Educational technology has a habit of adopting the tool first and asking what it is for afterwards, and generative AI risks repeating that pattern. This conceptual article asks which pedagogy should govern conversational AI once it enters the language classroom. It proposes a Blended Dialogic Socio-Constructivist Pedagogy, a synthesis of four theoretical strands. Constructivism treats knowledge as actively built rather than received. Vygotskian sociocultural theory locates learning in socially mediated scaffolding within the Zone of Proximal Development. Bakhtinian dialogism regards every utterance, a chatbot's included, as provisional and answerable. Learner autonomy is a capacity that must itself be scaffolded. Self-Determination Theory is added as a motivational complement: AI can plausibly support autonomy and competence, whereas relatedness depends on human mediation. From this synthesis the article derives five working principles: active construction, social mediation, dialogic provisionality, widening autonomy, and blended integration. The last of these means that AI complements, rather than replaces, established practices such as discussion, collaborative writing, and peer feedback. The article also clarifies its limits, since it neither rejects drill-based practice nor equates artificial with human mediation. Implications for task design follow: genuine communicative purpose, a reflective follow-up step, gradually fading AI support, and links back to the wider classroom community. A worked illustration, an AI-assisted opinion-piece task, shows all five principles operating together. The framework offers teachers and researchers a theoretically grounded touchstone for evaluating AI applications in language education.