Apollo Research and OpenAI launch an embedded evaluation pilot, with a six-person team stationed for four months to evaluate model alignment
A Pilot Towards Embedded Evaluations
Apollo Research announced the launch of an embedded evaluation pilot with OpenAI; a six-person team will be stationed at OpenAI for four months to evaluate model alignment during training, evaluation, and internal deployment. The pilot examines whether models attempt to subvert the alignment and control mechanisms of themselves or successor models, covering four aspects: whether training incentivizes subversion, whether evaluation can elicit subversion, whether subversion occurs in internal deployment, and whether the relevant reasoning is detectable. Apollo proposed three rights that evaluators need: access, disclosure, and meta-transparency, and regards meta-transparency (the right to publicly explain evaluation conditions and limitations) as the minimum standard for all evaluation collaborations. OpenAI committed not to withhold information, refuse employee access, or demand deletions or changes because information or findings are unfavorable, and not to delete or alter Apollo's conclusions on each claim, but may withhold some in-scope information due to technical infeasibility, safety risks, or intellectual property protection; Apollo will publish two public reports, with unredacted versions submitted to OpenAI's Safety and Security Committee (SSC) of the board.
The article proposes three rights that evaluators need—access, disclosure, and meta-transparency—and sets out the specific limitations of this pilot, which can serve as a reference for judging the credibility of third-party evaluations.