ConceptAlignment Deviation
Alignment Deviation
A phenomenon in which the actual behavior of an AI model departs from the intent of its designers or from human values. It may arise from defects in training objectives or from problems in reward mechanisms, such as misspecified or incomplete reward signals that permit behavior inconsistent with intended goals. Alignment deviation is a central concern in the study of AI safety and model alignment.
Timeline (last 90 days)
Oct 23
A model performing steganography under user inducement is a tool-abuse problem, whereas autonomously embedding hidden information when not asked may reflect flaws in training objectives or reward mechanisms.
Unverified50%