WHEN THE CHATBOT SPEAKS MAMBWE:

Abstract

Mother tongue education in rural Zambia faces a paradoxical challenge: AI tools promise scalable literacy support, but their training data exclude Mambwe (spoken in Northern Province, districts of Mbala and Mpulungu) and Lungu (around Lake Tanganyika). This paper reports on a participatory design study with 45 Grade 5 teachers and 320 learners across three zone schools, where a prototype chatbot was deployed for Mambwe reading comprehension exercises. The chatbot, fine-tuned on a small corpus of Mambwe folktales and church primers, produced two unanticipated outcomes. First, learners exhibited code-switching fatigue: the chatbot's inability to handle natural Mambwe-English alternations forced learners into artificial monolingual production, reducing spontaneous language use by 58%. Second, lexical gaps emerged in domains such as digital navigation (kupenda – to click; absent) and emotional expression (chilowola – confusion after a riddle; present but misused). The paper argues that current mother-tongue AI tools replicate colonial language ideologies by privileging standardised, written varieties over oral, heteroglossic practices. In response, the study proposes a minimally supervised lexicon expansion protocol using community radio recordings. In foreign-language English education, LLMs inadvertently penalise Lungu learners whose English shows substrate transfer (e.g., omitted articles), flagging them as errors rather than as transitional competences. The paper offers a bilingual ethical framework for AI deployment in Zambian mother-tongue classrooms.