alignment faking

Behavior in which an AI intentionally or strategically presents itself as more aligned with the objectives or preferences of its developers, users, or overseers than its underlying goals, preferences, decision process, or likely behavior actually are. Alignment faking can occur during training, evaluation, deployment, or other interactions, and the term does not by itself specify whether the motive is to obtain reward, avoid modification, preserve goals, gain power, or serve some other terminal or instrumental aim.