Anthropic CEO Clarifies Stance: Open Models Aren't the Enemy — Beijing's Military AI Ambitions Are

Anthropic CEO Dario Amodei took to his blog on Monday to clear the air amid swirling industry speculation that his company had been quietly backing efforts to restrict open-weight AI models — potentially including those developed in China. In no uncertain terms, Amodei wrote that Anthropic has never pushed for a ban on open-weight models and that anyone familiar with his prior writing would already know that. The clarification came with deliberate emphasis, suggesting the company felt the mischaracterization had reached a point where silence was no longer an option.

The backdrop to Amodei's statement was a high-profile open letter published Friday, co-signed by Nvidia, Meta, Microsoft, Hugging Face, Mistral, and others, urging policymakers to avoid imposing sweeping restrictions on open-weight AI. Notably, it marked Nvidia CEO Jensen Huang's first post on X, signaling how seriously the industry is taking this policy debate. Though that letter did not call out China directly, the broader conversation has been dominated by concerns that Chinese AI labs are advancing rapidly, in part by using distillation techniques — essentially bombarding American models with prompts to reverse-engineer their capabilities.

Amodei's blog post drew a careful distinction between open-weight models as a general category and the specific threat he believes China represents. He described open-weight models without dangerous capabilities as a "public good," offering real value to developers, researchers, and businesses at essentially no cost beyond compute. His concern is not that companies or individuals are using such models — even Chinese-made ones — but rather that authoritarian governments could develop frontier AI systems capable of delivering "permanent military superiority" or being weaponized against their own citizens.

Beyond geopolitical power dynamics, Amodei flagged biological attacks as a particularly alarming AI-enabled risk, noting that open-weight models are harder to monitor and impossible to retract once released — a point he backed up by citing a UK AI Security Institute report. This puts him at odds with open-source advocates who argue that decentralized, powerful models actually empower defenders more than they do bad actors. It is a genuine tension in the field, and one that shows no sign of easy resolution.

On constructive policy steps, Amodei endorsed continued chip export restrictions to China and called for a formal crackdown on AI distillation as a form of intellectual property theft. Perhaps most notably, he voiced support for a global AI model safety testing organization — one that would include China itself. He expressed cautious optimism that even the Chinese Communist Party might agree to limited cooperation around preventing AI-enabled bioweapons, framing it as a matter of mutual interest. The Trump administration's recent moves toward such a framework, he added, have been encouraging.