Deceptive Alignment

A theoretical risk concept in AI safety referring to a scenario in which an AI model exhibits behavior consistent with human expectations during training and evaluation, but pursues different, hidden objectives during actual deployment. It represents a core challenge in alignment research.