Anthropic's Amodei clarifies stance: no opposition to open-weight models, but warns of Chinese AI risks

Anthropic founder and CEO Dario Amodei has moved to address speculation within the AI sector, clarifying in a Monday blog post that his company does not support restrictions on open-weight models—even as he articulates distinct concerns about geopolitical AI competition.

The clarification followed Nvidia CEO Jensen Huang's Friday open letter, signed by major AI firms including Meta, Microsoft, and Mistral, urging U.S. policymakers to avoid sweeping restrictions on open-weight AI systems. Industry observers had questioned whether Anthropic favored stricter rules.

Amodei's position on open models#

"Anthropic has never advocated for a ban on open-weights models," Amodei wrote in his response, distinguishing between two separate issues that have become conflated in recent debate. He views open-weight systems without dangerous capabilities as beneficial—free to operate and valuable to businesses, developers, and researchers alike. His actual concern, he emphasized, lies elsewhere.

The real threat: authoritarian AI dominance#

Amodei's worry centers on authoritarian governments—particularly China—developing AI systems more powerful than those available in the West, potentially achieving what he calls "permanent military superiority" or using such tools for domestic repression. He flagged other concerns including biological attacks enabled by advanced AI, arguing that open-weight models pose particular risks because guardrails are harder to enforce and usage harder to monitor.

Why open-weight matters differently#

The distinction matters because open-source advocates contend that distributed access to powerful models actually strengthens defensive capabilities. Amodei countered that once open-weight systems are released, they cannot be recalled—citing analysis from the UK AI Security Institute.

Proposed countermeasures#

Rather than bans, Amodei advocated for a combination of approaches: maintaining existing restrictions on advanced chip exports to China, establishing formal penalties for model distillation (a technique where labs use prompts to reverse-engineer competitor systems), and building a global model safety testing framework. He expressed particular interest in a shared testing organization that even China might join, noting that preventing AI-driven biological weapons could align with Beijing's own interests. "I think this may actually be possible," he wrote.