Dario Amodei has denied that Anthropic advocates banning open-weights AI models, responding to speculation around potential US restrictions on Chinese models.

The CEO's statement comes after reports suggested US officials are considering banning American companies from using Chinese open-weights models. Tech companies have since signed a letter supporting open-weights development, with some accusing Anthropic of seeking protectionist measures.

"Anthropic has never advocated for a ban on open-weights models," Amodei wrote in a blog post. He described open-weights models without dangerous capabilities as "a public good" that provides value to businesses, developers, and researchers.

Instead, Amodei outlined two primary national security concerns. First, the risk that authoritarian governments could build more powerful AI models than the US and achieve military superiority or enable deep repression. Second, the potential for powerful models to enable cyberattacks or biological attacks through misuse.

Anthropic's policy recommendations

The company supports three specific measures rather than blanket bans. These include blocking powerful chip sales to China and cracking down on smuggling operations that circumvent export controls.

Anthropic also wants action against industrial-scale distillation operations, which allow China to build better models with fewer chips. Amodei noted these operations can bring Chinese AI capabilities to within "a few months" of the US frontier.

The third recommendation calls for mandatory safety testing of all sufficiently capable models, both open and closed. Amodei said such testing should be global and could include cooperation with China on preventing AI biological weapons.

Regarding the industry letter supporting open-weights models, Amodei agreed with points about expanding access and strengthening competition. However, he disputed claims that open-weights models necessarily make safeguards easier to develop or help defenders more than attackers.

The CEO emphasized that questions about model safety should be "empirically answered by rigorous pre-release testing, not assumed in advance." He referenced recent research on modular training strategies as potentially promising for improving open-weights model safety.

Anthropic's position reflects broader industry tensions over balancing AI development openness with national security concerns as competition with China intensifies.