Anthropic CEO Dario Amodei clarified that he does not oppose open-weight AI models, but expressed concern about China's advancements in AI, highlighting geopolitical risks.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Dario Amodei explicitly states Anthropic has never advocated banning open-weight models, but warns that authoritarian governments like China could use more powerful AI for military advantage or oppression, and calls for global cooperation on model safety testing.
Anthropic CEO Dario Amodei responded on July 27, 2026, to industry rumors, explicitly stating the company has never advocated banning open-weight models. He emphasized that open-source models without dangerous capabilities are a public good, requiring only compute costs to run, and are valuable for businesses, developers, and researchers.
Previously, Nvidia CEO Jensen Huang, along with multiple AI companies (including Hugging Face, Meta, Microsoft, Mistral, etc.), published an open letter urging policymakers not to impose 'premature restrictions' on open-source AI models. Amodei's response aims to clarify that his stance is unrelated to an open-source ban.
Amodei's fear does not stem from corporate use of open-source models (including Chinese models), but from the possibility that authoritarian governments could build more powerful models than the US to achieve 'permanent military advantage' or oppress their populations. He notes that the Chinese Communist Party (CCP) is not the only concerning authoritarian government, but it is 'the most capable.'
He also worries that AI will enable biological attacks, not just cybersecurity threats. In dangerous scenarios like bioweapons, open-source models are harder to guardrail or monitor for use. He cites a UK AI Safety Institute report stating that once open weights are released, they cannot be recalled.
Amodei lists actions to contain China, including restricting its access to advanced chips (a long-standing US policy) and formally cracking down on model distillation. Distillation is a technique that learns a model's workings by bombarding it with prompts; the US has threatened sanctions if IP theft involving US models is detected.
Notably, Amodei supports efforts to establish a global model safety testing organization, especially if all countries, including China, agree to be tested. He believes this could be a consensus, and the Trump administration has moved in this direction in recent months. He adds that limited cooperation on preventing AI bioweapons may be feasible, as it aligns with China's own interests.
This article is based on TechCrunch's reporting on Amodei's blog post, a reliable first-hand source. All views are explicitly attributed to Amodei, with no external knowledge introduced.
Amodei's stance: open-source models themselves are not a threat, but authoritarian governments (especially China) could use more powerful AI for military and oppressive purposes; countermeasures should focus on chip restrictions, cracking down on distillation, and global safety testing cooperation, rather than a blanket ban on open-source.
Primary report
Primary source
Also covered by 1 sources