Dario Amodei, founder and CEO of Anthropic, responded Monday afternoon to industry speculation that his company backs U.S. government efforts to ban open-weight Chinese models or open-weight models more broadly. Amodei wrote plainly that “Anthropic has never advocated for a ban on open-weights models,” and reiterated that he does not view such bans as a useful policy.
His post followed a related industry move: Jensen Huang, founder and CEO of Nvidia, used X on Friday — his first post on the platform — to share an open letter in which Nvidia and other AI firms, including Hugging Face, Meta, Microsoft and Mistral, urged policymakers not to impose broad, premature restrictions on open-weight AI models.
Distinguishing public models from state-driven threats
In his Monday blog entry, Amodei argued that open-weight models should be treated differently from the national-security threat posed by certain states. He said open-weight models that lack dangerous capabilities function as a public good: aside from the compute required to run them, they impose no direct cost and offer value to businesses, developers and researchers.
Amodei said he has long-standing concerns about AI, but he does not regard businesses using open-weight models — even those produced in China — as the principal worry. Instead, his fear is that authoritarian governments could build models more powerful than those in the United States to secure permanent military advantage or to use AI to repress their populations. He noted that while the Chinese Communist Party (CCP) is not the only authoritarian government of concern, it is the “most capable.”
Worries about biological misuse and irreversibility
Amodei also warned that AI could enable biological attacks in addition to cybersecurity threats. He argued that open-weight models are particularly problematic in these scenarios because it is difficult to impose guardrails on them or to monitor their use. Referring to a UK AI Security Institute report, he pointed out that once open-weights are released they cannot be fully withdrawn.
That position contrasts with open-source advocates who maintain that access to powerful, non‑proprietary models helps defenders, researchers and the public better understand and mitigate risks.
Policy steps Amodei supports: chips, distillation bans, and global testing
To counter risks from China, Amodei listed measures he believes would be effective: restricting access to high-performance chips — a long-standing U.S. policy — and formally cracking down on distillation, a technique that can be used to replicate model capabilities by repeatedly querying a target model.
The U.S. has previously threatened sanctions if it determined that intellectual property theft involving U.S. models had occurred.
Amodei also expressed support for growing efforts, some led by the U.S., to establish a model-safety testing organization. He said such testing should apply to the most capable models regardless of country of origin or whether the models are open or closed, while exempting less capable models from startups and academia from these requirements. Crucially, he underscored that testing would need to be global to be effective, which would require the cooperation of actors including the CCP. He suggested limited cooperation to prevent AI-enabled biological weapons could be achievable because it aligns with China’s interests as well.
Conclusion
Amodei’s statement clarifies that Anthropic does not support a blanket ban on open-weight models. At the same time, he framed a set of targeted policy responses to address risks from authoritarian states — particularly China — and potential AI-enabled biological threats, including chip export controls, action against distillation, and a globally applied model safety testing regime that would include participation from China.
