When does AI assistance undermine relational authenticity? A mini review of conditional effects in AI-mediated interpersonal communication

Research on AI-mediated communication (AI-MC) has produced seemingly contradictory findings on how AI assistance affects interpersonal relationships. Disclosed use lowers trust, behavioral measures show null effects, and under realistic conditions receivers do not suspect use at all. This mini review integrates experimental studies conducted in interpersonal contexts and argues that these conflicting findings may be reconciled by a conditional-effects framework. Thirteen core experimental studies and four auxiliary studies from adjacent domains, identified through multi-database searches, were coded on four moderators and on the mediators they tested. Relational costs concentrate under four conditions: disclosure or suspicion of use, message emotionality, relational closeness, and level of assistance. Because the moderators are confounded across studies, the framework is offered as an organizing structure rather than a settled resolution; within-study evidence supports the moderators to differing degrees, with level of assistance the most consistently supported and relational closeness the least. Diminished perceived authorship, comprising a cognitive attribution component and a relational-obligation component, is proposed as a candidate mechanism; it is indirectly supported by the mediators tested to date but has not been tested directly. The review states falsifiable propositions for each moderator, flags anomalous findings, notes measurement problems, and outlines priorities for longitudinal, psychometric, and cross-cultural research.