Anthropic rejects calls to ban open-weights AI models

Anthropic CEO Dario Amodei has clarified that the company does not support banning open-weights AI models, amid debate over Chinese models and reports that US officials may consider restricting their use by American businesses.

The company’s position is more targeted. Anthropic says non-dangerous open-weights models are a public good because developers, businesses, and researchers can use them without paying for access beyond the compute required to run them. However, it also warns that releasing model weights can make safeguards and monitoring harder, especially for models with cyber, biological, or alignment risks.

Anthropic’s preferred approach

Amodei argues that banning Chinese open-weights models would not address the company’s primary national-security concern: authoritarian governments developing more powerful AI systems for military use or domestic repression. It also would not stop malicious actors who are unlikely to operate as legitimate US companies.

Instead, Anthropic supports three measures:

• Restricting China’s access to advanced chips and chipmaking equipment, while pursuing smuggling and other workarounds.
• Targeting industrial-scale model distillation, which can help organizations build stronger systems with less computing capacity.
• Requiring sufficiently capable models—whether open or closed—to undergo safety testing before release, including evaluations for cyber, biological, and alignment risks.

The company says these policies should focus on capabilities and behavior rather than whether a model is open or closed, and that testing should be global. Anthropic also agrees with parts of an industry letter supporting open weights, including their potential to expand access and customer control, but disputes claims that broad availability automatically benefits defenders more than attackers.

For AI builders, the statement points toward a policy environment centered on capability thresholds, compute access, provenance, and pre-release evaluations—not a blanket ban on open model weights.

Source: Anthropic News