Anthropic 执行长 Dario Amodei 主张政府应仿照银行监管机制，让独立评估员进驻前沿人工智慧企业以进行即时监控，而非仅在系统发布前作表面测试。这项提议借鉴了美国联准会、联邦存款保险公司等机构派驻专员深入银行日常营运与风险管理的成熟经验，能有效因应 AI 领域瞬息万变且具跨产业扩散性的特点。

在此监管架构下，具备技术专长且派驻于实验室的监管人员，能从模型设计初阶段即把关以防止危害行为，并监督各项压力测试，确保企业在未事先知悉的极端情境（如重要基础设施遭干预）下仍具备手动覆盖或关闭系统的能力。相较于过去监管社群媒体时因议题过于分散而停滞的困境，国会应设立专门的「人工智慧监督办公室」，聚焦于核心的模型安全规范。

为防止如过往金融监管般出现监管俘虏或流于形式清单的问题，AI 监管机构应由专业人员组成并推行更短期的轮调制度以维持超然独立。当前国会研议中的《前沿法案》虽提议由公司聘请持照评估员，但仍有评级机构利益冲突的隐忧，唯有建立如同银行的公权力常驻监管机制，才能兼顾 AI 的庞大效益与系统性风险控管。

Anthropic CEO Dario Amodei proposes that governments should formalize an oversight mechanism similar to bank supervision by embedding independent evaluators directly inside frontier AI companies for real-time monitoring rather than relying solely on pre-release evaluations. This approach draws on established practices from agencies like the Federal Reserve, where supervisory teams reside inside commercial banks to examine day-to-day operations and evaluate comprehensive risk management.

Under such a regime, technically proficient supervisors embedded within labs would ensure safety standards from the initial stages of model design and oversee mandatory stress tests, confirming that staff can manually override or terminate systems even in surprise crisis scenarios involving critical infrastructure. Learning from the stalled efforts to regulate social media due to overly broad agendas, Congress should focus immediately on model safety and establish an agency like the Office of AI Supervision.

To counter risks of regulatory capture and bureaucratic box-checking seen in financial history, an AI regulatory body must employ capable experts and enforce rapid staff rotations between companies to safeguard independence. While the bipartisan Frontier Act introduces licensed assessors, allowing companies to hire their own reviewers risks financial-crisis-style conflicts of interest, making embedded public supervision essential to fairly balance AI innovation with systemic safety.