Aligning to a Moving Target: When AI Gazes into the Human, the Human Also Adjusts into AI

When an AI is fitted to a human, interaction with that AI can change the human’s preferences, capacities, and expectations, and the human to whom alignment refers is then no longer fixed. This paper asks under what conditions judging alignment successful by reference to a changed human is legitimate. Part 1 sets out a three-stage structure, names the phenomenon in which the criterion of success moves inside the interaction Endogenous Alignment Criterion Shift, and proposes four diagnostic items that separate recursion from harm to safety: the independence and legitimate updating of the reference criterion, the effectiveness of independent verification, the substantive preservation of options, and the possibility of exit. Part 2 draws on a Japanese national survey (N = 11,939), examining how the association between emotional dependence on AI and loneliness varies by social support. Contrary to the prediction, this association was stronger among respondents with more support. Some who identified an AI as a best friend or partner did not affirm that any current AI was conscious. Part 3 places legitimacy in the process of updating, specifies five review conditions, reads the diagnostic items as design requirements for coexistence environments, and proposes a longitudinal design.