Anthropic’s Dario Amodei Says Open Weights Are Not the Enemy, but Geopolitical AI Risk Is

Lead

Anthropic founder and CEO Dario Amodei has moved to clarify his position in a heated industry debate over open-weight AI models. In a new blog post, he said Anthropic has never advocated a ban on open-weight models, responding to speculation that the company supported U.S. restrictions on Chinese open-weight systems or open-weight AI more broadly.

The clarification came after Nvidia CEO Jensen Huang shared an open letter on X from a group of AI companies and organizations, including Hugging Face, Meta, Microsoft, Mistral, and Nvidia. The letter urged policymakers not to impose broad, premature restrictions on open-weight models. Although the letter did not specifically name China, much of the surrounding debate has focused on China’s AI progress, alleged intellectual property theft, and the use of model distillation to learn from U.S.-built systems.

Key points

• Open weights and China risk are separate issues: Amodei said open-weight models without dangerous capabilities can function as a public good. They are useful to businesses, developers, and researchers, and their main cost is the compute required to run them.
• His core concern is powerful AI in authoritarian hands: Amodei worries that authoritarian governments could build models stronger than those available in the U.S. and use them to secure lasting military advantage or intensify domestic repression.
• China is not the only concern, but it is central: He said the Chinese Communist Party is not the only authoritarian actor he worries about, but he described it as the most capable.
• Open weights become harder to govern at the frontier: Once model weights are released, they cannot easily be withdrawn. Amodei argued that this makes guardrails and monitoring more difficult, especially for models that could enable cyber or biological attacks.
• Policy should target capabilities and misuse pathways: His proposed measures include restricting access to advanced chips and taking formal action against distillation when it amounts to stealing model capabilities or intellectual property.
• Global safety testing may be a path forward: Amodei backed efforts to create safety testing for the most capable AI systems, regardless of whether they are open or closed and regardless of country of origin, while exempting less capable startup and academic models.

Why it matters

Amodei’s response attempts to separate two debates that are often collapsed into one: whether open models benefit the AI ecosystem, and how governments should respond to national-security risks from frontier AI. Open-source advocates argue that accessible models improve transparency, competition, and defensive capacity. Safety-focused leaders counter that once the most powerful models are released, they may be impossible to control.

Rather than calling for a blanket ban, Amodei is arguing for a capability-based approach. In that framing, smaller or less dangerous open models remain valuable, while the most advanced systems require testing, controls, and international coordination.

The most notable element may be his openness to global cooperation. He suggested that even China could have an interest in preventing AI-enabled biological weapons. That does not remove the geopolitical rivalry, but it points to a possible narrow area of shared interest: preventing the most catastrophic forms of AI misuse.

Source: TechCrunch AI