Pre-Deployment Safety Framework for AI Tutoring Systems Serving Minors in Game-Based Learning Environments

The integration of AI tutoring into educational games presents unprecedented opportunities for personalized learning, yet introduces distinctive risks when serving minors (ages 13–17). Unlike traditional classroom AI tools, game-based AI tutors foster extended engagement, emotional investment in creations, and parasocial relationships with AI companions—amplifying psychological and privacy vulnerabilities. This paper proposes a comprehensive safety framework addressing the unique intersection of gaming dynamics, AI tutoring, and child protection. Drawing from regulatory requirements (COPPA 2025 updates, GDPR Article 8, EU AI Act 2024–2026), established ethical guidelines (UNICEF Policy Guidance v3.0, IEEE standards), and industry best practices (Khan Academy Khanmigo, Duolingo), we synthesize seven core principles: child rights prioritization, regulatory compliance by design, transparency and AI literacy, fairness and autonomy promotion, safety monitoring and crisis response, content and interaction safeguards, and oversight accountability. The framework employs a layered safeguards model—technical (RAG grounding, PII detection, injection protection), interaction (age-tiered modes, boundary prompts), and oversight (parental dashboards, auditing)—operationalized through a developer checklist suitable for resource-constrained independent developers. We demonstrate application through AMAIG, a water infrastructure engineering simulation serving learners ages 13–65+. The novel contribution lies in addressing the specific risks of informal, immersive gaming contexts that existing classroom-focused frameworks overlook. We argue that pre-deployment safety frameworks are essential for accountability amid rapidly evolving regulations, providing practitioners with actionable guidance for ethical AI tutoring in educational games.