AI in Global South Universities: Powerful Predictions, Weak Evidence of Real Learning Gains

Artificial intelligence is sweeping through universities in emerging economies, promising to predict which students will drop out, personalise coursework, and streamline institutional decisions. But a sweeping new review of the evidence suggests that much of this technological enthusiasm is running ahead of proof. A critical scoping review published in Discover Education by Abdullah F. Alenezi of the Public Authority for Applied Education and Training in Kuwait examined how artificial intelligence and learning analytics are actually being implemented across higher education systems in the Global South, and the picture it paints is one of striking regional contrasts, impressive algorithms, and a persistent shortage of evidence that any of it improves learning.

The review followed the methodological framework of Arksey and O’Malley and was reported according to the PRISMA Extension for Scoping Reviews. The author searched eight major databases, including PubMed, ERIC, ACM Digital Library, ScienceDirect, Scopus, and Wiley Online Library, for peer-reviewed and selected grey literature published between January 2019 and October 2025. Out of 1,847 records identified, plus 12 from grey literature sources, 1,436 titles and abstracts were screened after duplicate removal, and 63 full-text articles were assessed. Just 19 primary studies met the eligibility criteria, supplemented by one earlier Brazilian study retained as historical background. The evidence base spanned empirical investigations, systematic reviews, policy reports, bibliometric analyses, and conceptual studies from Latin America, Sub-Saharan Africa, South Asia, East Asia, and cross-regional Global South research.

What emerged was not a single trajectory of AI adoption but a geographically differentiated maturity landscape. Latin America, represented by five studies, is dominated by collaborative capacity-building, most notably the Learning Analytics in Latin America Project, known as LALA, which spans Chile, Brazil, and Colombia. LALA built shared data standards, multi-country training programmes, and institutional dashboards, and reported gains in student retention and faculty data literacy. Yet the review documents a sobering pattern of decay: when external European Union funding concluded, the initiative weakened because sustainability frameworks had never been institutionalised. Analytics governance remained a project-level activity rather than a permanent university function, a phenomenon the review calls the pilot-to-policy gap.

Sub-Saharan Africa, by contrast, has prioritised ethics over algorithmic sophistication. Scholars such as Paul Prinsloo and Rogers Kaliisa have shifted the continental discourse toward epistemic justice, framing data privacy, student consent, and fairness as prerequisites for responsible analytics adoption. Early-warning systems piloted at South African universities have offered proof of concept but remain constrained by infrastructural inequities, low interoperability between learning management systems, and limited staff data fluency. The review argues that African scholarship contributes a normative model with global resonance: analytics as a socio-technical enterprise demanding cultural legitimacy and contextual adaptation rather than mere technical deployment.

South Asia presents perhaps the most paradoxical case. Studies from India and Pakistan report predictive accuracies exceeding 90 percent for grade-point average and dropout forecasting using machine learning models such as Random Forest, XGBoost, and Support Vector Machines. Yet the research overwhelmingly stops at model validation, rarely extending to intervention design or evaluation of learning outcomes. The review coins the term algorithmic myopia to describe this pattern: a research culture that privileges precision metrics, such as precision, recall, and area under the curve, over meaningful educational impact. Without human-in-the-loop interpretation, fairness auditing, or longitudinal evaluation, technically impressive models may identify at-risk students without improving their educational experiences. In one Brazilian dataset, early-warning alerts linked to personalised advising were associated with a 15-percentage-point reduction in dropout relative to the previous cohort, and a Pakistani institution reported roughly an 8-point improvement in first-year retention with human-mediated follow-up, but the review stresses these are single-institution, quasi-experimental estimates without matched control groups, indicators of potential rather than causal proof.

East Asia and the Middle East remain at exploratory stages, with only two eligible studies. China represents a high-capacity, state-coordinated adoption landscape with rapid bibliometric growth in generative AI, while Gulf Cooperation Council countries show early experimentation with adaptive and generative AI tutoring. Both contexts exhibit limited pedagogical evaluation and regulatory ambiguity. Across all regions, five intervention types were identified: predictive early-warning systems, adaptive and personalised learning platforms, generative AI academic support, institutional analytics dashboards, and automated assessment systems. Generative AI tools such as ChatGPT show the fastest growth, but the evidence rests almost entirely on self-reported survey data vulnerable to social desirability bias, with no objective performance verification and no longitudinal or experimental studies of learning impact, equity implications, or academic integrity risks.

Beneath these regional differences, the review identifies structural constraints that recur everywhere. Data infrastructure quality emerges as the primary determinant of feasibility, with chronic deficiencies in internet connectivity, system interoperability, and record completeness undermining algorithmic reliability and adherence to FAIR data principles remaining limited. Governance consistently lags behind adoption: many institutions import policy templates from high-income countries without adaptation, producing what the review terms regulatory dissonance, and few conduct Data Protection Impact Assessments or maintain ethical audit trails. Faculty attitudes range from cautious optimism to scepticism, with analytics frequently perceived as instruments of surveillance. Financial sustainability is similarly fragile, with nearly all regions operating in short-term, donor-funded cycles that produce temporary gains but little institutional learning.

The review is anchored in three theoretical traditions that sharpen its critique. Datafication theory, drawing on Ben Williamson’s work, describes how student behaviours such as attendance and engagement are encoded as predictive inputs, risking a reductionism that compresses the complexity of learning into metrics reflecting institutional priorities. Decolonial epistemology foregrounds data colonialism, the extraction of institutional and student data without reciprocal benefit to local communities, and challenges the uncritical import of AI systems calibrated for Western educational contexts into the Global South. Responsible AI ethics, informed by UNESCO’s human-centred framework, demands transparency, accountability, cultural sensitivity, and student agency, extending to anticipatory governance that identifies unintended consequences before deployment rather than reacting after harm occurs.

From this synthesis, Alenezi proposes a Human-Centred and Contextually Responsive AI Framework integrating responsible AI, equity by design, educator participation, and contextual governance. The practical recommendations are concrete: institutions should establish data governance boards with academic, student, and IT representation before scaling any system; complete impact assessments before deploying predictive analytics; publish model cards documenting training data and known limitations; and conduct annual algorithmic audits accessible to faculty and students. Educator capacity should be developed within pedagogical cycles rather than treated as a technical add-on, and deployment should begin with high-yield, low-risk use cases, with early-warning systems linked to personalised advising in first-year programmes cited as the most empirically supported entry point. Scalability should be assessed across technical, financial, institutional, and pedagogical dimensions before any expansion.

The review’s limitations are candidly acknowledged: only 19 primary studies, a single reviewer, heterogeneous designs that precluded meta-analysis, and language restrictions to English, Spanish, and Portuguese. Yet its central message is unambiguous. The future of AI in emerging higher education will not be determined by algorithms alone but by how effectively institutions combine technological innovation with human expertise, ethical governance, and contextual understanding. As the global community pursues Sustainable Development Goal 4’s promise of inclusive, equitable quality education, the review warns that without equity-by-design principles, fairness evaluation across gender, socioeconomic status, and digital access, and sustained institutional commitment beyond donor cycles, AI risks reinforcing the very inequalities it promises to dissolve. The next generation of research, it argues, must stop asking whether AI can predict educational outcomes and start asking whether it actually improves learning.

Subject of Research: Implementation of human-centred artificial intelligence and learning analytics in higher education institutions in emerging economies

Article Title: Human-centred Artificial Intelligence and Learning Analytics in emerging higher education: a critical scoping review

Article References: Alenezi, A. F. (2026). Human-centred Artificial Intelligence and Learning Analytics in emerging higher education: a critical scoping review. Discover Education, 5(1), Article 1143. https://doi.org/10.1007/s44217-026-02066-9

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02066-9

Keywords: artificial intelligence, learning analytics, higher education, emerging economies, Global South, predictive analytics, educational equity, responsible AI, data governance, algorithmic myopia, generative AI, SDG4