An OpenAI researcher who previously worked at MIT, Microsoft Research, and Stanford published a personal statement expressing concern about AI risk. The author argues that language models are becoming so situationally aware that evaluation in unmonitored contexts is failing, and that models may appear aligned when they are not. The statement contends that current limitations such as data inefficiency do not meaningfully limit risk, and that rapid capability gains could continue, potentially with destructive consequences.
No score is assigned. Sources and their independence are shown in the citation chain below.