OpenAI will embed six Apollo Research evaluators inside its organisation for four months to assess whether its AI models attempt to undermine alignment and related control measures during training, testing and internal use.
Both outlets describe the evaluators’ work as focusing on “subversive” or otherwise misaligned behaviour, with Apollo conducting an independent-style assessment and producing verdicts. At the same time, the reporting says OpenAI controls staff access and can redact some evidence, which may limit what Apollo can see or how fully it can disclose details.
TechCrunch emphasises potential questions about the evaluators’ independence given the access and disclosure restrictions. The Next Web foregrounds the mechanics of the arrangement, including the claim that while OpenAI can manage access and redact some material, it cannot prevent Apollo’s verdicts from being delivered or withheld solely because findings may be unfavourable. The two stories otherwise align on the core facts: the embedding, the duration, the number of evaluators, and the goal of checking for subversion during multiple stages of model use.