Kwang Sik Jeong, Sarah Choi, Keon Chul Park, Journal of Internet Computing and Services, Vol. 26, No. 4, pp. 55-68, Aug. 2025
10.7472/jksii.2025.26.4.55, Full Text:  HTML
Keywords: AI Companion, Service Architecture, Harmful Language, Filtering, Age-Appropriateness

Abstract

This paper presents a novel AI companion service architecture that supports safe conversations for young children, emphasizing harmful language filtering tailored to children's developmental stages. The research examines developmental theories, including Vygotsky's cognitive development and Bandura's social learning theories, to understand how children process harmful language. Through analysis of six AI services (Amazon ALEXA, Google Assistant, Naver CLOVA, Embodied MOXIE, ROYBI Robot, Miko3), the study identifies limitations in existing systems regarding intent classification accuracy, age-appropriate filtering, and educational feedback. Based on these findings, our research proposes a service architecture for harmful language filtering with three main components: a context-aware language processing module that detects harmful intent beyond keyword matching; a developmentally appropriate classification module that adjusts filtering sensitivity according to age and developmental stage; and an educational feedback module providing explanations and alternative expressions rather than merely blocking inappropriate language. The architecture also comprises three additional components: user profile database, knowledge base, monitoring analytics. This approach addresses technical limitations while incorporating age and developmental considerations essential for child-oriented AI design. Furthermore, this study explores children’s perception of harmful language across developmental stages, providing empirical insights to guide adaptive filtering strategies. The proposed architecture is expected to enhance not only safety but also educational value in AI-mediated interactions, with implications for ethical AI design in child-oriented services.

Statistics
Cite this article