Who Gets to Define Safety? A Systematic Review of How Generative AI Research Addresses Youth Online Safety

Paper Title

Who Gets to Define Safety? A Systematic Review of How Generative AI Research Addresses Youth Online Safety

Publication Info

• Topic area: Youth online safety in the context of Generative AI systems
• Keywords: Generative AI, youth safety, online risks, participatory design, AI ethics, human-computer interaction, developmental psychology, multimodal AI, risk evaluation, stakeholder inclusion

Background and Problem

• Problem / challenge: Existing safety frameworks for online youth interactions are inadequate for addressing the unique risks posed by Generative AI (GenAI), such as misinformation, emotional dependency, and biased outputs. Current research often frames safety as a technical feature rather than a relational or developmental concern.
• Significance: With GenAI systems increasingly embedded in youth’s digital lives, addressing safety is critical to prevent sociotechnical harms and ensure systems support youth’s developmental and emotional needs.
• Motivation and related work: Prior research on youth online safety has emphasized participatory approaches and contextual risk evaluations, but GenAI research predominantly focuses on technical metrics and lacks meaningful youth engagement. This paper addresses these gaps by systematically reviewing how youth safety is conceptualized, studied, and evaluated in GenAI research.

Solution

• Proposed approach: A systematic review of 30 GenAI studies (2014–2025) to analyze how youth safety is defined, what risks are prioritized, and how safety is evaluated across the GenAI lifecycle.
• Novelty:
  1. Systematic synthesis of gaps in risk definitions, evaluation practices, and stakeholder inclusion in GenAI research.
  2. Introduction of a working definition of youth online safety as a sociotechnical construct emphasizing developmental alignment, social awareness, and youth agency.
  3. Proposal of a multistakeholder, lifecycle-based participatory agenda for GenAI safety, advocating for interdisciplinary collaboration and youth inclusion.
• Procedure and key techniques:
  • Conducted a systematic literature review using PRISMA guidelines across five databases (ACM Digital Library, Web of Science, IEEE Xplore, ACL Anthology, arXiv).
  • Analyzed 30 studies using reflexive thematic analysis to identify patterns in stakeholder involvement, risk prioritization, and evaluation methodologies.
  • Adapted Roger Hart’s Ladder of Youth Participation to assess the depth of youth engagement in the research.

Results

• Concrete findings:
  • Most studies (N=26) were authored by AI researchers, with limited input from youth development experts.
  • Risks studied were predominantly informational (e.g., misinformation, bias), with less attention to interpersonal, psychological, or privacy-related harms.
  • Safety evaluations were largely reactive, occurring post-deployment, with minimal youth involvement beyond consultation.
• Advantage over baselines: Highlights the need for participatory, developmentally grounded approaches to safety, contrasting with the current system-centered, technical focus in GenAI research.
• Experiments / evaluation:
  • Youth participation was limited to pseudo-participatory roles in most studies (N=15), with only one study reaching shared decision-making (Stage 6 on Hart’s Ladder).
  • Formal contexts (e.g., education, health) dominated research settings, with informal spaces like social media and gaming underexplored.
  • Text-based GenAI modalities were overrepresented, while multimodal systems (e.g., image, audio, video) received limited attention.
• Limitations and future work:
  • Lack of youth-specific risk taxonomies and developmental frameworks.
  • Minimal exploration of informal GenAI interactions and multimodal risks.
  • Future work should focus on early-stage participatory design, longitudinal evaluations, and interdisciplinary safety frameworks.

Summary

This paper systematically reviews how youth online safety is addressed in Generative AI research, revealing significant gaps in stakeholder inclusion, risk conceptualization, and evaluation practices. Current approaches are predominantly technical and reactive, with limited attention to youth’s developmental and relational needs. The authors propose a participatory, multistakeholder agenda to redefine safety as a sociotechnical construct and advocate for youth-centered, developmentally informed frameworks. These findings highlight the need for interdisciplinary collaboration and proactive safety evaluations to ensure GenAI systems support youth well-being and agency.