Anthropic CEO Defends Advanced AI Push Amid Safety Policy Debate

Anthropic CEO Dario Amodei defended the company’s pursuit of advanced AI models on Oct. 21, arguing that work at the technological frontier helps it push for stronger safety practices across the industry.

His response followed criticism of a draft AI safety bill Anthropic shared with U.S. lawmakers and changes to the company’s Responsible Scaling Policy. The bill remains a proposal. The debate centers on how Anthropic can continue developing increasingly capable models while meeting its stated safety commitments.

Nathan Calvin, Encode AI’s strategy lead, questioned whether those commitments had kept pace with Anthropic’s development efforts. He pointed to the company’s computing investments, its growing use of Claude in AI research and a February revision to its safety policy.

The revised policy separates safeguards Anthropic plans to pursue as a company from more ambitious recommendations for the wider AI industry. Anthropic said it could not commit to following those industry-wide recommendations on its own. Under its earlier policy, Anthropic said it would refrain from training or deploying a model that posed catastrophic-harm risks unless safeguards reduced those risks to an acceptable level.

Amodei acknowledged that Anthropic had made mistakes but maintained that competing to build leading models supports its safety mission. He described nine decisions that he said carried commercial costs. Among them were limiting access to Mythos-class models for several months and restricting biology-related requests.

The dispute also reaches beyond decisions about individual model releases. Anthropic’s Oct. 19 safety roadmap says policies focused on the deployment of dangerous models may eventually need to be accompanied by measures addressing the pace of AI development itself. That would place the speed of research alongside model safeguards in the policy debate.

Amodei has previously urged international safeguards against AI-assisted biological weapons. Anthropic plans to publish a policy framework on the pace of AI development by Jan. 15, 2027, or explain why it decided against publishing one.