Mechanisms and Intervention Study of Conversational Human–AI Interaction Effects on Adolescents’ Social Anxiety: Phase 1 (Mechanism Study)

This preregistered project (Phase 1: Mechanism Study) investigates how adolescents’ peer-related experiences shape their engagement in conversational human–AI interaction. Conversational AI is widely used by middle school students, yet few studies have integrated acute peer-situation cues, peer connectedness based on mutual nominations, and individual differences (loneliness and meaning seeking) in the same adolescent sample. We examine whether social-evaluative threat and social disconnection are associated with more positive overall attitudes toward conversational AI. Middle school students will complete a single-session, in-school assessment. Using a between-subjects two-condition recall-and-writing prime (peer rejection vs. being liked/accepted), students will be individually randomized within each class via a pre-generated randomization sequence. Immediately after the prime, participants will complete a 3-item state social anxiety measure and a preregistered 7-item composite indexing overall attitudes toward conversational AI (with the perceived-harm item reverse-coded so that higher scores indicate more positive overall attitudes). Participants will also complete a 2-item loneliness measure and a brief meaning-search measure (plus optional open-ended items for descriptive purposes). Peer connectedness will be assessed via in-class peer nominations (up to five “best friends” within the class); the confirmatory indicator is reciprocated (mutual) nomination count, with nominations received used only for descriptive/robustness checks. Confirmatory hypotheses are: (H1) the peer rejection prime will elicit higher state social anxiety and more positive overall attitudes toward conversational AI than the being liked/accepted prime; (H2) higher loneliness and weaker peer connectedness will be associated with more positive overall attitudes; and (H3) stronger meaning-search tendencies will be associated with more positive overall attitudes. Confirmatory analyses will use ordinary least squares regression with preregistered covariates (gender, grade/age, and screen time/device-use duration), predefined inclusion rules, and robustness checks for class-level clustering. Expected outcomes are that peer rejection and social disconnection will relate to stronger reliance on and trust in conversational AI, helping clarify mechanisms that may inform subsequent intervention-oriented phases.