Background: AI-supported classrooms increasingly depend on chatbots for feedback, explanation, and language practice, yet multilingual Indonesian learning environments raise unresolved questions about whose language, meaning, and cultural norms these systems can reliably understand. Objective: This study examines multilingual reliability, cultural alignment, and pedagogical risk in AI-supported Indonesian classrooms by analysing how chatbot responses handle Bahasa Indonesia, English, code-mixed input, informal registers, regional-language elements, and culturally situated classroom prompts. Method: This study uses a qualitative-dominant mixed-method design based on 24 document-derived prompt-response episodes collected from verified public educational, curricular, linguistic, platform, and AI-governance sources, analysed through multilingual reliability audit, cultural-alignment analysis, and pedagogical-risk mapping. Results: Findings show that chatbot reliability is strongest in formal Bahasa Indonesia and English-oriented classroom tasks, but becomes less stable when prompts involve code-mixing, informal language, politeness nuance, or regional-language meaning. Cultural alignment appears mostly partial because responses often recognise general classroom intent while simplifying local examples, religious-social pragmatics, and translanguaging practices. Implication: Pedagogical risk is concentrated in assessment overreach, local meaning erasure, and standard-language dominance, indicating that teacher mediation remains necessary before chatbot output enters classroom use. Novelty: This study contributes a sociolinguistic accountability framework for evaluating chatbot use in multilingual education.