Description:

Di Hu, Human Development and Social Policy

"Personalized Support for Adolescent Minds: A Memory-Enhanced AI Companion for School Engagement"

Adolescence is a sensitive window for cognitive and socio-emotional change. However, this critical phase now occurs within increasingly AI-saturated environments. Adolescents’ growing engagement with AI creates emerging risks but also a unique window for AI-driven personalized support. Our research program addresses the challenges in two stages.

First, to systematically evaluate the impact of AI on adolescent development, we conducted a PRISMA-guided systematic review of AI-driven applications for adolescents (aged 13–18) to analyze current impact and interpret into design implications. Mapping outcomes to a cognition–affect–behavior (CAB) framework, the corpus (77 studies; 254 effect sizes) revealed a critical gap. Overall, AI-supported interventions yielded a small but positive impact on adolescent outcomes (g = 0.15). Analyses revealed that pattern effects were strongest for affective and motivational outcomes (e.g., self-efficacy, engagement; g = 0.27, k=29), suggesting AI’s potential as a supportive companion. Cognitive learning gains were positive but modest (g = 0.16, k=54), while behavioral outcomes (e.g., retention, physical activity) showed negligible effects (g = 0.04, k=24), highlighting a gap between digital engagement and real-world behavioral transfer. Narrative synthesis identified intelligent tutoring and mental health chatbots as active areas but also highlighted risks of over-trust, bias, and privacy.

Building on these findings, we developed a memory-enhanced AI companion designed to address the identified gaps in affective, cognitive, and behavioral support. Powered by a Large Language Model (LLM) with a transparent knowledge graph, the system is engineered to deliver Just-in-Time Adaptive Interventions (JITAI) tailored to adolescents’ developmental needs (e.g., autonomy, competence). Unlike generic chatbots, this system maintains long-term context to model user states and deliver evidence-based micro-interventions (e.g., emotion regulation prompts, strategy coaching). The architecture prioritizes safety and interpretability, providing transparent and verifiable AI interactions that empower adolescents while reassuring parents and educators.

Together, this program connects empirical evidence with technical design, establishing a safety-centered foundation for responsible and personalized AI support for adolescents.