Observed facts: CNBC reports that METR, Apollo Research, and Transluce are among small groups being asked to evaluate model capabilities and risks. It reports that funding, model access, and reporting structure remain unsettled. CNBC also says OpenAI stated it was finalizing contracts with third-party assessors; two former employees alleged their dismissal related to evaluator communications, a characterization OpenAI disputes.
Mechanism: Evaluation requires access and resources, while labs and infrastructure partners that evaluators scrutinize may also be sources of support. Governance terms shape whether evaluators can independently scope and report findings. A nominally external review can be weakened if the sponsor controls scope, evidence access, or publication.
Why now: Why now: CNBC reports that voluntary external assessment is receiving renewed attention as federal oversight remains limited and labs face pressure to demonstrate safety practices. OpenAI’s reported contracting work makes the question immediate, while the dispute over employee communications shows that boundaries around engagement are already consequential.
Implications: Evaluator independence depends on concrete funding, access, reporting, and staff-engagement protections, not labels alone. A workable system may require diversified funding and explicit protections for access and reporting, but the source does not establish that such safeguards have been adopted.
Evidence boundary: Evidence boundary: this note relies on CNBC’s reported account and attributed statements; it does not independently verify the allegations about employee dismissals or establish the final terms of any evaluator contracts.
Watch: Watch whether forthcoming contracts specify funding sources, scope and continuity of model access, publication or reporting rights, and safeguards for staff engagement with external evaluators.