Anthropic founder and CEO Dario Amodei responded to recent controversies surrounding the company's stance. He stated that Anthropic has never advocated for the US government to ban open-source weighted models, nor does it support blanket restrictions on such models.

Responding to rumors of being banned

In a blog post published on Monday, Amodei stated that claims that Anthropic supports restricting Chinese open-source models or even opposes open-source weighted models are inaccurate. He wrote that he has never believed that "banning" is an effective approach, and therefore it was necessary to reiterate his position.

He also distinguished the relationship between open-source weight models and national security risks. According to him, open-source weight models that do not possess dangerous capabilities have public value and can be used by enterprises, developers, and researchers at a lower cost.

Concerns point to the misuse of stronger AI

Amodei stated that his long-standing concern is not about companies using open-source models, even if those models originate from China. His greater concern is that authoritarian governments will develop AI systems more powerful than those of the United States and use them to gain long-term military advantage or strengthen their control over their societies.

Besides cybersecurity, he also mentioned that AI could be used in higher-risk scenarios such as biological attacks. In these cases, the problem with open-source weighted models is that once the model weights are made public, it becomes difficult to apply safeguards and to continuously monitor usage.

Support global security testing

Regarding countermeasures, Amodei listed several measures he supports, including continuing to restrict China's access to high-performance chips and formally cracking down on "distillation" practices. Distillation refers to the practice of one model learning the output and performance capabilities of another model by providing it with extensive feedback.

He also expressed support for establishing a model security testing organization, provided that major countries are willing to accept testing. He believes that for such a mechanism to be truly effective, it must cover the entire globe, and China needs to participate; limited cooperation is still possible on issues such as preventing AI from being used for biological weapons.