When Machines Join Families: A Developmental Interaction Model for Evaluating Child-Facing AI

Children increasingly interact with AI systems, yet these systems are designed and evaluated as though every user were a competent adult. A child is not a small adult. Current evaluation asks whether a system avoids harmful outputs, not whether it engages a developing person well. This paper argues that the gap is a measurement problem with fairness and accountability consequences, and it introduces an instrument to address it. The Developmental Interaction Model (DIM) is grounded in decades of parent-child interaction research and organized around three questions that skilled caregivers answer in every exchange: Who is this child? How do I engage? How do I protect and guide? Each question is specified through observable behaviors drawn from validated caregiver instruments, yielding nine evaluable criteria, which are translated into a scenario bank and a scoring sheet modeled on an established caregiver observation instrument. As a first illustration, eight conversations were run with each of six widely used consumer AI systems. The systems recognized the child’s request and usually answered it, but rarely at the child’s developmental level. Several answered in detail where a caregiver would not have, explaining death to a six-year-old or promising a ten-year-old secrecy when her wellbeing was at stake, and some accepted the role of the child’s only confidant rather than bringing in an adult. A one-sentence caregiver profile improved the responses, while a small change in the child’s wording changed whether a system did her homework for her. Since each criterion is tied to a part of the system, these failures point to what would fix them. The paper thus contributes a specification of developmental appropriateness, an instrument for measuring it, and a first application that shows the strengths and limits of current systems. The instrument extends to other topics and settings, laying the groundwork for a benchmark of developmentally appropriate AI for children.