Anthropic CEO's Position on Open-Weights Models

In a statement on July 27, 2026, Anthropic CEO Dario Amodei made it clear that the company has never advocated for a total ban on open-weight models. He noted that recent reports suggested U.S. officials were considering banning U.S. companies from using Chinese open-source models, while some tech companies signed an open letter supporting open-source models, and some accused Anthropic of trying to protect its own business by banning open-source models. Amodei emphasized that such accusations are not true.

Amodei believes that open-source models without dangerous capabilities are a public good, incurring no additional cost beyond the computing power required to run them, and providing value to businesses, developers, and researchers.

Amodei outlined two core concerns. The primary concern is that authoritarian governments (especially the Chinese Communist Party) could build AI models more powerful than those of the U.S. and use them to achieve permanent military superiority or deep suppression of their own people. He cited Vice President Vance's warning in Paris and the 2026 Intelligence Community Annual Threat Assessment, noting that AI progress in other global powers is challenging U.S. economic competitiveness and national security advantages. Amodei stressed that whether a model is open source is irrelevant; the most dangerous models could be trained in secret and used solely by the People's Liberation Army or the Ministry of State Security.

The secondary concern is that powerful AI models could be misused for cyber attacks or biological attacks, and there are serious alignment issues. Open-source models (whether from China or elsewhere) pose higher risks because it is difficult to impose guardrails, monitor usage, and once weights are released, they cannot be revoked. However, banning U.S. companies from using these models does not solve the risk, as malicious actors are unlikely to be legitimate U.S. companies.

Amodei proposed three measures he has long advocated. First, not selling advanced chips or chip manufacturing equipment to China, and cracking down on rampant smuggling and workarounds. Due to China's limited domestic production capacity and the constraints of scaling laws, without U.S. chips, China cannot build models more powerful than those of the U.S. This is the most direct and effective way to address the primary threat, while also indirectly helping to address the secondary threat.

Second, cracking down on industrial-scale distillation operations. Distillation is much more efficient than training models from scratch, allowing China to build better models despite limited chip availability, thus partially circumventing chip bans. While distillation cannot give China AI capabilities equal to or better than those of the U.S., it can bring Chinese frontier models close to U.S. frontier levels (within a gap of months). Amodei noted that many companies performing distillation release open-source models, but the open weights are far less important than the fact that these operations are supported by authoritarian states seeking to surpass the U.S. A total ban on open-source models is neither the correct remedy nor what Anthropic is calling for.

Third, mandatory safety testing for all sufficiently powerful models (both open source and closed source). This is the best way to address the secondary threat. Amodei expressed relief at recent moves by the Trump administration and industry proposals that apply testing to the most powerful models, regardless of their country of origin or whether they are open source, while exempting weaker models from startups and academia. Testing should apply globally, including China, and may be achieved through limited cooperation, as preventing AI-driven biological weapons also aligns with China's interests.

Amodei agreed with some points in the open letter: open weights expand access to the AI economy, enhance competition in certain use cases, and give customers more control. The distillation issue should be addressed through targeted legal and commercial frameworks. However, he disagreed with the assertion in the open letter that open-source models are inherently easier to develop safety measures for, and that broad access to capabilities necessarily benefits defenders. He believes the opposite may be true.

He gave an example that there may be a strong attacker-defender asymmetry in the biological domain: a sufficiently powerful model could use widely available materials to rapidly weaponize a pandemic-level virus, while defense requires years of operation (as shown by 'Operation Warp Speed'). Such issues should be answered empirically through rigorous pre-release testing, not assumed in advance.