Anthropic CEO Proposes "Bank-Style Regulation": Experts Say Evaluators Lack Authority to Halt AI; You Cannot Be Both Player and Referee

Anthropic CEO Dario Amodei recently proposed a regulatory framework for frontier AI companies, suggesting the introduction of an "embedded supervision" model similar to the banking industry. The proposal advocates for independent third-party safety evaluators to be stationed long-term within companies, granting them access close to internal risk teams, and allowing them to independently publish their findings with limited redactions.

With the rapid iteration of large model capabilities, AI safety issues have become an industry focus. Previously, Anthropic researchers resigned due to concerns that frontier labs were racing to develop systems that could go out of control. Amodei's proposal reflects leading AI companies' attempts to alleviate public and regulatory concerns about AI risks by introducing external oversight mechanisms while driving technological development.

Although the proposal aims to strengthen AI safety reviews, several banking regulatory experts and AI evaluation practitioners point out its core flaw: the lack of enforcement power. Unlike bank regulators, these AI evaluators cannot stop the training or release of models, making it difficult for them to truly act as "referees." This discussion will prompt the industry to further consider how to establish an AI regulatory framework with substantive binding force.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.