not established findings.

The counterarguments do not come only from Beijing. Sriram Krishnan, a former White House AI policy adviser, said on Thursday that ChatGPT and Claude were themselves created by training on human content from the internet, that distillation is a core part of computer science, and that it remains unclear how much it actually contributes to performance. China's commerce ministry called the US allegations groundless and legally invalid. Neil Shah, a partner at Counterpoint Research, added the engineering angle: chip restrictions forced Chinese developers to build more efficient architectures with less compute, and dismissing that as simple imitation misreads the competitive landscape.

Why it matters

This is not an academic quarrel. The theft framing underwrites export controls, agency warnings and how Western labs describe their threat models, and CISA has now put a government stamp on the industrial-scale claim. If Gomez and Krishnan are closer to the mark, policy aimed purely at stopping copying will misjudge an industry that has, on Shah's account, spent years engineering around constraints. If Anthropic and CISA are right, relaxed talk of efficient engineering becomes complacency about unauthorised access at scale.

Either way, the useful shift is where the disagreement now lives: inside the US industry, in public, rather than only between Washington and Beijing.

Our read

Gomez's narrow point looks hard to dent. Imitation can carry you to the frontier, but it does not move it. Note, though, what he does not dispute: he accepts China partly relied on copying and distillation. The live question is not whether unauthorised training happened but how much it explains, and right now that question is being answered with adjectives. Treat both "theft" and "world-class" as claims to be tested against published benchmarks and disclosed evidence, and assume the loudest voices in this row each have something to protect.

What to watch

• The full interview
  Whether the CNBC conversation or a transcript names the benchmarks where Gomez says Chinese models lead, giving the claim something testable.
• Evidence over summaries
  Whether Anthropic or CISA publish technical detail for the alleged distillation campaigns, such as documented access patterns, rather than characterisations.
• Policy language
  Whether Washington's framing bends towards efficiency or hardens in response to the pushback.
• The next releases
  Whether upcoming Chinese models back "surpassed on certain benchmarks" with reproducible, independently checked results.

Discussion spark: Gomez says copying can close a gap but never open one. If that is technically true, how much of the theft framing survives, and what evidence would settle the question for you?

not affiliated with or endorsed by Cohere