Anthropic CEO Dario Amodei has clarified that he does not support banning open-weight artificial intelligence models, while maintaining that increasingly powerful AI systems developed by authoritarian governments could pose serious global security risks.
Amodei addressed the issue following growing debate across the technology industry about potential government restrictions on open-weight AI. These models allow developers and organisations to download, modify, and operate their underlying parameters with fewer controls from the original creator.
The Anthropic co-founder said his company has never advocated for a general ban on open-weight models. Instead, he argued that policymakers should distinguish between widely useful models with limited dangerous capabilities and highly advanced systems that could be misused for military, biological, or surveillance purposes.
Amodei described less capable open-weight models as a public benefit because businesses, researchers, and independent developers can use them without paying for access to a proprietary platform.
Such models can support experimentation, academic research, local deployment, and the development of specialised AI applications. They can also provide organisations with greater control over their data and reduce their dependence on a small number of major AI companies.
His comments place Anthropic in a more nuanced position than critics had suggested. The company appears to support continued access to open models while calling for greater scrutiny of systems that reach potentially dangerous levels of capability.
The clarification came after several major technology companies and AI organisations urged US policymakers to avoid broad or premature restrictions on open-weight technology. The group included Nvidia, Meta, Microsoft, Hugging Face, and Mistral.
Amodei’s primary concern is not ordinary business use of open-weight models, including models developed in China. His warning focuses on the possibility that authoritarian governments could build AI systems more powerful than those available in democratic countries.
He argued that such systems could be used to gain military advantages, strengthen domestic surveillance, suppress political opposition, or expand state control over citizens.
China is a central part of this concern because of the country’s technological resources and rapidly developing AI sector. However, Amodei said the broader risk applies to any authoritarian government capable of developing advanced artificial intelligence.
He also warned that highly capable AI could increase biological and cybersecurity threats. Open-weight systems may present additional challenges in these areas because their behaviour cannot easily be monitored or restricted after release.
Once a model’s parameters are publicly distributed, its original developer may have little ability to withdraw the system or prevent users from modifying its safeguards.
Another major issue is AI distillation, a process through which one model can learn from the outputs of another model by sending it large numbers of prompts.
American AI companies have expressed concern that overseas laboratories could use this technique to reproduce valuable capabilities developed by US companies. Chinese AI organisations have faced allegations of using outputs from Western models, although the companies involved have disputed or not publicly confirmed many of these claims.
Amodei called for stronger action against unauthorised distillation and supported continued restrictions on China’s access to the most advanced AI chips.
US export controls already limit the sale of certain high-performance processors to China. Supporters say these restrictions can slow the development of advanced military and surveillance technologies, while critics argue that they may encourage China to accelerate its domestic semiconductor industry.
Amodei also expressed support for an international organisation that would test the most powerful AI systems for dangerous capabilities.
Under such a framework, advanced models could be evaluated regardless of whether they are open-weight or closed, and regardless of the country in which they were developed. Smaller models created by startups, universities, and independent researchers could remain exempt to avoid placing excessive regulatory pressure on innovation.
For the system to work, Amodei said participation would need to extend beyond the United States and its allies. China and other major AI powers would also need to cooperate.
Although broader geopolitical cooperation may be difficult, he suggested that countries could share an interest in preventing AI-enabled biological attacks. Even strategic competitors could benefit from common testing standards designed to reduce threats capable of crossing national borders.
The debate highlights the challenge facing governments as they attempt to regulate artificial intelligence without restricting useful research and competition.
Broad restrictions on open-weight models could benefit large technology companies by making it harder for smaller developers to compete. At the same time, releasing increasingly powerful systems without safeguards could make dangerous capabilities widely available.
Amodei’s position calls for regulation based on a model’s capabilities and potential risks rather than whether its technology is open or proprietary.
As AI systems become more advanced, policymakers will need to determine where that threshold should be placed. The outcome could shape not only the future of open AI development but also the balance of technological power between the United States, China, and the rest of the world.
Share your thoughts about this article.
Be the first to post a comment!
