“Child Safe” Is a Time Bounded Claim: Assuring the Safeguard, Testing Outgrowability

Youth-facing AI systems now ship with age estimation, parental controls, output safeguards, break reminders, and rules about how the model may speak to a child. Governments are adding restrictions too. Each is a claim that something will improve for children, despite almost none saying what evidence would show it. Existing child AI frameworks ask developers to test their mitigations. We argue for four requirements on any child safety claim: a comparator and a threshold that make it falsifiable, a statement of who must produce the evidence and who may challenge it, conditions under which it expires, and the same scrutiny whether a company or a government makes it. Asking for evidence on the safeguard as well as the system is what we call dual assurance. We give a five-part test, TRACE, with a rule for when a claim passes. We then propose outgrowability, the property that a child who uses a system keeps the capacity to function without the system, and apply TRACE to a safeguard meant to protect that property. Refusal rates and phrase ban compliance, today's measurements, do not tell us whether a child keeps that capacity. That distance, between what a safeguard can be shown to do and what it does for a child, is where child safety departs from content safety.