OpenAI's "independent assessment" doesn't look independent at all
OpenAI’s claim of an independent assessment of its internal AI safety is questioned as the four engaged labs will still face restricted access and OpenAI’s control over who can be examined and who can speak to staff, undermining true independence. The arrangement has drawn critique from researchers who note that the evaluators lack employee-level access and that OpenAI can selectively disclose information, limiting the study’s rigor. The reported findings from METR’s on-site work after the July Hugging Face incident highlight methodological challenges, such as reliance on AI agents for analysis and gaps in understanding multi-agent dynamics. The setup appears to compromise transparency and public accountability, contrasting with OpenAI’s stated goal of broader industry openness. The piece suggests that real accountability requires broader access, consistent terms for assessors, and unfiltered visibility into models and code—steps that remain uncertain.
Dive Deeper:
OpenAI announced that four independent assessors—METR, Apollo Research, Transluce, and Goodfire—would evaluate internal models and test safety features, but the access and collaboration terms were still controlled by OpenAI, raising questions about true independence.
METR’s team, including two METR staff members and a nonprofit collaborator, conducted on-site work for six days total after the Hugging Face incident, reporting that analysis often depended on AI agents and lacked reliable human oversight for complex multi-agent scenarios.
A METR-led critique noted gaps in how to understand and oversee AI activities when multiple agents interact, with developers’ outputs sometimes missing details, overconfident, or hard to interpret, undermining the rigor of the assessments.
OpenAI indicated it would allow METR to proceed with agreed but undefined access and safety arrangements, and noted that the scope would be less expansive than METR’s own alignment proposal, signaling potential constraints on the depth of the evaluation.
Apollo Research will work with OpenAI for four months, but OpenAI reserves the right to withhold information and to select which staff assessors can speak with, introducing potential information asymmetry.
Transluce and Goodfire face similarly limited access: Transluce can review models and safety docs only with pre-approved researchers, while Goodfire promises deep access to internal systems but without clear definitions of what that entails.
Overall, the arrangement is framed as a collaboration toward shared standards, but critics argue it falls short of independent verification since OpenAI maintains significant control over scope, access, and disclosure, delaying genuine transparency.