Reframing artificial intelligence and learning analytics in emerging higher education: towards a human-centred and contextually responsive future

Abstract

Introduction:

Artificial intelligence (AI) and learning analytics (LA) are increasingly used to enhance student outcomes and support institutional decision-making in higher education. However, evidence from emerging economies remains fragmented and unevenly distributed across regions. This critical scoping review synthesises current evidence on AI and LA implementation across higher education institutions in Latin America, Sub-Saharan Africa, South Asia, and the Middle East.

Methods:

The review followed the Joanna Briggs Institute Population-Concept-Context (PCC) framework and was reported in accordance with the PRISMA-ScR guidelines. Systematic searches of eight academic databases and selected grey literature sources identified 31 eligible publications published between 2019 and 2026. The evidence base comprised 20 empirical studies, 6 conceptual papers, and 5 policy or grey literature sources. To enhance transparency, studies were classified using a structured evidence-grading framework ranging from Grade A (high evidence) to Grade D (conceptual and policy evidence).

Results:

Four dominant application areas emerged: predictive early-warning systems, adaptive learning platforms, generative AI tools, and institutional analytics dashboards. Considerable regional variation was identified. Latin America demonstrated comparatively greater implementation maturity, while Sub-Saharan Africa placed greater emphasis on governance, ethics, and data justice. South Asia showed strong technical sophistication but limited pedagogical integration, whereas evidence from the Middle East reflected predominantly early-stage implementation. Across all regions, evaluation focused primarily on technical performance, with comparatively limited attention given to educational impact, equity outcomes, and long-term sustainability. Through a three-stage analytical synthesis, the review develops the Human-Centred AI and Learning Analytics Implementation Framework (HCAI-LIF), comprising four interdependent dimensions: Governance, Pedagogy, Equity, and Sustainability.

Discussion:

The review reframes AI and LA as socio-technical systems that require context-sensitive adaptation and responsible implementation rather than technology-driven adoption alone. The proposed framework provides an evidence-informed foundation for policymakers, institutional leaders, and researchers seeking to promote equitable, sustainable, and human-centred digital transformation in higher education across emerging economies.

1 Introduction

Artificial intelligence (AI) and learning analytics (LA) have become increasingly central to contemporary higher education, particularly in efforts to improve student retention, personalise learning experiences, enhance academic support, and strengthen institutional decision-making (Hilliger et al., 2024; Holmes et al., 2019; Williamson et al., 2020). Advances in predictive modelling, early-warning systems, adaptive learning environments, and more recently generative AI have accelerated the adoption of data-driven approaches across HEIs worldwide (Dempere et al., 2023; Ouyang et al., 2022; UNESCO, 2023). These technologies are frequently presented as scalable solutions capable of improving educational quality, operational efficiency, and student success. However, growing evidence suggests that their implementation raises significant pedagogical, ethical, organisational, and governance challenges that extend well beyond technical performance (Crawford, 2021; Khalil et al., 2023; Prinsloo, 2024).

These challenges are particularly pronounced within emerging and developing economy contexts, where institutional capacity, digital infrastructure, data governance mechanisms, regulatory frameworks, and educational resources often remain uneven or underdeveloped (Prinsloo and Kaliisa, 2022b; UNESCO, 2023). In such settings, AI and LA may function either as enabling innovations that enhance educational access and effectiveness, or as mechanisms that inadvertently reinforce existing inequalities and structural disadvantages (Guzmán-Valenzuela et al., 2021; Salas-Pilco et al., 2020). Concerns surrounding algorithmic bias, inadequate transparency, fragmented data ecosystems, weak interoperability, insufficient educator preparedness, and limited institutional oversight have become increasingly prominent in the literature (Khalil et al., 2023; Villasenor, 2019). Consequently, there is growing concern that technological adoption may be progressing faster than institutions' capacity to ensure responsible, equitable, and pedagogically meaningful implementation (Holmes et al., 2023; Prinsloo, 2024).

Despite the rapid growth of AI and LA research, much of the existing literature originates from well-resourced institutions in developed economies and frequently prioritises technical validation metrics over broader educational outcomes (Hilliger et al., 2024; Salas-Pilco et al., 2020). Studies commonly report measures such as accuracy, precision, recall, and model performance while paying comparatively limited attention to educational impact, equity outcomes, stakeholder experiences, implementation sustainability, and long-term institutional integration (Ahmed, 2024; Khalil et al., 2023). Evidence from emerging economies remains fragmented, geographically uneven, and characterised by heterogeneous methodologies, making it difficult to draw coherent conclusions regarding effective implementation practices across diverse contexts.

The diversity of socio-economic, cultural, political, and institutional conditions across emerging higher education systems further complicates attempts to generalise findings from technologically advanced settings. Factors such as resource constraints, governance maturity, digital literacy, policy environments, and local educational priorities substantially influence the effectiveness and sustainability of AI and LA initiatives (Prinsloo and Kaliisa, 2022b; Simbeck, 2024; UNESCO, 2023). Addressing these gaps requires moving beyond purely technical perspectives towards a broader socio-technical understanding that recognises the interdependence of technology, pedagogy, governance, institutional culture, and human agency.

To address these limitations, this study undertakes a critical scoping review of AI and LA implementation in emerging higher education systems. The review is informed by a Population–Concept–Context (PCC) framework consistent with Joanna Briggs Institute (JBI) guidance for scoping reviews. The Population comprises students, educators, institutional leaders, and support staff within HEIs. The Concept focuses on the implementation, evaluation, governance, and educational implications of AI and learning analytics systems, including predictive analytics, early-warning systems, adaptive learning platforms, generative AI applications, and institutional analytics dashboards. The Context encompasses HEIs operating within emerging and developing economies, including Latin America, Sub-Saharan Africa, South Asia, the Middle East and North Africa (MENA), and the broader Global South.

2 Materials and methods

2.1 Review design and analytical framework

This study adopted a critical scoping review design to map, synthesise, and critically evaluate evidence concerning the implementation of AI and LA in HEIs across emerging and developing economies. The review followed the methodological framework proposed by Arksey and O’Malley (2005) and refined by Levac et al. (2010), with reporting guided by the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) checklist (Tricco et al., 2018). Unlike conventional systematic reviews, which seek to aggregate homogeneous evidence and estimate intervention effects through meta-analysis, critical scoping reviews facilitate the exploration of heterogeneous evidence and support interpretive synthesis across emerging, multidisciplinary fields—particularly appropriate given the rapidly evolving nature of AI and LA research and the importance of examining pedagogical, ethical, governance, and contextual dimensions alongside technical performance.

2.2 Search strategy and information sources

The search covered the period from January 2019 to April 2026. This timeframe was selected to capture the rapid expansion of AI-enabled educational technologies following accelerated digital transformation associated with the COVID-19 pandemic, increased adoption of platform-based learning environments, and the emergence of generative AI applications in higher education.

Eight major academic databases were searched:

The core Boolean search string applied across all databases was:

(“artificial intelligence” OR “machine learning” OR “learning analytics” OR “predictive analytics” OR “adaptive learning” OR “generative AI” OR “algorithmic decision-making”) AND (“higher education” OR “university” OR “tertiary education” OR “HEI”) AND (“emerging econom*” OR “developing countr*” OR “Global South” OR “Latin America” OR “Sub-Saharan Africa” OR “South Asia” OR “Middle East”)

Database-specific adaptations were applied where required, including field-tag modifications and thesaurus-controlled vocabulary. For example, ERIC descriptors were used in the ERIC database; MeSH (Medical Subject Headings) terms were applied in PubMed; and the ACM Computing Classification System was used in the ACM Digital Library. Although search syntax was adjusted to reflect the indexing requirements of individual databases, the underlying search concepts, keywords, and Boolean structure remained consistent throughout the review process.

To strengthen policy relevance and reduce publication bias, a structured grey literature search was also undertaken. Sources included UNESCO, OECD, World Bank, and Deloitte Insights repositories. Grey literature was eligible if it: (a) was published between 2019 and 2026; (b) focused explicitly on AI, learning analytics, or digital transformation in higher education; (c) included substantial discussion of implementation, governance, equity, or institutional readiness; and (d) originated from recognised international organisations, governmental agencies, or institutional research bodies. Marketing materials, opinion pieces, and documents lacking methodological transparency were excluded. While grey literature broadened contextual understanding, it was not used as primary evidence of intervention effectiveness (see Section 2.3 on evidence classification).

2.3 Eligibility criteria and evidence classification system

Studies were eligible for inclusion if they: (1) examined AI or LA systems within higher education institutions; (2) focused on emerging or developing economy contexts consistent with the PCC framework; (3) reported educational, institutional, implementation, or governance outcomes; (4) were published between January 2019 and April 2026; and (5) were available in English, Spanish, or Portuguese. Studies were excluded if they: (a) focused exclusively on K–12, vocational, or corporate training settings; (b) examined AI applications unrelated to learning, teaching, or educational decision-making; (c) provided insufficient methodological or analytical information; or (d) were duplicate publications.

2.4 Study selection and screening process

Figure 1

2.5 Data extraction and evidence charting

A standardised extraction template was developed and piloted on five studies prior to full extraction. Extracted information included: author(s) and publication year; country and geographic region; study type and evidence classification (Type 1/2/3); AI or LA intervention type; methodological design; sample characteristics; educational and institutional outcomes; and governance, equity, and sustainability considerations. All extracted data were compiled in a comprehensive evidence table (Table 1) organised by evidence type and geographic region.

2.6 Evidence appraisal and grading framework

While formal risk-of-bias assessment falls outside the standard requirements of scoping review methodology, a structured evidence-grading framework was implemented to facilitate transparent differentiation in evidence quality across the empirical literature. This framework supports interpretive weighting rather than formal quality assessment, and all claims throughout the manuscript are explicitly linked to evidence grades to ensure proportional interpretation.

2.7 Framework development: three-stage analytical synthesis

Figure 2

2.8 Researcher reflexivity and analytical transparency

As a critical scoping review, interpretation inevitably involved researcher judgement during thematic synthesis and framework development. To enhance transparency, all analytical decisions were documented and linked explicitly to evidence grades and source classifications. The framework development process followed established JBI scoping review procedures with systematic documentation of each analytical stage. All included sources were independently coded by evidence type and grade using predefined criteria to minimise subjective interpretation.

2.9 Methodological limitations

Several methodological limitations should be acknowledged at the outset. First, the exclusion of Scopus and Web of Science may have reduced retrieval sensitivity, though discipline-specific databases were prioritised for depth. Second, language restrictions (English, Spanish, Portuguese) may have excluded relevant studies published in other languages. Third, substantial heterogeneity in study designs, outcome measures, and institutional contexts precluded meta-analysis and limited causal inference. Fourth, grey literature selection may favour international organisational perspectives over local or regional viewpoints. Finally, the evidence base remains geographically uneven, with Latin America and South Asia more strongly represented than the Middle East or parts of Sub-Saharan Africa. These limitations are further discussed in Section 8.1.

3 Results

3.1 Study selection and sample characteristics

The systematic search process identified 2,847 records across academic databases and grey literature repositories. Following duplicate removal (n = 924), 1,923 records underwent title and abstract screening. A total of 142 full-text s were assessed for eligibility, resulting in a final sample of 31 sources published between 2019 and April 2026. The final evidence base comprised 20 Type 1 empirical studies, 6 Type 2 conceptual and theoretical contributions, and 5 Type 3 policy and grey literature documents. Among the 20 empirical studies, four were classified as Grade A (high evidence), ten as Grade B (moderate evidence), and six as Grade C (exploratory evidence). The evidence base revealed considerable variation in methodological rigour, institutional maturity, implementation focus, and evaluation quality (Figure 3).

Figure 3

Table 1 summarises the 31 included sources, classified by evidence type and grade, with details on geographic region, intervention category, methodological design, and key findings. Table 1 enables readers to trace specific regional patterns and evidence grades back to individual studies.

3.2 Types of AI and learning analytics interventions

Four main categories of AI and LA intervention were identified across the included studies (Figure 2 and Table 3).

Predictive early-warning systems constitute the most prevalent intervention type, present in approximately 45% of empirical studies (n = 9 Type 1 studies; Grade A-B evidence). These systems utilise machine learning to identify at-risk students on the basis of academic and behavioural data. Grade A and B evidence indicated that, while predictive accuracy was frequently high—exceeding 85%–90% in several South Asian studies (Ahmed, 2024; Aslam et al., 2024)—only a limited number of studies evaluated how predictions translated into improved student outcomes or retention. The dominant evaluation focus remained on model validation metrics such as precision, recall, and AUC, with limited attention to downstream pedagogical impact or intervention effectiveness.

Adaptive and personalised learning systems, present in approximately 25% of empirical studies (n = 5 Type 1 studies; Grade B evidence), tailor content delivery, pacing, and assessment difficulty based on learner interaction data. Grade B evidence from Latin America and South Asia suggested moderate improvements in engagement and learning satisfaction, particularly in large-scale STEM courses. However, scalability and contextual adaptation remained persistent challenges, with many systems imported from high-income contexts without meaningful localisation for emerging economy settings.

Generative AI academic support tools, including large language model-based assistants, were increasingly documented in the literature, particularly in Grade C evidence. Studies relying predominantly on self-reported perceptions (Dempere et al., 2023; Ouyang et al., 2022) raised concerns about response bias and the absence of empirically validated learning outcomes. Issues of equitable access, academic integrity, and epistemic trust received limited systematic attention in the included studies.

Institutional analytics for decision-making, deployed at governance level, support enrolment forecasting, resource allocation, and programme performance evaluation. Grade B evidence indicated measurable efficiency gains in select case studies (Deloitte, 2025), though implementation frequently occurred in departmental silos with limited integration into institutional strategic frameworks.

3.3 Regional patterns in AI and learning analytics adoption

3.3.1 Latin America: institutional capacity building and collaborative implementation

Within the reviewed evidence base, Latin America (n = 6 empirical studies, Grade A–B evidence; supported by Type 2 and Type 3 evidence) appeared to demonstrate comparatively greater implementation maturity than other regions (Figure 3), although conclusions should be interpreted cautiously given the limited number of empirical studies. Grade A and Grade B evidence documented collaborative learning analytics programmes, institutional dashboard development, and analytics literacy initiatives. Pioneering initiatives such as the Learning Analytics in Latin America (LALA) project demonstrated improvements in retention monitoring, student-support coordination, and data-informed decision-making (Hilliger et al., 2024; Salas-Pilco et al., 2020). However, sustainability challenges were evident across multiple cases, with many initiatives dependent on short-term external funding and international partnerships. While implementation maturity appeared comparatively advanced, institutionalisation remained incomplete in many settings.

3.3.2 Sub-Saharan Africa: governance, ethics, and data justice

Sub-Saharan African studies (n = 4 empirical studies, Grade B evidence; supported by Type 2 contributions) exhibited a distinct emphasis on governance, ethics, and contextual legitimacy, foregrounding concerns regarding privacy, fairness, algorithmic bias, and data justice. Grade B evidence consistently highlighted that implementation outcomes were influenced as strongly by governance capacity and stakeholder trust as by technological sophistication (Prinsloo, 2024; Prinsloo and Kaliisa, 2022b). Common barriers included fragmented infrastructure, limited technical expertise, and financial constraints. Consequently, African scholarship framed AI and LA primarily as socio-technical and governance challenges rather than purely technological innovations. This regional focus reflects both the genuine importance of governance in emerging contexts and the research priorities of scholars in the region.

3.3.3 South Asia: technical sophistication and limited pedagogical integration

South Asian studies (n = 3 empirical studies; mixed evidence grades) demonstrated the highest levels of algorithmic sophistication. Grade A evidence reported machine-learning models achieving predictive accuracies exceeding 85%–90% for academic performance and dropout forecasting (Ahmed, 2024; Aslam et al., 2024). Despite strong technical validation, evidence regarding educational impact remained limited. Most Grade A studies terminated at model development or validation stages without evaluating intervention effectiveness, behavioural change, or learning improvement. Grade C evidence (survey-based studies) provided exploratory data on student and educator perceptions. As a result, technical capability consistently exceeded pedagogical integration. This pattern represents a significant finding from a limited sample (n = 3) and warrants replication and extension in future research.

3.3.4 Middle East and emerging contexts: increasing experimentation

Evidence from the Middle East (n = 2 empirical studies; Grade B–C evidence) reflected early-stage institutional experimentation. Grade B and Grade C evidence indicated increasing institutional interest in adaptive learning systems, institutional analytics, and emerging generative AI applications. However, given the very small sample (n = 2), most studies lacked longitudinal evaluation, comprehensive governance assessment, or rigorous outcome measurement, meaning conclusions regarding effectiveness remain preliminary. This region requires substantially more empirical investigation before robust patterns can be identified.

3.3.5 Cross-regional and global south perspectives

Cross-regional studies (n = 5 empirical; supplemented extensively by Type 2 and Type 3 sources) contributed significantly to the conceptual and policy dimensions of the review. These studies consistently highlighted the importance of fairness, transparency, accountability, and contextual responsiveness. Grade D evidence underscored the risks of data colonialism—the importation of AI models and governance frameworks developed in high-income contexts without meaningful contextual adaptation. Successful implementation was repeatedly linked to stakeholder participation, governance maturity, cultural alignment, and institutional readiness.

3.4 Cross-cutting findings: educational impact, equity, and sustainability

A key finding across the empirical literature was the disparity between technical validation and educational impact assessment. Grade A evidence demonstrated that AI and LA systems can support improved retention monitoring and institutional decision-making when combined with proactive human intervention. However, only approximately 20% of studies measured educational outcomes directly, with most relying on technical metrics—predictive accuracy, model sensitivity, and performance indices—as proxies for effectiveness. The review consequently finds moderate Grade B evidence supporting institutional efficiency gains but considerably weaker evidence regarding educational transformation or learning improvement.

Only a small proportion of empirical studies explicitly examined equity outcomes. Grade B and Grade C studies that addressed fairness reported concerns regarding algorithmic bias, unequal predictive performance across student groups, and the potential reproduction of existing institutional inequalities. Cross-regional conceptual studies (Type 2) and policy documents (Type 3) consistently argued that equity cannot be assumed as an automatic outcome of technological innovation and requires intentional design, governance oversight, and continuous monitoring.

Sustainability emerged as one of the most consistently reported implementation challenges. Eleven of the twenty empirical studies described pilot projects or short-term initiatives, while only four documented evidence of institutionalisation beyond initial implementation phases. Common barriers included dependence on external funding, insufficient technical capacity, limited educator engagement, weak governance structures, and inadequate organisational support. These findings suggest that long-term success depends less on technological sophistication and more on institutional commitment, governance maturity, and capacity-building investments.

4 Human-centred AI and learning analytics implementation framework (HCAI-LIF)

4.1 Analytical basis for framework development

The Human-Centred AI and Learning Analytics Implementation Framework (HCAI-LIF) was developed through the three-stage analytical synthesis process described in Section 2.7. Across the 20 empirical studies, six conceptual contributions, and five policy documents, a consistent pattern emerged: implementation success was influenced less by technological sophistication alone and more by the interaction among organisational, pedagogical, ethical, and institutional factors. Four dimensions appeared consistently across evidence types, geographic regions, and intervention categories—Governance, Pedagogy, Equity, and Sustainability—and these therefore constitute the analytical foundation of the proposed framework (Figure 4 and Table 4).

Figure 4

The inductive first stage identified recurring implementation enablers (e.g., educator capacity, governance frameworks, data quality) and barriers (e.g., infrastructure deficits, funding dependency, algorithmic bias) across the empirical evidence (Table 3). The deductive second stage aligned these themes with responsible AI principles articulated in UNESCO (2023), socio-technical systems theory, and international governance frameworks including the African ethical analytics frameworks advanced by Prinsloo (2024). The convergent third stage identified four dimensions appearing robustly across empirical, conceptual, and policy evidence, forming the basis of the HCAI-LIF.

4.2 Dimension one: governance

Governance emerged as the most consistently reported implementation requirement across all 31 sources. Grade B evidence from Sub-Saharan Africa (n = 4 studies), Latin America (n = 6 studies), and cross-regional policy literature highlighted the importance of accountability, transparency, privacy protection, ethical oversight, and stakeholder participation. A recurring finding was that technological implementation frequently advanced more rapidly than governance capacity, creating accountability gaps. Consequently, institutions deployed predictive analytics, adaptive systems, or generative AI applications without corresponding mechanisms for ethical oversight, algorithmic auditing, or accountability to affected learners.

The Governance dimension of the HCAI-LIF therefore encompasses: data governance and stewardship; ethical oversight and accountability mechanisms; privacy and consent management; transparency and explainability requirements; and stakeholder participation and trust-building processes. Operationally, institutions should develop governance frameworks codifying consent procedures, data minimisation principles, access controls, and algorithmic risk management prior to system deployment. Transparent documentation through model cards, algorithmic audits, and data protection impact assessments can operationalise UNESCO's (2023) principles of human-centred AI.

4.3 Dimension two: pedagogy

A recurring finding across all regional evidence clusters was the disconnect between technical capability and educational impact. Grade A and Grade B studies demonstrating high predictive accuracy (South Asia, Latin America) rarely evaluated corresponding improvements in learning outcomes, behavioural change, student engagement, or educational transformation. The framework consequently positions pedagogy as a central implementation dimension, emphasising that AI and learning analytics generate educational value only when technological outputs are translated into meaningful interventions through educator interpretation and learner support.

The Pedagogy dimension encompasses: human-in-the-loop decision-making that preserves educator agency; professional development and analytics literacy for teaching staff; curriculum alignment ensuring that data insights connect to instructional design; learner-centred design processes; and responsive, timely intervention systems. Bridging the gap between algorithmic output and instructional action requires co-design workflows in which educators collaborate with data scientists to ensure contextual alignment and foster institutional ownership.

4.4 Dimension three: equity

Equity concerns appeared consistently across conceptual (Type 2) and policy (Type 3) evidence and were increasingly evident in empirical studies (Type 1) across all regions. Issues of algorithmic bias, fairness, accessibility, representation, and contextual adaptation emerged as critical implementation considerations. The framework positions equity as a design principle rather than a post hoc evaluation criterion. Responsible implementation requires disaggregated analysis of differential impacts across student subgroups defined by gender, socio-economic status, disability, first-generation student status, and digital access.

The Equity dimension encompasses: fairness and bias mitigation at the design stage; accessibility and inclusion considerations; contextual responsiveness to local educational realities; student agency and participation in system design; and continuous monitoring of differential outcomes across population subgroups. Grade B evidence indicated that systems deployed without fairness auditing risk reproducing or amplifying existing institutional inequalities under a veneer of algorithmic objectivity.

4.5 Dimension four: sustainability

Sustainability represented one of the most frequently reported implementation barriers across empirical evidence. Most initiatives remained confined to pilot projects, and only four of the twenty empirical studies documented evidence of institutionalisation beyond initial implementation phases. Common barriers included funding dependency, insufficient technical capacity, limited educator engagement, weak governance structures, and inadequate organisational support. Grade B evidence repeatedly demonstrated that short-term technological success did not translate into long-term institutional transformation.

The Sustainability dimension encompasses: diversified and long-term funding models reducing dependency on external grants; technical and organisational capacity development; adequate and interoperable infrastructure investment; professional development supporting educator readiness; and organisational learning and iterative adaptation capabilities. Addressing the pilot-to-policy gap requires institutional commitment at leadership level alongside governance mechanisms capable of embedding AI and LA systems within strategic institutional frameworks.

4.6 Integrated framework

The HCAI-LIF conceptualises AI and learning analytics implementation as a socio-technical process operating at the intersection of Governance, Pedagogy, Equity, and Sustainability (Figure 4). Rather than treating technology as the primary driver of change, the framework positions technological capability within a broader ecosystem of institutional, educational, ethical, and organisational conditions. The four dimensions are interdependent: governance creates the conditions of trust enabling pedagogical integration; pedagogical embedding generates equity-relevant evidence; equity considerations inform governance refinement; and sustainability ensures that all dimensions are institutionally embedded over time (Table 4).

Framework dimensions emerged through convergence across empirical (Type 1), conceptual (Type 2), and policy (Type 3) evidence sources and should be interpreted as mutually reinforcing rather than sequential implementation stages.

5 Discussion

5.1 The dominance of technical performance over educational impact

The findings reveal a persistent and structurally embedded emphasis on predictive accuracy rather than educational impact across the reviewed literature. While Grade A and Grade B studies frequently reported high-performing predictive models—with South Asian studies reporting accuracies exceeding 85%–90%—few demonstrated corresponding improvements in student learning, engagement, or well-being. This constitutes a fundamental gap between technological capability and pedagogical application, characterisable as algorithmic myopia (Khalil et al., 2023). High AUC values and precision-recall metrics, while technically meaningful, provide limited insight into whether analytics systems are meaningfully improving educational practice.

This technocentric bias reflects a broader form of technological determinism in which computational capability is conflated with educational improvement. As Khalil et al. (2023) emphasise, predictive systems detached from pedagogy risk becoming instruments of stratification rather than empowerment, over-identifying marginalised students without providing equitable or human-centred support mechanisms. These patterns are supported by Grade B evidence from Latin America and Sub-Saharan Africa, which consistently indicated that implementation success depended on human interpretation and intervention rather than algorithmic accuracy alone. Advancing the field requires a fundamental reorientation toward impact-centred evaluation frameworks that integrate computational outputs with pedagogical reflection, stakeholder feedback, and equity monitoring.

5.2 Contextual and institutional misalignment

A second critical dimension concerns the transferability and contextual validity of AI and LA systems developed predominantly within Global North institutional contexts. Grade B evidence from multiple regions—though based on a limited number of studies—indicated that many systems were calibrated using datasets, platform architectures, and behavioural assumptions that do not reflect the lived realities of students in emerging HEIs. When deployed without meaningful localisation, such models encountered infrastructural, cultural, and pedagogical mismatches that compromised both reliability and relevance (Hilliger et al., 2024; Prinsloo and Kaliisa, 2022b).

This pattern reflects epistemic dependency—a reliance on external paradigms that limits regional agency in AI design and perpetuates transplanted innovation, in which imported solutions are only partially aligned with local socio-educational ecosystems (Prinsloo, 2024). Grade D evidence from cross-regional studies consistently argued that addressing this requires investment in localised feature engineering, participatory co-design processes involving educators and students, and the embedding of cultural, linguistic, and socio-economic variables into model architectures.

5.3 Governance as a foundation for responsible implementation

Across all regions and intervention categories, governance emerged from the synthesis as a critical determinant of implementation success. Grade B evidence from Sub-Saharan Africa most clearly demonstrated that institutions frequently adopted AI-enabled systems faster than governance mechanisms evolved to support responsible implementation. Without effective governance structures, technological innovation may generate unintended consequences including privacy concerns, algorithmic bias, reduced transparency, and declining stakeholder trust. These findings are particularly significant given that the majority of included studies reported no formal governance framework or ethical oversight mechanism at the time of implementation.

5.4 Equity beyond technological innovation

Although many studies were motivated by goals of improving student success and reducing dropout risk, Grade B and Grade C evidence indicated that relatively few explicitly evaluated equity outcomes or disaggregated results by student subgroup characteristics. Consequently, assumptions that AI and learning analytics automatically promote inclusion are not currently supported by strong empirical evidence from emerging economy contexts. Grade D conceptual evidence consistently argued that equity represents both an ethical requirement and a research priority, requiring intentional design, fairness assessment, differential outcome analysis, and contextual adaptation throughout the implementation lifecycle.

5.5 The pilot-to-policy gap

The review identified a persistent gap between pilot implementation and institutional transformation. Grade B evidence across Latin America, Sub-Saharan Africa, and South Asia indicated that many initiatives demonstrated promising short-term outcomes but failed to progress toward sustainable institutional integration. Common barriers identified across multiple evidence types included funding dependency, insufficient technical expertise, weak governance structures, limited organisational readiness, and inadequate educator engagement. These findings suggest that sustainable implementation depends less on technological sophistication than on institutional capacity and sustained organisational commitment—a conclusion supported consistently across evidence grades.

While these patterns are suggestive, the small number of empirical studies per region (2–6 studies) limits the robustness of regional comparisons. These patterns should be interpreted as emerging trends requiring validation in future research rather than definitive geographic characterisations.

5.6 Illustrative regional cases

Two regional exemplars illustrate how contextual forces mediate implementation success. The Learning Analytics in Latin America (LALA) Project (Hilliger et al., 2024) exemplifies a bottom-up regional collaboration that sought to mainstream LA literacy across universities in Chile, Brazil, and Colombia. By developing shared data standards and training programmes, LALA enhanced academic analytics maturity and improved student retention through evidence-based advising. Grade B evidence from this project demonstrated that capacity building and regional cooperation can partially overcome resource asymmetries. However, once external funding concluded, the initiative's sustainability weakened, revealing the fragility of externally financed implementation ecosystems and the absence of institutionalised analytics governance within partner universities.

In South Africa, Grade D and Grade B evidence from Prinsloo and Kaliisa (2022b) and Prinsloo (2024) shaped continental discourse by advancing ethical frameworks for learning analytics contextualised to African social realities, reframing analytics as tools for social justice and foregrounding data privacy, student consent, and institutional accountability. Pilot programmes implementing these principles influenced policy revisions at major universities. However, uneven infrastructure and fiscal constraints hindered systemic implementation, underscoring that ethical foresight must be accompanied by resource commitment for scalable transformation. Together, these cases confirm that implementation sustainability and equity are determined less by technological capability than by governance quality, cultural adaptation, and institutional vision.

5.7 Contribution to knowledge

This study contributes to the literature in four principal ways. First, it provides a cross-regional synthesis of AI and LA implementation evidence from emerging economies, addressing a significant gap in existing reviews that have predominantly focused on developed economy contexts. Second, by introducing a tripartite evidence classification system and structured evidence-grading framework (Grades A–D), it supports more proportionate interpretation of findings than previous narrative reviews, enabling readers to calibrate the strength of conclusions relative to the quality and type of supporting evidence. Third, it reframes AI and LA as socio-technical systems rather than purely technological solutions, highlighting the critical importance of governance, pedagogy, equity, and sustainability as interdependent implementation dimensions. Fourth, the HCAI-LIF presents an operationally grounded implementation framework developed through the systematic integration of empirical, conceptual, and policy evidence, offering a transparent and theoretically informed foundation for implementation across diverse higher education contexts.

6 Implications for policy and practice

7 Future research directions

Building upon these insights, a forward-looking research agenda for AI and LA in emerging HEIs should address five priority directions.

First, longitudinal studies and quasi-experimental evaluations capable of establishing causal relationships between AI and learning analytics interventions and educational outcomes are needed. Current evidence remains predominantly observational or correlational, limiting causal inference.

Second, context-sensitive fairness and privacy frameworks should be co-created with local stakeholders, reflecting cultural, linguistic, and legal norms rather than importing Western governance templates. This is particularly critical in Sub-Saharan Africa and South Asia, where existing governance frameworks remain nascent.

Third, empirical comparisons of teacher-in-the-loop design paradigms can clarify how human interpretation mediates algorithmic efficacy and ethical integrity, illuminating the pedagogical value of human oversight in automated systems.

Fourth, economic analyses of sustainability and total cost of ownership are required to identify financially viable implementation models for long-term analytics infrastructure in resource-constrained emerging HEIs.

Fifth, as generative AI tools proliferate, longitudinal research should assess their implications for learning outcomes, critical thinking, academic integrity, and epistemic autonomy, particularly within diverse and multilingual student populations in emerging economies.

8 Limitations and conclusions

8.1 Limitations

8.2 Conclusion

This study provides a critical synthesis of AI and learning analytics implementation in emerging higher education systems, examining evidence across Latin America, Sub-Saharan Africa, South Asia, the Middle East, and the broader Global South. Through a systematic PRISMA-ScR process guided by a PCC framework, 31 sources were identified and classified by evidence type (Type 1/2/3) and grade (A/B/C/D) to support proportionate and transparent interpretation of findings.

A central finding is the persistent imbalance between technological capability and educational application. The rapid advancement of predictive modelling and AI-enabled analytics has not been matched by equivalent progress in pedagogical integration, equitable design, rigorous impact evaluation, or long-term sustainability. Grade A and Grade B evidence consistently indicated that technical performance is relatively well documented, while educational impact, equity outcomes, and sustainability remain substantially under-evaluated across all regions.

The proposed HCAI-LIF offers a structured, evidence-grounded pathway for aligning AI and LA adoption with educational goals. By positioning Governance, Pedagogy, Equity, and Sustainability as interdependent and constitutive implementation dimensions derived from systematic synthesis of empirical, conceptual, and policy evidence, the framework provides actionable guidance for responsible, contextually appropriate, and educationally meaningful implementation. The framework seeks to address a significant limitation in the existing literature, namely the absence of structured implementation approaches that integrate human-centred principles, equity considerations, and institutional embedding alongside technical optimisation.

For AI and LA to realise their emancipatory potential in emerging higher education, the field must move beyond algorithmic optimisation toward models of governance-first, educator-engaged, equity-monitoring, and sustainably resourced implementation. The evidence reviewed suggests that the future success of AI and learning analytics in emerging higher education systems will depend less on algorithmic sophistication than on institutions' ability to align technological innovation with governance capacity, pedagogical purpose, equity objectives, and long-term sustainability. Responsible implementation therefore requires a shift from technology-centred adoption towards human-centred institutional transformation.

References

• 1

  AhmedE. (2024). Student performance prediction using machine learning algorithms. Appl. Comput. Intell. Soft Comput.2024, 4067721. 10.1155/2024/4067721

  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  ArkseyH.O'MalleyL. (2005). Scoping studies: towards a methodological framework. Int. J. Soc. Res. Methodol.8 (1), 19–32. 10.1080/1364557032000119616

  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  AslamM. A.MurtazaF.HaqM. E. U.YasinA.AzamM. A. (2024). A human-centered approach to academic performance prediction using personality factors in educational AI. Information15 (12), 777. 10.3390/info15120777

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  CeleN. (2021). Big data-driven early alert systems as means of enhancing university student retention and success. S. Afr. J. High. Educ.35 (2), 56–72.

  • Google Scholar
  • View reference in article

• 5

  Cengage Group. (2024). Digital Learning Pulse Survey: AI in Higher Education. Boston, MA: Cengage Group.

  • Google Scholar
  • View reference in article

• 6

  CrawfordK. (2021). Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. New Haven, CT: Yale University Press.

  • Google Scholar
  • View reference in article

• 7

  DaiK.LiuY.ZhangX. (2026). Generative AI in higher education: a bibliometric review of emerging trends, power dynamics, and global research landscapes. Comput. Educ. Artif. Intell.10, 100544. 10.1016/j.caeai.2026.100544

  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  de OliveiraC. F.SobralS. R.FerreiraM. J.MoreiraF. (2021). How does learning analytics contribute to prevent students’ dropout in higher education: a systematic literature review. Big Data Cogn. Comput.5 (4), 64. 10.3390/bdcc5040064

  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  Deloitte. (2025). 2025 Higher Education Trends: Digital Transformation and Institutional Analytics. New York, NY: Deloitte Insights.

  • Google Scholar
  • View reference in article

• 10

  DempereJ.AllamK.GatzaliL.SeifanS. (2023). Student perceptions of generative artificial intelligence tools in higher education. Front. Educ.8, 1206936. 10.3389/feduc.2023.1206936

  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  GournaS.RigouA.KyriaziF.MarinagiC. (2024). The added value of learning analytics in higher education. Int. J. Educ. Inf. Technol. 18,133–142. 10.46300/9109.2024.18.13

  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  Guzmán-ValenzuelaC.de la VegaI.SalinasJ. (2021). Mapping learning analytics research in Latin America: a bibliometric analysis. Heliyon7 (6), e06948. 10.1016/j.heliyon.2021.e06948

  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  Heredia-JiménezV.Jiménez MacíasA. A.Ortiz RojasM.Imaz MarínJ.Moreno-MarcosP. M.Muñoz MerinoP. J.et al (2020). “An early warning dropout model in higher education degree programs: a case study in Ecuador,” in CEUR Workshop Proceedings(Vol.2704) (Aachen: CEUR-WS.org),58–67.

  • Google Scholar
  • View reference in article

• 14

  HilligerI.OchoaX.MirandaC.Pérez-SanagustínM. (2024). Learning analytics capacity-building in Latin America: the LALA project experience. J. Learn. Anal.11 (1), 45–63. 10.18608/jla.2024.8409

  • CrossRef
  • Google Scholar

• 15

  HolmesW.BialikM.FadelC. (2019). Artificial Intelligence in Education: Promises and Implications for Teaching and Learning. Boston, MA: Center for Curriculum Redesign.

  • Google Scholar
  • View reference in article

• 16

  HolmesW.BialikM.FadelC. (2023). Artificial intelligence in education: promises and implications for teaching and learning. Int. J. Artif. Intell. Educ.33 (1), 13–31. 10.1007/s40593-021-00239-1

  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  Joanna Briggs Institute. (2020). JBI manual for Evidence Synthesis. Adelaide: Joanna Briggs Institute.

  • Google Scholar
  • View reference in article

• 18

  KhalilM.PrinslooP.SladeS. (2023). Fairness and transparency in learning analytics: a global review. J. Learn. Anal.10 (2), 67–84. 10.18608/jla.2023.7983

  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  LevacD.ColquhounH.O'BrienK. K. (2010). Scoping studies: advancing the methodology. Implement. Sci.5, 69. 10.1186/1748-5908-5-69

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  MangalM.PardosZ. A. (2024). Implementing equitable and intersectionality-aware ML in education: a practical guide. Br. J. Educ. Technol.55, 2003–2038. 10.1111/bjet.13484

  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  MarcolinoM. R.PortoT. R.PrimoT. T.TarginoR.RamosV.QueirogaE. M.et al (2025). Student dropout prediction through machine learning optimization: insights from moodle log data. Sci. Rep.15 (1), 9840. 10.1038/s41598-025-93918-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  MogoaleP. D.PretoriusA.MogaseR. C.SegooaM. A. (2025). Integrating artificial intelligence within South African higher learning institutions. S. Afr. J. Inf. Manag.27 (1), 1939. 10.4102/sajim.v27i1.1939

  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  OliveiraD.CunhaP. C.NakayamaC. J. C. A.KM. (2021). Learning analytics and dropout prediction in higher education: a systematic review. Big Data Cogn. Comput.5 (4), 64. 10.3390/bdcc5040064

  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  OuyangF.ZhengL.JiaoP. (2022). Artificial intelligence in online higher education: a systematic review of empirical research from 2011 to 2020. Educ. Inf. Technol.27 (6), 7893–7925. 10.1007/s10639-022-10925-9

  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  PrinslooP. (2024). Contextualising learning analytics ethics in African higher education. Learn. Media Technol.49 (3), 321–338. 10.1080/17439884.2024.2321437

  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  PrinslooP.KaliisaR. (2022a). Learning analytics on the African continent: an emerging research focus and practice. J. Learn. Anal.9 (2), 218–235. 10.18608/jla.2022.7539

  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  PrinslooP.KaliisaR. (2022b). Data privacy on the African continent: opportunities, challenges and implications for learning analytics. Br. J. Educ. Technol.53 (4), 894–913. 10.1111/bjet.13226

  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  PrinslooP.KaliisaR. (2025). Dimensions of privacy and its implications for learning analytics: preliminary insights from/for African higher education. Learn. Media Technol.50 (4), 432–447. 10.1080/17439884.2024.2321437

  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  Salas-PilcoS. Z.XiaoK.Rodríguez-TrianaM. J. (2020). Learning analytics in Latin America: a systematic literature review. Br. J. Educ. Technol.51 (4), 973–998. 10.1111/bjet.12952

  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  SalesA.BalbyL.CajueiroA. (2016). Exploiting academic records for predicting student drop out: a case study in Brazilian higher education. J. Inf. Data Manag.7 (2), 166–166.

  • Google Scholar
  • View reference in article

• 31

  SimbeckK. (2024). They shall be fair, transparent, and robust: auditing learning analytics systems. AI Ethics4, 555–571. 10.1007/s43681-023-00292-7

  • CrossRef
  • Google Scholar
  • View reference in article

• 32

  SmitM.Bond-BarnardT.WagnerR. F. (2025). Artificial intelligence in South African higher education: survey data of master’s level students. Data Brief61, 111813. 10.1016/j.dib.2025.111813

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 33

  TriccoA. C.LillieE.ZarinW.O'BrienK. K.ColquhounH.LevacD.et al (2018). PRISMA Extension for scoping reviews (PRISMA-ScR): checklist and explanation. Ann. Intern. Med.169 (7), 467–473. 10.7326/M18-0850

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 34

  UNESCO. (2023). Guidance for Generative AI in Education and Research. Paris: UNESCO Publishing.

  • Google Scholar
  • View reference in article

• 35

  VillasenorJ. (2019). Artificial Intelligence and Bias: Four Key Challenges. Washington, DC: Brookings Institution.

  • Google Scholar

• 36

  WilliamsonB.BayneS.ShayS. (2020). The datafication of teaching in higher education: critical issues and perspectives. Teach. High. Educ.25 (4), 351–365. 10.1080/13562517.2020.1748811

  • CrossRef
  • Google Scholar
  • View reference in article

• 37

  YuensookT.JantakoonT.LimpinanP. (2025). AI-driven adaptive learning systems in higher education: a systematic review. J. Educ. Learn.15 (2), 117–132. 10.5539/jel.v15n2p117

  • CrossRef
  • Google Scholar
  • View reference in article