AI and children's rights: A narrative review of evidence

Research shows that harms to children are present across the lifecycle an AI system, from design and training data through to deployment.

• Bias recurs across race, gender, age and income wherever systems go unexamined for children, deepening inequality within and between countries, including for children seeking refuge.
• Generative AI has put harmful tools within easy reach, from nudification to peer and adult cyberbullying, while children's images and biometric data are collected and reused to train models.
• Consent frameworks built for adults rest on parental authority and overlook children's evolving capacities, and permissions given in early childhood follow children through the systems they meet for life.
• Companion apps and chatbots use emotionally adaptive, attention-seeking design that can support lonely children but they also foster dependency, manipulation and a loss of agency.
• Automated systems in social care and benefits (in the UK and the Netherlands), and AI-assisted targeting that has killed children in Gaza, means that children who never use AI directly still face life-altering harms.
• The water, energy and mineral extraction behind AI falls hardest on the communities and children closest natural resources.

Rights-respecting practice is emerging, giving states and businesses something concrete to build on.

• Accountability and transparency tools are maturing, from proportionate, accessible data protection to risk-based assessment of where AI should be used, limited or refused.
• Governance is beginning to account for cultural and community context, pairing individual data protection with community-oriented safeguards.
• Child rights impact assessments and human rights due diligence turn broad commitments into something auditable, requiring attention to accessibility, affordability and the needs of disabled, remote and marginalised children.
• Consulted meaningfully, children can identify risks that adult designers miss. They propose their own recommendations for governance, and their expectations of AI are mixed, holding both hope and concern.

• Governance stays reactive where it needs to be preventive: impact assessments and due diligence belong before a system reaches children, alongside clear routes to remedy and redress.
• Frameworks default to adult baselines and rarely embed child-specific risk assessment, while age assurance can be privacy-invasive, easily bypassed or harmful to children's best interests.
• Participation is affirmed in principle and neglected in practice; it needs mandated mechanisms, resources and auditable design requirements rather than one-off consultation.
• Critical AI literacy, and a say in how markets shape technology, remain out of reach for most children worldwide, widening the skills gap.
• Commitments to train officials, professionals and developers rarely come with evaluations of whether training changes practice, and resources seldom reach the jurisdictions that need them most.
• The evidence base itself is partial: it is mostly short-term and drawn from the Global North, leaving policy exposed to fear and hype without sustained, independent, child-centred research.