Yaman Yu's Dissertation Defense

Generative AI is becoming part of youths' everyday digital lives, supporting schoolwork, creative expression, social interaction, emotional support, and AI companionship. Yet youth face risks that are not fully captured by existing online safety frameworks, AI risk taxonomies, or general-purpose moderation systems. Because generative AI produces dynamic, personalized, and socially responsive content, risks often depend on context, emerge through interaction, and intersect with youth’s developmental needs, family relationships, and social environments.

This dissertation develops a youth-centered approach to understanding, evaluating, and designing for generative AI safety. Across four studies, I first examine how parents and teenagers perceive generative AI safety, revealing gaps between parental assumptions and teens’ actual uses, as well as challenges in mediation, monitoring, and communication. I then introduce Youth-Centered GAI Risks (YAIR), an empirically grounded taxonomy that identifies developmentally relevant risks and interaction pathways in youth–GAI use. Building on YAIR, I present YouthSafe, a youth-centric safety benchmark and safeguard model that detects subtle, context-sensitive risks often missed by existing moderation systems. Finally, I examine safe AI companions for youth through parent and expert perspectives, showing that safety judgments depend on youth maturity, AI character design, interaction patterns, youth agency, and family values. This dissertation contributes empirical evidence, a risk taxonomy, a benchmark and safeguard model, and design principles for building generative AI systems that are safer and more developmentally appropriate for youth.

Questions? Contact Yaman Yu.