Anthropic CEO Dario Amodei has proposed a three-stage AI safety framework, advocating for a controlled pace in developing frontier models to prioritize safety and alignment research. The company plans to grant third-party evaluators long-term internal access to its facilities, allowing them to monitor safety protocols, report incidents, and assess training processes.
Under this initiative, external evaluators will receive office space, security credentials, and access to internal tools comparable to those used by the company's own risk teams. While Anthropic may redact sensitive information related to legal privileges, customer privacy, or trade secrets, it cannot block the publication of findings simply because they are unfavorable to the firm.
The second phase of the proposal calls for AI companies in democratic nations to coordinate on common safety standards and curb unconstrained capability races. Amodei suggests that government support is necessary to navigate antitrust regulations, aiming to establish verifiable slowdown agreements once a sufficient number of firms adopt external oversight.
The third stage focuses on international cooperation, starting with common interests like prohibiting the use of AI in developing biological weapons. Amodei notes that current models are increasingly capable of assisting in the creation of next-generation systems, which could accelerate technical iteration beyond human control.
Amodei estimates that substantial progress in model interpretability, safety, and operational standards can be achieved within one to two years. OpenAI CEO Sam Altman and Elon Musk have publicly expressed support for the introduction of independent evaluators and the proposed approach to slowing down development.
