Anthropic proposes embedded safety assessors, but experts question independence and enforcement power

Anthropic CEO Dario Amodei proposed embedding independent third-party safety assessors within leading AI companies, granting them access to internal risk teams and the authority to publish findings. He cited banking regulators as a model, promising system access equivalent to internal teams. However, banking regulators have greater enforcement power, including the ability to halt operations or shut down institutions, according to Julie Andersen Hill of the University of Wyoming Law School. Experts like Albert Ziegler of XBOW noted that black-box testing may miss complex risks, and assessors lack veto power. Deborah Raji of UC Berkeley criticized the arrangement for potential conflicts of interest, arguing that METR, a potential assessor, has focused mainly on Anthropic and OpenAI, raising concerns about independence. Raji emphasized that true oversight requires an independent body to decide assessments and enforce consequences, not just access. Sarah Heck of Anthropic acknowledged that self-regulation through 'honorable codes' is insufficient, highlighting the need for external oversight.