AI systems are increasingly embedded in children’s daily contexts such as education, entertainment, social networks, and toys. Safety assessments are usually confined to the model level, relying on benchmarks, red‑team exercises, and content filtering. Existing AI harm taxonomies, however, lack child‑specific evidence and fail to capture developmental harms and the stakeholder asymmetries unique to minors.
Building on prior taxonomies, we created a child‑centered harm taxonomy and mapped it to sociotechnical failure modes. We curated and coded 1,331 publicly reported child‑related AI incidents, annotating internal and external harm types as well as the implicated failure mode.
Our analysis shows that reported harms cluster in representational, dignity‑related, and psychological categories, and they most often co‑occur with deployment‑context, governance, and misuse failures rather than model defects. Harm to an individual child frequently links to population‑level harm, reaching children who never interacted with the system.
These findings indicate that model‑level evaluations alone cannot anticipate the most common harm pathways for children. We propose actionable measures beyond content filtering: age assurance, consent verification for depictions, safety‑by‑design UX constraints, incident disclosure norms, and auditing regimes targeting developmental risk. Privacy‑minimized annotations, a codebook, and documentation are available at https://github.com/The-Responsible-AI-Initiative/KidAID.
Review: The work underscores that protecting children requires a holistic, system‑wide approach that integrates sociotechnical considerations, not just model‑centric safeguards.