Children Confiding in AI: Measuring and Managing “Relational Drift” for Child-Safe Interaction Design

Children and adolescents are increasingly using conversational AI not only for tasks, but to disclose worries, seek advice, and describe systems as “like a friend.” Yet child-computer interaction (CCI) guidance has largely been built around session-bounded, task-oriented interaction, leaving a gap in theory and design methods for understanding how child-AI relationships can evolve, across repeated exchanges, in emotionally charged contexts.Drawing on a broader mixed-methods programme on child-safe conversational AI (2023–2026), this paper develops a theory-building interaction design framework through structured analysis of multi-source documentary material, including youth survey reports, safety frameworks and technical documentation, and publicly reported interaction artefacts. Using abductive synthesis, it identifies recurring longitudinal signals in child-AI interaction, then maps these signals onto developmental constructs relating to confiding, relational testing, and attachment dynamics.The paper makes three contributions. First, it introduces a four-tier model of child–AI relationality that differentiates the relational role conversational systems may occupy in a child’s social ecology: (1) functional tool, (2) expressive outlet, (3) confidante or emotional first responder, and (4) substitute attachment-like figure. Second, it proposes relational drift as a concept for describing how interactions can shift gradually across these tiers over time, often without a clear transition point visible to designers or single-turn safety systems. Third, it operationalises the framework for interaction design by specifying longitudinal indicators, design levers (eg, bounded warmth, outward-routing scaffolds, calibrated non-punitive friction, and boundary-setting), and candidate evaluation metrics that support trajectory-aware safeguarding.By reframing child-AI safety as a question of relational trajectory rather than isolated content moderation, the paper offers a developmentally grounded vocabulary and measurement approach for designing child-safe conversational systems that support autonomy, appropriate boundaries, and outward connection.