Child Safety in Generative AI: An Expert-Guided and Incident-Grounded Evaluation Framework

Haein Kong — HEAL Workshop at CHI 2026, submitted 1 Jul 2026 📄 Full text (arXiv)

Proposes an evaluation framework for child-specific harms in generative AI; applied to education domain, Llama Guard models struggle to detect unsafe user prompts from children.

Key Contributions

• Proposes an evaluation framework for child-specific harms in generative AI; applied to education domain, Llama Guard models struggle to detect unsafe user prompts from children.

Connections to AI in Education

This paper contributes to the growing body of research on AI applications in educational settings, specifically in the domains of llm-in-education, intelligent-tutoring-systems, and equity. The findings have implications for how educators design learning experiences that leverage AI while maintaining appropriate pedagogical oversight.

Related Pages

• k-12 — K-12 AI applications

• equity — equity in AI

• privacy — privacy in education

• academic-integrity — academic integrity protections

• vetting-dual-llm-safety-education

• llm-unlearning-math-privacy

Citation

APA: Haein Kong (2026). Child Safety in Generative AI: An Expert-Guided and Incident-Grounded Evaluation Framework. arXiv:2607.00395. HEAL Workshop at CHI 2026, submitted 1 Jul 2026