ICIBEMBA SYNTAX AGAINST MACHINE PREDICTION:

Abstract

This paper investigates the friction between Large Language Models (LLMs) and the morphosyntactic structures of Icibemba (M42), a dominant Bantu language of northern Zambia. While LLMs excel in Germanic and Romance languages, their transformer-based attention mechanisms struggle with Bantu noun class systems and agglutinative verbal morphology. Using a dataset of 1,200 Icibemba sentences annotated for locative applicatives and progressive aspectual markers, we demonstrate that GPT-4–derived models systematically overgeneralise subject-verb-object patterns, producing grammatically impossible predictions (e.g., misplacing the locative prefix ku- as a prepositional object). The study further contrasts LLM output with classroom performance of L1 Bemba-speaking primary learners in Mpulungu District, revealing that models fail to replicate child-like acquisition of chi-/fi- class agreement. Pedagogically, the paper argues against uncritical adoption of LLM-based tutoring for Bantu-language literacy and proposes a hybrid model: human-led construction grammar instruction augmented by token-level visualisations of model uncertainty. The paper concludes that without Bantu-specific training corpora and attention-head modifications, LLMs risk reinforcing standard language ideologies that erase rural Bemba, Mambwe, and Lungu varieties. This has direct implications for Zambia's language-in-education policy, which recognises seven local languages but provides no digital infrastructure for their computational linguistics.