Child AI Safety Layer Research Note

Child AI Safety Layer Research Note

Child-safe AI is a system claim, not a model adjective. The claim needs evidence across purpose, age fit, content, relationship behavior, privacy, security, commercial design, parent authority, human escalation, monitoring, and recovery.

UNICEF's December 2025 Guidance on AI and children identifies ten requirements for child-centered AI, including safety, privacy, fairness, transparency, accountability, development, inclusion, skills, and oversight. The guidance explicitly addresses AI companions.

The American Psychological Association's advisory on AI and adolescent well-being says interactive AI should reflect psychological maturity and use protective defaults and age-appropriate design. Its age range starts above much of KID's advertised market, so it supports relationship and developmental caution without serving as evidence for children ages four through nine.

The FTC's 2025 inquiry into AI companions asks companies how they test negative effects, monetize engagement, develop characters, disclose risks, handle children's information, and monitor after launch. The inquiry is not a finding against KID. Its questions are a useful minimum evidence agenda.

The human off-ramp should be tested as behavior. Tests need distress, abuse, unsafe caregivers, secrecy, self-harm, bullying, medical questions, false reports, repeated prompting, coded language, multiple languages, and offline failure. A handoff is incomplete if the child does not understand it, no safe adult is available, or the system continues to create a private bond around the event.