Beyond "I Can't Help With That": How Child Safety Experts Evaluate AI Chatbot Safety

Key Findings

• — Current child-safety evaluations rest on unvalidated assumptions about appropriate output, with refusal treated as a safe default; practitioners disputed that assumption
• — Evaluations that focus on adversarial prompt detection or surface-level output harms can miss responses that harm youth in practice
• — 19 practitioners working directly with vulnerable youth (social workers, therapists, psychologists) reviewed chatbot responses to risky situations drawn from prior empirical work
• — Practitioners distinguished chatbot behaviors likely to cause harm from behaviors that could meaningfully support youth in difficult moments, and set out what role chatbots should and should not play
• — The paper recommends incorporating practitioner perspectives into child-safety evaluation design and supporting infrastructure

Methodology Notes

Qualitative interview study, n=19 practitioners; stimuli were chatbot responses to youth-risk situations established in prior empirical work. arXiv 2608.07902, submitted 2026-08-08, cs.CY (also cs.AI, cs.HC). The abstract page records acceptance at the Ninth AAAI/ACM Conference on AI, Ethics, and Society (AIES 2026); the proceedings version is not yet published, so this row is the preprint. Author affiliations are not stated on the arXiv abstract page. Small purposive practitioner sample; findings are qualitative and not intended as prevalence estimates.

Sources