AI evaluation for children must measure development, not only harm
UNICEF has called for children to be placed at the centre of independent AI evaluation. It argues that assessment should account for age, language, culture, circumstances, repeated use and long-term effects, and should test whether AI supports dignity, learning, development and wellbeing rather than checking only for obvious harmful responses.
What it could improve
The approach moves AI governance beyond one-time technical tests. It supports independent evidence, child participation and continuing evaluation of how AI affects children over time. This is particularly important for conversational systems, learning tools and emotionally engaging services that may shape behaviour and relationships through repeated interaction.
What remains unresolved
The article establishes the right evaluation questions, but a voluntary evaluation ecosystem still needs common benchmarks, access to evidence, enforceable consequences and a reliable way to know when a child-appropriate response is required.
- A model may perform well in a laboratory while producing different effects across ages, languages, cultures, disabilities and real-world settings.
- Risks such as dependency, manipulation, distorted advice or developmental displacement may emerge only through repeated use.
- Independent evaluators need meaningful access to systems and evidence without exposing children’s conversations or identities.
- AI services may rely on a declared account age or have no reliable child classification at all.
- Provider-by-provider age checks can create repeated verification, inconsistent classifications and unnecessary identity exposure.
- Evaluation must cover beneficial outcomes, autonomy and participation as well as the absence of prohibited content.
The Child-Friendly Internet perspective
The Child-Friendly Internet view is that AI should adapt to the child’s developmental context and prove that it supports safety, dignity, learning and growing autonomy. Independent evaluation should happen before deployment and continue as systems change. Participating AI services could receive a short-lived, purpose-limited child assurance signal, without learning the child’s identity or linking activity across services.
- Evaluate complete experiences and repeated use, not isolated model answers.
- Benchmark performance across developmental stages, languages, cultures and accessibility needs.
- Require independent testing before launch and continuing monitoring after material changes.
- Use privacy-preserving assurance to trigger age-appropriate behaviour without recurring identity checks.
- Measure learning, wellbeing, agency and beneficial participation alongside harm prevention.
The central question
How can independent evaluation prove that AI helps children flourish while the service learns only the minimum information needed to respond appropriately?