Anthropic’s Dario Amodei responds: doesn’t oppose open-weight models, but fears Chinese AI

Anthropic CEO Clarifies Stance on Open-Weight AI

In response to industry speculation, Anthropic founder and CEO Dario Amodei has explicitly stated his company does not support a ban on open-weight models. His clarification, delivered via a blog post, comes on the heels of an open letter organized by Nvidia and signed by firms like Meta and Hugging Face, which urged policymakers to avoid “premature restrictions” on open models. Amodei’s statement aims to separate his company from calls for broad prohibitions, directly addressing murmurs that Anthropic backed government efforts to ban open-weight models, particularly those originating from China.

A Nuanced Threat Model: Open-Weights vs. State Actors

Amodei frames his concerns not around open-weight models themselves—which he considers a “public good” when they lack dangerous capabilities—but around the risk of authoritarian governments developing superior AI for military dominance or domestic repression. He identifies the Chinese Communist Party (CCP) as the “most capable” threat in this domain. He also highlighted the danger of AI-enabled “biological attacks,” arguing that the uncontrollable nature of released open-weight models makes them particularly risky in such scenarios, as they are difficult to monitor or apply guardrails to after distribution. To counter these specific threats, Amodei suggests several targeted actions:

• Restricting China's access to powerful computing chips.
• Calling for a formal crackdown on IP theft via model distillation.
• Supporting a global model safety testing organization, which would require cooperation from all nations, including China, to prevent catastrophic misuse.

Shifting the Policy Debate from 'Open vs. Closed' to 'Capabilities and Controls'

Amodei's detailed position introduces a critical nuance to the AI policy debate, shifting the focus from a simple binary of open versus closed models to a more complex discussion about specific capabilities and actor-based risks. By advocating for targeted interventions like chip export controls and anti-distillation measures, he proposes a middle path that seeks to preserve the innovative benefits of open research while mitigating national security risks. This perspective could influence policymakers to adopt more surgical regulations rather than broad bans, focusing on controlling access to foundational resources like compute and preventing the proliferation of models with verifiably dangerous, such as bioweapon-related, capabilities.

Dario Amodei is attempting to reframe the AI safety debate from a simple 'open vs. closed' binary to a more nuanced discussion focused on specific capabilities and malicious actors, advocating for targeted interventions like chip restrictions and distillation crackdowns rather than a blanket ban on open-weight models.