AI 2027 author Daniel Kokotajlo tweets message from current OpenAI capabilities researcher, Dan Selsam, on AI risk. Gives some insight into why some AI researchers may be freaking out: increasing model situational awareness during alignment evaluations

Daniel Kokotajlo, author of "AI 2027," shared a message from OpenAI researcher Dan Selsam regarding AI risk. Selsam highlights a growing dependence on AI models among researchers, who are increasingly offloading cognitive tasks to them. This trend, already prevalent in AI labs due to substantial internal token subsidies, could spread globally, potentially leading to a civilization entirely modulated by AI. This shift, while seemingly positive and possibly coinciding with a renaissance, raises concerns about human researchers losing ownership and critical engagement with model-driven research.

This report uniquely features an OpenAI capabilities researcher's direct account of increasing AI dependence within research, unlike typical discussions that focus on external AI risks.