Anthropic CEO Dario Amodei denied that the company supports proposals to ban artificial intelligence (AI) models with open weights, including those developed in China.
The statement was published on Monday, following increased discussions in the United States about possible restrictions on AI models whose weights can be accessed and modified.
"Anyone who has read my previous texts should know that I do not consider these bans a useful measure, but let me make it clear so there is no doubt: Anthropic has never advocated banning open-weight models," Amodei wrote.
The statement came days after Jensen Huang, CEO of Nvidia, published an open letter on X against possible "premature restrictions" on open-weight AI models.
The document was supported by companies such as Hugging Face, Meta, Microsoft, and Mistral. The discussion, however, took on a geopolitical component due to the advances of Chinese AI labs.
Amodei separated the debate over open models from concern about China's technological capabilities. According to him, models without dangerous capabilities can generate benefits for developers, researchers, and companies.
"Open-weight models that do not possess dangerous capabilities are a public good: they cost nothing beyond the computing power needed to run them and add value to companies, developers, and researchers."
The executive said that his main concern is related to the use of advanced systems by authoritarian governments. Among the risks cited are military applications, internal repression, and biological attacks.
For Amodei, open-weight models present additional control difficulties because, once made available, their weights cannot simply be withdrawn.
As an alternative, the CEO defended measures against Chinese access to advanced chips and actions to combat model distillation, a technique that may allow one system to learn from another.
Amodei also supported the creation of international safety tests for the most capable models, regardless of whether they are open or closed.
"I think this idea is actually close to a consensus: I was encouraged by the fact that the Trump administration has moved in this direction in recent months, and by the industry's recent proposals that would apply these tests to the most capable models, regardless of their country of origin or whether they are open or closed (fully exempting less capable models, such as those from startups and academia)," Amodei wrote.
"Note that, to be effective," he added, "the tests would need to be global, which means that even the CCP would need to be involved. I think this may be possible… limited cooperation around preventing biological weapons with AI may be possible because it is also in China's interest."