Anthropic CEO Rejects Calls To Ban Public-Weight AI, Warns Of State-Led Threats

In a post Monday, Anthropic founder Dario Amodei pushed back against suggestions his company wants sweeping prohibitions on models whose parameters are publicly available, while warning that the real risk comes from powerful systems built by authoritarian states.

Amodei’s clarification followed a high-profile letter circulated last week by Nvidia and a coalition of large AI firms urging regulators not to move quickly to limit access to models with released weights. The original appeal, shared on social media by Nvidia’s chief executive, called for caution about broad restrictions—an argument that set off a wider industry debate about how to balance openness and security. Much of the talk has centered on Chinese AI teams, which some U.S. companies accuse of accelerating progress by appropriating others’ work. One method often cited involves repeatedly querying a target system to reconstruct its behavior and replicate its capabilities.

The Anthropic chief drew a bright line between two categories of risk. He argued that models made available for anyone to run, when they lack hazardous functions, deliver public benefits: they’re inexpensive to operate and help startups, researchers and small companies build tools and diagnose problems. One independent policy analyst said that stance reflects a common tension in technology regulation—protecting innovation while guarding against misuse.

But Amodei stressed he’s most alarmed by state-directed efforts to outbuild U.S. capabilities. He warned that some governments may pursue more advanced models to secure lasting military advantages or to tighten repression at home, and he singled out China as the most advanced among those regimes. He also flagged the danger of AI lowering barriers to biological attacks, arguing that systems whose internals are exposed are harder to gate and to track once they spread. A recent U.K. security review, he noted, argues that once model parameters are published they’re effectively impossible to withdraw—a point that clashes with advocates who say open access helps defenders learn and prepare.

To blunt those threats, Amodei outlined several policy moves he supports: continued limits on the sale of top-tier accelerator chips to certain countries, stronger enforcement against techniques that effectively clone proprietary models, and the potential for sanctions when intellectual property is misappropriated. He also voiced support for an international model-safety testing body, provided participation is genuinely global. He said recent proposals from U.S. officials and industry groups—to subject the most capable systems to scrutiny regardless of their origin or whether they’re openly released—appear to be gaining traction.

Whatever one makes of his proposals, Amodei’s intervention highlights a widening split in the AI field between defenders of openness and those urging tighter controls on state-related threats. Creating a universally accepted testing regime would be difficult, one academic observer noted, but may be one of the few workable paths toward reducing the risk that advanced models become tools of coercion or harm.