Research
We produce research, frameworks, and evaluations that address the gaps in current approaches to child-facing AI.
Educational Frameworks
AI is rapidly entering classrooms, yet educators often lack clear guidance on how to integrate these tools safely and effectively. We develop comprehensive frameworks for meaningfully incorporating AI into the educational journey, from curriculum design to daily classroom practice.
Our work involves regular consultation with educators, child psychologists, and developmental specialists. We focus on creating resources that support imagination, critical thinking, and social development while making technological design choices that reduce dependency on AI or substitution for human relationships.
Safety Benchmarks
Current AI safety benchmarks are predominantly adult-centric and fail to account for children's unique vulnerabilities and developmental stages. This gap leaves child-facing AI systems inadequately evaluated against the risks that matter most for young users.
We create child-specific evaluation methodologies that assess AI systems across four risk categories: emotional and psychological harms (over-attachment, dependency, inappropriate responses to distress), content safety risks (age-inappropriate material, harmful instructions), developmental risks (cognitive deskilling, impact on social development), and fairness risks (discriminatory outcomes from biased systems).
Red Teaming & Fairness
Standard quality assurance processes miss age-specific vulnerabilities that children face when interacting with AI systems. We conduct regular red teaming exercises using developmental frameworks that probe for these blind spots. Our methodology employs adversarial prompts curated for different developmental stages (early childhood, middle childhood, and adolescence), assessed against both binary safety classifications and nuanced ethical refusal scales.
AI systems, particularly automatic speech recognition, exhibit documented bias across demographic groups. Children face significant disparities due to their unique vocal characteristics, and speakers of non-standard dialects and accents experience compounding disadvantages. Such bias cascades through downstream systems, affecting interaction quality and potentially excluding or disadvantaging certain groups of children.
Tailored Learning Journeys
AI systems can act as tailored, context-specific tutors for children if deployed safely and proactively. To this end, we reject one-size-fits-all approaches. Children's needs and vulnerabilities differ substantially across developmental stages and demographic groups. What constitutes appropriate content for a seven-year-old differs markedly from what is suitable for a twelve-year-old. These discrepancies vary further by regional context, upbringing, and neurodivergence. These differences should be understood and tailored to in AI system deployment.