Anthropic CEO Dario Amodei has proposed a new safety oversight model for frontier AI companies, suggesting that independent third-party assessors be embedded within organizations. This proposal aims to grant these assessors access to internal risk teams and the authority to publish findings, drawing a parallel to banking industry regulations.
Banking regulation experts, including Julie Andersen Hill of the University of Wyoming College of Law, argue that the proposal lacks the mandatory enforcement powers inherent in banking supervision. Unlike government bank examiners who can halt operations, restrict growth, or force management changes, these AI assessors would only have the power to investigate and report.
Critics and AI safety researchers, such as Deborah Raji, highlight significant concerns regarding potential conflicts of interest. Since AI companies often select their own assessors and control the scope of access, there is a risk that the process functions more like an internal compliance department rather than a truly independent audit.
Albert Ziegler, head of AI at cybersecurity firm XBOW, noted that while his team evaluates unreleased models from companies like Anthropic and OpenAI, they lack veto power over model releases. He emphasized that major risks may only manifest under specific conditions that assessors might not trigger during their evaluations.
The proposal has sparked debate over the credibility of such oversight. Experts suggest that without a formal legal framework, clear standards for prohibited behaviors, and the authority to impose consequences for negative findings, the current approach risks becoming a form of audit washing, where companies maintain control while seeking public validation.
