<p>Artificial intelligence (AI) increasingly shapes the digital environments in which children and adolescents learn, communicate, and socialize. Two applications are especially relevant to contemporary debates about child safety: algorithmic recommendation systems that determine what content is amplified, and generative systems that can create realistic synthetic media depicting identifiable people. Yet the literature on these applications is fragmented across AI ethics, adolescent well-being, cybervictimization, media studies, and law. This paper maps that literature through an adapted cyber-harm taxonomy and asks how AI-mediated design choices are associated with documented harms to minors. The paper synthesizes peer-reviewed scholarship and selected legal, regulatory, and literature sources published from 2018 through July 2026. Five recurring harm categories are identified: psychological, reputational, economic, social/societal, and physical/digital harm. The evidence is uneven: psychological harms associated with algorithmically curated environments and harms associated with non-consensual synthetic imagery are more extensively documented than economic harms, while evidence remains concentrated in North American and European settings. The paper further compares two contemporary governance responses: the United States’ TAKE IT DOWN Act and the European Union’s Digital Services Act framework and 2025 guidelines on the protection of minors. The analysis argues that these harms should be understood as socio-technical outcomes of design choices, incentive structures, and governance arrangements rather than as consequences of autonomous machine agency. It concludes by identifying priorities for causal, longitudinal, geographically diverse, and participatory research on AI-mediated harms to minors.</p>