Anthropic’s Dario Amodei Says He Doesn’t Oppose Open-Weight Models but Fears Chinese AI

Anthropic CEO Dario Amodei says his company does not support banning open-weight artificial intelligence models, pushing back against growing speculation in the AI industry while maintaining that increasingly capable Chinese AI systems could pose serious national security risks.

Amodei addressed the debate Monday, writing that Anthropic has never advocated for a ban on open-weight models. His comments followed an industry letter urging U.S. policymakers to avoid broad restrictions on openly available AI technology.

The letter was backed by companies including Nvidia, Hugging Face, Meta, Microsoft and Mistral. Nvidia CEO Jensen Huang shared the statement on X as the technology industry debates how Washington should respond to increasingly capable Chinese AI developers.

Amodei separates OpenAI from China concerns.

Amodei argued that open-weight models without dangerous capabilities can provide significant benefits to businesses, developers and researchers. Once released, these models can generally be downloaded and operated using the computing resources available to users, giving organisations more control over how they deploy AI.

He also indicated that ordinary business use of open-weight models, including models developed in China, is not the central concern behind his warnings about advanced AI.

Instead, Amodei said he is worried that authoritarian governments could develop systems more powerful than those available in the United States and use them to gain military advantages or strengthen domestic repression. While his concerns are not limited to China, he described the Chinese Communist Party as the most capable authoritarian government in this area.

Powerful open models could create additional risks

Amodei also warned that highly capable AI could increase risks involving biological attacks and cybersecurity. In those scenarios, he argued, open-weight models could become more difficult to control because developers cannot easily monitor how downloaded models are being used or reliably enforce safeguards.

Once model weights are publicly released, they also cannot simply be recalled. Amodei cited research from the U.K. AI Security Institute highlighting the permanent nature of releasing model weights.

That position differs from arguments made by open-source AI advocates, who contend that giving researchers and cybersecurity defenders access to powerful models can help them identify vulnerabilities and develop defences without depending on a small number of proprietary AI providers.

Distillation and advanced chips remain concerns.

Rather than broadly restricting open models, Amodei outlined other measures he believes could limit China’s ability to develop frontier AI systems. Those include maintaining restrictions on access to advanced computing chips and creating stronger measures against unauthorised model distillation.

Distillation allows one AI system to learn from the outputs of another model, potentially enabling developers to reproduce some capabilities without independently completing the same expensive training process. U.S. officials have considered stronger responses when such techniques involve alleged theft of intellectual property from American AI companies.

Amodei backs global AI safety testing.

Amodei also expressed support for efforts to establish organisations capable of independently testing the safety of the world’s most advanced AI models. He suggested that the approach should apply to powerful systems regardless of whether they are developed in the United States, China or elsewhere and regardless of whether they are open or proprietary.

Less capable models developed by startups and academic researchers could be exempt, allowing regulation to concentrate on systems presenting the greatest potential risks.

For such a framework to work, Amodei said participation would ultimately need to extend beyond the United States and its allies to include China. He suggested limited international cooperation may be achievable in areas where countries share a common interest, particularly preventing AI from enabling biological weapons.

His position distinguishes restricting open AI as a category and controlling the risks associated with the world’s most capable systems. Amodei argues that open-weight models can remain valuable to the broader technology ecosystem. At the same time, governments focus their attention on frontier capabilities, advanced computing resources and potential misuse by powerful state actors.