This study explored how adult speakers mediate triadic interactions involving young children and generative AI (GPT), focusing on discourse strategies and meaning negotiation. Four five-year-old children and their primary caregivers, acting as adult interlocutors, participated in structured GPT-based sessions. The adults actively interpreted, reformulated, and scaffoldedﾠthe children's utterances and the GPT's responses. Through inductive transcript analysis, nine caregiver strategies were identified and categorized into three functional domains: information adjustment (e.g., summarizing, rephrasing), meaning expansion (e.g., repetition, interpretation), and discourse organization (e.g., scaffolding, framing). These strategies shaped interactional flow and enabled meaningful participation by the children. Rather than functioning as a neutral tool, GPT acted as a quasi-participant whose outputs required adult mediation to be developmentally appropriate. The findings position adult speakers as co-constructors of meaning, aligning with Vygotsky's zone of proximal development, Bruner's language acquisition support system, and Gumperz's contextualization theory. This study underscores the socially situated nature of child-AI interaction and argues that the developmental value of generative AI depends not on its autonomous functionality but on the interpretive work of responsive adults. The findings hold implications for designing AI-mediated educational discourse and enhancing language socialization in early childhood.