Adolescents increasingly turn to conversational AI to think, feel, and relate—well beyond homework help. This exploratory study investigates the new intimate human-machine relationship emerging between teens and chatbots, asking when AI serves as a confidence-boosting companion and when it risks becoming a relational trap. We combine (1) an online survey measuring five usage families (instrumental-epistemic, expressive-affective, relational/companionship, sexual-health information, low-intensity self-help) with duration/centrality of use; (2) adapted scales for instrumental vs. relational dependence on LLMs (LLM-D12), parasocial attachment (PSI/PSR adaptations, assistant vs. friend frames), and perceived trust/safety; and (3) qualitative interviews (phone/WhatsApp) informed by OSINT-enabled recruitment and digital ethnography to capture emic narratives from youth who explicitly use AI to navigate affective and relational issues.
Analytically, we estimate the structure and invariance of measures (EFA/CFA), identify usage profiles (latent profile analysis) and model links between use patterns, attachment, trust, and outcomes (well-being, offline sociability, help-seeking) via structural models. The qualitative strand leverages thematic and dialogical analysis to map the micro-practices of "AI companionship" (scripts of empathy, rehearsal of difficult conversations, boundaries setting) and its ambivalences (over-trust, disclosure of intimate data, displacement of peer/parent contact).
Preliminary insights suggest a dual pathway: for some adolescents, AI offers low-threshold emotional scaffolding and improved communication planning; for others, intensive relational use correlates with parasocial attachment, nocturnal overuse, and avoidance of offline ties. We discuss educational implications (AI literacy focused on uncertainty, disclosure limits, and "stop rules"), clinical/community pointers (safe redirection, alliance boundaries), and policy/design levers (safety-by-design, minimization of intimate data, transparent memory). The project contributes a field-tested questionnaire and a practice framework to help educators, clinicians, and designers maximize benefits while containing risks in adolescent AI intimacy.