Amodei Pushes Back: Anthropic Never Called for a Ban on Open-Weight AI

Dario Amodei wants to set the record straight. In a statement posted to Anthropic's site, the company's CEO rejected the growing narrative that his firm has been lobbying to outlaw open-weight AI models, calling the characterization inaccurate.

"Anthropic has never advocated for a ban on open-weights models," Amodei wrote, directly countering critics who had accused the company of pushing for blanket restrictions. The clarification arrived the same week Nvidia CEO Jensen Huang published an open letter, co-signed by Meta, Microsoft, Hugging Face, and Mistral, urging regulators to avoid sweeping limits on openly released models.

Amodei's essay draws a sharper line than the one critics assumed he held. He describes open-weight models that lack dangerous capabilities as a genuine public good, benefiting startups, independent developers, and academic researchers who rely on freely available weights to build and study AI systems.

His actual worry sits elsewhere: the possibility that authoritarian states, chiefly the Chinese Communist Party, could develop AI systems that outpace American capabilities and deploy them for military dominance or internal surveillance and repression. That geopolitical framing, not a general hostility to open models, is what shapes his policy recommendations.

A secondary but related concern involves misuse. Amodei flagged the danger that highly capable models, once released, could be exploited for cyberattacks or biological weapons development, compounded by unresolved alignment risks that make model behavior hard to guarantee at scale.

Because open weights, once published, can never be recalled, Amodei leaned on a UK AI Security Institute finding to argue that release decisions carry irreversible consequences. That logic underpins two concrete policy positions he supports: tighter export controls on advanced chips and chipmaking tools sold to China, and a crackdown on distillation, the technique by which one model learns from another's outputs through automated querying.

He also floated a more cooperative measure: a global testing organization for model safety that would include China alongside the US and other nations, an attempt to manage the risk without shutting down cross-border research entirely.

The timing matters. Amodei's clarification followed reports that some US officials were weighing a ban on American companies using Chinese open-weight models, a policy direction he did not explicitly endorse but whose underlying anxiety he shares. He also cited the Intelligence Community's 2026 Annual Threat Assessment, which found that rival powers' AI progress is eroding the economic and security edge the US has historically held.

None of this is a sudden pivot. Amodei's essay echoes positions he laid out six months earlier in "The Adolescence of Technology," suggesting the current statement is less a course correction than a clarification aimed at industry peers who read his earlier remarks as anti-open-source.

"Anthropic has never advocated for a ban on open-weights models."

Whether that clarification satisfies Huang, Meta, and the other signatories of the joint letter remains an open question, but the disagreement now looks narrower: not whether open weights should exist, but how to handle the subset of models capable of causing serious harm.