OpenAI’s usage analysis included 1,819 scientists who had sent a scientific message and had enough observation time. The most active tenth accounted for 42% of scientific message activity after adjustment for observation length. The frequently cited time estimate came from a different instrument: 251 respondents who consented to analysis estimated how long the same work at the same quality would have taken without AI. Their estimate was about 1.9 times as long. A lottery allocated workspaces; it did not randomly assign identical tasks to AI and non-AI arms.
OpenAI’s separate external-assessment agreements describe access, not findings. METR’s initial scope covers one selected frontier reinforcement-learning run; Transluce will examine post-training evaluations. The Safety and Security Committee receives unredacted reports. What becomes public can be narrower because of confidentiality and responsible-disclosure provisions.
Sources: openai.com · cdn.openai.com · openai.com · openai.com