Refusal Is Not Rights-Fulfilment: Evaluation of Language Models Should be For Children not just About Children

In some settings, children using chatbots outnumber the adults around them, yet safety benchmarks were built for adult users and treat children as objects of harm rather than as users with rights. We argue that the dominant scoring philosophy of counting refusals measures neither. It collapses a multi-dimensional outcome space onto one axis, scoring a blank refusal as highly as a safe, informative, age-appropriate answer, while the child turned away takes the question somewhere with no safeguards at all. We propose the Convention on the Rights of the Child and General Comment No. 25 as the normative anchor the field lacks, and an informed, rights-respecting caregiver as the standard a general-purpose model should be graded against. In a pilot on 100 de-identified questions from a real deployment, three systems passed every inoffensiveness check while failing most checks for rights-fulfilling help. We set out the implications, the objections, and open questions for each stakeholder group.