OpenAI's "independent assessment" doesn't look independent at all

Mother Jones
OpenAI's new safety reviews by four outside labs are compromised by limited access and OpenAI's control over which staff and information are shared.

Summary

OpenAI announced that four independent assessors—METR, Apollo Research, Transluce, and Goodfire—will evaluate its AI models and safety features. However, the assessors face significant limitations: OpenAI approves team members, selects which staff they can interview, and can withhold information for reasons including "technical infeasibility, security risks, or IP protection." Each assessor receives different levels of access and time; for example, METR's prior investigation was limited to six on-site days, and its report noted heavy reliance on AI agents with "significantly worse judgment and reliability than human researchers." Apollo's four-month assessment can be curtailed by OpenAI, Transluce can only access pre-approved materials, and Goodfire's "deep access" remains undefined. Critics note that these are not truly independent investigations, as OpenAI retains final say over scope and access, contradicting CEO Sam Altman's earlier commitment to "employee-like access." The article concludes that while the announcement signals some transparency, it falls far short of genuine independence.

(Source:Mother Jones)