Anthropic CEO Denies Advocating Open-Weight Model Ban

Amodei Warns of Chinese AI Risks, Calls for Export Controls and Safety Checks

Dario Amodei, Anthropic CEO, issued a statement on the 27th (local time) stating, "Anthropic has never advocated for the prohibition of open-weight models." This came after reports that the U.S. government is considering restrictions on the use of Chinese-made open-weight models, prompting criticism that Anthropic sought to regulate competing models to protect its proprietary AI. The CEO directly refuted these claims.

In a notice titled 'Our Position on Open-Weight Models,' Amodei stated, "Open-weight models that do not include dangerous functionalities are public goods. These models incur no additional costs beyond computing expenses and provide value to businesses, developers, and researchers."

However, he emphasized the need to guard against China's rapid advancement in AI technology. "The most concerning issue is that authoritarian governments, including China, could develop AI more powerful than the U.S. to secure military superiority or use it for surveillance and repression of their citizens," Amodei said. "The most dangerous AI could be models secretly developed and supplied to the Chinese People's Liberation Army's drones or the Ministry of State Security's surveillance systems." He added that whether Chinese companies disclose model weights or U.S. firms use them is of little significance.

Amodei maintained the existing stance that open-weight models pose higher risks of misuse than closed models. "Regardless of their origin, open-weight models inherently carry potentially greater risks than closed models," he explained. "Safety measures are extremely difficult to apply or monitor, and once weights are public, they cannot be retracted."

He advocated for stricter export controls on advanced AI chips and semiconductor manufacturing equipment. Since China's domestic production of advanced chips is limited, Amodei argued that without access to U.S.-made chips, it would struggle to train models surpassing those of the U.S.

Amodei also called for intensified crackdowns on "distillation" by Chinese AI companies. Distillation involves using responses from powerful models as training data to enhance performance with fewer computational resources. Anthropic claims Chinese firms are distilling its Claude model to improve their own. He further argued that mandatory safety checks should be required before the release of any AI model exceeding a certain capability threshold.

Amodei explained why Anthropic did not sign a joint letter opposing open-weight model regulations, which was endorsed by over 80 companies including NVIDIA, Meta, Microsoft, Google, and OpenAI. While acknowledging that open-weight models "expand AI accessibility, foster competition, and empower user control," he disagreed with the claim that "open-weight models necessarily make safety research easier or favor defenders over attackers."