Artificial Intelligence, Large Language Models, and Machine Learning in African Health Professions Education: A Systematic Review of Adoption, Literacy, Institutional Feasibility, and Ethical Perceptions

Large language models (LLMs) are being adopted across African health professions education (HPE) faster than institutions, infrastructure, or evaluation practice can track them, producing a natural large-scale case study in uncalibrated human AI interaction under real resource constraint. We synthesize evidence from 19 studies (PRISMA 2020; PROSPERO CRD420261356111; n > 8,000 participants) spanning eight countries and two multinational cohorts to characterize three deployment-relevant phenomena rarely visible in high-income LLM evaluation settings: (i) a reproducible knowledge–perception paradox in which self-assessed AI competence exceeds objectively tested competence; (ii) an infrastructure gate that decouples stated adoption intention from sustained use, independent of model quality; and (iii) a positive relationship between objective AI literacy and ethical apprehension, contrary to the assumption that literacy monotonically increases comfort. We formalize these as a three-layer framework extending UTAUT with an explicit resource access gate, and translate the findings into design targets for low-bandwidth interfaces, in-product calibration tooling, and trust-aware safety design– an evidence base for AI system builders that existing African-language and low-resource-NLP work does not by itself provide.