We conducted a systematic review and meta-analysis following the PRISMA 2020 reporting guidelines ([54]). The review aimed to identify quantitative studies reporting the use of artificial intelligence among students and academic staff in African higher education institutions.

A structured search was performed across six major academic databases, including Scopus, Web of Science, ERIC, ScienceDirect, DOAJ, and AJOL. The full Boolean search strategy for PubMed was: (“artificial intelligence” OR “AI” OR “generative AI” OR “ChatGPT”) AND (“higher education” OR “university” OR “tertiary education”) AND (“Africa” OR names of individual African countries). Searches were conducted between January 2026 and March 2026, and results were limited to studies published in English. Reference lists of included studies were manually screened to identify additional relevant studies. Searches were limited to studies published between 2019 and 2026. Studies were eligible if they: 1) reported quantitative or mixed-methods empirical data, 2) involved university students or academic staff, 3) were conducted in African higher education settings, and 4) provided sufficient data to estimate prevalence. Qualitative-only studies, editorials, commentaries, and studies conducted outside higher education were excluded.

Two reviewers (AS and SM) independently screened titles, abstracts, and full texts for eligibility. Data extraction and quality appraisal were also conducted independently by two reviewers using a standardized form that captured study characteristics, including publication details, country, participant group, sample size, AI tools assessed, and the numerical data required for meta-analysis. Disagreements were resolved through discussion and, where necessary, consultation with a third reviewer. Methodological quality was assessed using the Joanna Briggs Institute (JBI) Critical Appraisal Checklist for Prevalence Studies.

The analysis focused on pooled prevalence estimates of self-reported AI use. AI use was defined as self-reported engagement with AI tools for academic purposes, including generative AI systems such as ChatGPT. Where studies differed in measurement (e.g., ever use, current use, or general exposure), these were harmonized as binary prevalence outcomes (use vs non-use) to enable pooled analysis. These differences were considered a potential source of heterogeneity. Meta-analysis was performed in R (version 4.5.1) using random-effects models to account for expected variation across countries, institutional settings, study populations, and definitions of AI use. Proportions were analyzed using logit transformation, and the between-study variance was estimated using restricted maximum likelihood (REML). Subgroup analyses were conducted by population type and region, where data permitted. Sensitivity analysis was performed using a leave-one-out approach, and influence diagnostics were explored to examine the contribution of individual studies to the pooled model.

Although only seven studies met the inclusion criteria, quantitative synthesis was undertaken to provide a structured summary of the emerging evidence rather than a precise continental estimate. Because the included studies varied in population, setting, and how AI use was defined, the pooled estimates were interpreted cautiously as summary measures across heterogeneous contexts.

Statistical heterogeneity was assessed using Cochran’s Q and the I² statistic, with higher I² values indicating greater inconsistency across studies. Because fewer than ten studies were included, formal statistical testing for publication bias or small-study effects was not considered reliable. Funnel plot assessment was therefore treated as exploratory only ([45]).

This study involved secondary analysis of published data and did not require ethical approval.

The database search yielded 60 records. After the removal of 20 ineligible records, 40 titles and abstracts were screened. Of these records, 20 records were excluded, leaving 20 full-text articles assessed for eligibility. Thirteen full-text articles were excluded due to the absence of quantitative prevalence data (n = 7), non-African study settings (n = 4), or ineligible study populations (n = 2). Seven studies met all inclusion criteria and were included in the meta-analysis. The study selection process is illustrated in Figure 1.

Figure 1. PRISMA 2020 flow diagram of study selection.

A total of seven cross-sectional studies published between 2024 and 2025 were included in the review. These studies were conducted in five African countries, Zambia, Nigeria, Tanzania, South Africa, and Uganda, and together involved 1877 participants. Of these, 1373 reported using artificial intelligence tools, giving a crude prevalence of 73.1% across the combined sample. This crude prevalence represents the unweighted proportion across all included participants and differs from the pooled estimate derived from the random-effects meta-analysis, which accounts for between-study variability. At the individual study level, the reported prevalence varied widely, ranging from 27.1% in a Nigerian academic staff study to 99.0% in a South African student study.

The majority of the studies included focused on students. Four studies involved student populations, with a total of 1289 participants, while three studies examined lecturers, faculty members, or academic staff, comprising a combined total of 588 participants. The studies were distributed across three African regions: two from Southern Africa, two from West Africa, and three from East Africa. East Africa accounted for the highest proportion of participants, with Southern Africa and West Africa contributing smaller shares respectively. All studies employed cross-sectional survey designs and utilized self-administered questionnaires for data collection. In most cases, surveys were conducted online, although a few studies used institution-based data collection methods. The research primarily assessed the use of generative AI tools, especially ChatGPT, with some also exploring broader categories of AI tools used for teaching, learning, research, or academic writing.

Reporting of participant characteristics was inconsistent across studies. Some studies provided age categories rather than summary measures such as mean or median age, while sex distribution was reported in only some papers. Where these details were available, male participants tended to be more represented in certain groups, especially in studies involving engineering, health, or technical education settings. Overall, the included studies reflected a diverse but still limited evidence base on AI use in African higher education. A summary of the study characteristics is presented in Table 1 and Table 2.

Table 1. Characteristics and findings of included studies on AI use prevalence in African higher education institutions.

Total N = 1877 from 5 countries; Overall crude prevalence = 73.1% (1373/1877); Population: students (n = 1289, 4 studies), lecturers/academic staff (n = 588, 3 studies); Regions: Southern Africa (n = 487, 2 studies), West Africa (n = 364, 2 studies), East Africa (n = 1026, 3 studies).

Table 2. Included studies and their bibliographic details.

Methodological quality was assessed using the Joanna Briggs Institute (JBI) Critical Appraisal Checklist for Prevalence Studies (Table 3). Overall, the risk of bias across the included studies was judged to be moderate. The overall “moderate” risk of bias rating reflects consistent limitations across studies, including non-probability sampling, unclear response rates, and reliance on self-reported data. While these limitations reduce confidence in the precision of prevalence estimates, the consistency of findings across studies supports cautious interpretation of the overall pattern of AI use. Only one study employed a probability-based sampling strategy.

None of the included studies reported formal sample size calculations or response-rate analyses. Assessment of AI use relied solely on self-reported data, which may introduce recall and social desirability bias. Nevertheless, all studies adequately described their settings, participant characteristics, and data collection tools. No study was excluded based on methodological quality. The predominance of non-probability sampling likely contributed to the high between-study heterogeneity observed.

Table 3. Risk of bias assessment of included studies using the Joanna Briggs Institute Critical Appraisal Checklist for prevalence studies.

Interpretation: Yes: criterion met (low risk of bias); No: criterion not met (high risk of bias); Unclear: insufficient information to judge; Overall: based on consensus across criteria.

Using a random-effects model with REML estimation and logit-transformed proportions, the pooled prevalence of AI tool use across the seven included studies was 79.7% (95% CI: 56.6% - 92.2%). The common-effect model gave a lower estimate of 71.2% (95% CI: 68.9% - 73.5%). Heterogeneity was very high (I² = 97.4%, τ² = 2.0925, p < 0.001), suggesting marked differences in the reported prevalence of AI use across studies. The study-specific estimates and the pooled random-effects estimate are shown in Figure 2.

Figure 2. Forest plot showing individual study prevalence estimates and the pooled prevalence of AI use under the random-effects model.

Considerable heterogeneity was observed across the included studies. The I² statistic was 97.4% (95% CI: 96.2% - 98.3%), while the estimated between-study variance under REML was τ² = 2.0925 (τ = 1.4466). Cochran’s Q-test confirmed that this heterogeneity was statistically significant (Q = 231.49, df = 6, p < 0.0001). This level of variation is likely to reflect real differences between studies, including differences in participant groups, institutional settings, countries, and the specific way AI use was assessed. As such, the pooled prevalence should be interpreted cautiously as an average across heterogeneous contexts rather than a precise continental estimate.

3.6.1. Population Type

When the studies were grouped by population, students showed a higher pooled prevalence of AI tool use than lecturers. Under the random-effects model, the pooled estimate was 87.1% (95% CI: 43.1% - 98.4%) for students and 64.2% (95% CI: 3.2% - 99.0%) for lecturers. However, this difference was not statistically significant (p = 0.2509). There was still substantial heterogeneity within both groups, especially among lecturers, which means these subgroup estimates should be interpreted with caution. A forest plot stratified by population type, illustrating pooled estimates for students and lecturers, is shown in Figure 3.

Figure 3. Forest plot of pooled prevalence of AI use stratified by population type (students vs lecturers).

3.6.2. Region

When grouped by region, the pooled prevalence was highest in Southern Africa (94.3%), followed by East Africa (76.0%) and West Africa (64.8%) under the random-effects model. However, the regional differences were not statistically significant (p = 0.5743), and the confidence intervals were very wide, especially for Southern and West Africa. This suggests that the regional subgroup findings are unstable and should be interpreted as exploratory rather than conclusive. Regional subgroup estimates and their associated confidence intervals are displayed in Figure 4.

Figure 4. Forest plot of pooled prevalence of AI use stratified by African region.

3.7.1. Leave-One-Out Analysis

Leave-one-out sensitivity analysis showed that the pooled prevalence remained broadly stable, ranging from 73.0% to 84.1% after sequential omission of individual studies. Heterogeneity remained high across most iterations, although omission of the Nigerian academic staff study (NGA2) produced the largest change, increasing the pooled prevalence to 84.1% and reducing heterogeneity to 91.6%. Overall, no single study substantially altered the direction of the findings as shown in Figure 5.

Figure 5. Leave-one-out sensitivity analysis of pooled prevalence estimates.

3.7.2. Influence Diagnostics

Influence diagnostics based on a random-effects logit-transformed prevalence model detected variation in how individual studies contributed to the pooled estimate and heterogeneity. The South African student study (ZAF1) and the Nigerian academic staff study (NGA2) were the most influential, with Cook’s distances of 0.439 and 0.281, respectively. The Ugandan faculty study (UGA2) showed a more moderate influence, while the other studies had comparatively small Cook’s distances. Consistent with this, leave-one-out sensitivity analysis indicated that excluding NGA2 led to the greatest reduction in heterogeneity and an increase in the pooled prevalence estimate; excluding ZAF1 also reduced heterogeneity but did not significantly alter the overall interpretation. Overall, no single study changed the conclusion that AI tool use in African higher education was widespread. The influence diagnostics plot of the included studies is shown in Figure 6.

Figure 6. Cook’s distance influence diagnostics for the included studies. Higher values indicate greater influence of individual studies on the pooled random-effects model.

Publication bias could not be assessed with confidence because the meta-analysis included only seven studies. Although a funnel plot was explored, its interpretation was limited by the small number of studies and the use of transformed prevalence estimates; therefore, any assessment of asymmetry was considered exploratory rather than conclusive.

This study has several strengths. It represents, to our knowledge, the first meta-analysis to quantitatively synthesize evidence on AI use prevalence within African higher education. The use of appropriate transformation methods, random-effects modelling, and extensive sensitivity analyses enhances the methodological rigor and transparency of the findings.

Several limitations should nonetheless be acknowledged. The relatively small number of available studies limits the precision of subgroup analyses and constrains formal assessment of publication bias. Reliance on self-reported measures of AI use may overestimate actual engagement, and the predominance of single-institution, non-probability samples limits generalizability. High contextual and measurement variability across studies further complicates interpretation. Definitions of “AI use” were not standardized: some studies assessed broad AI tool use, others measured practical or current academic use, and several focused specifically on generative systems such as ChatGPT. These operational differences mean that the pooled prevalence combines related but not fully identical constructs, likely contributing to the very high between-study heterogeneity observed. This limitation reflects the early and rapidly evolving nature of empirical research in this field rather than a weakness specific to the present review.

The widespread and often weakly governed adoption of AI tools in African higher education carries important implications for policy and institutional practice. Evidence from recent governance-focused studies indicates that institutional preparedness has not kept pace with the rapid diffusion of AI technologies ([59]). In this context, reactive approaches, such as blanket prohibitions or fragmented guidance, are unlikely to be effective. Instead, universities must shift toward structured, proactive governance approaches that recognize AI as an already normalized component of academic practice.

The policy implications and strategic responses emerging from this review are synthesized in Table4, which highlights a clear disconnect between the rapid, organic uptake of AI tools and the current state of institutional readiness within African higher education. The recommendations outlined emphasize the need for comprehensive governance frameworks that move beyond enforcement toward integration, focusing on ethical use, assessment design, and academic integrity. Importantly, the table underscores that high prevalence of use, in the absence of formal policy, poses risks not only to academic standards but also to equity and consistency across institutions.

Capacity-building emerges as a central priority. Although AI use is evident among both students and lecturers, existing evidence suggests that staff engagement is more variable and constrained by limited training opportunities, unclear institutional expectations, and uncertainty around appropriate academic use. Embedding AI literacy within curricula and professional development programs is therefore essential to support responsible and pedagogically meaningful use. Without parallel investments in staff capacity and digital infrastructure, regulatory efforts alone are unlikely to succeed.

Table4 highlights that while generative AI adoption in African higher education is widespread, its integration remains largely unregulated and driven by informal practices. This underscores the urgent need for structured institutional governance frameworks to guide ethical and effective use. Variability in AI competencies among lecturers and students points to gaps in capacity building, necessitating targeted training and curriculum integration. Additionally, persistent infrastructural limitations and digital inequalities continue to hinder equitable access to AI technologies. The fragmented nature of existing research further limits evidence-based policymaking, emphasizing the need for more robust and standardized studies. Importantly, the growing use of AI raises critical concerns around academic integrity, requiring policy reform and assessment redesign. Overall, a coordinated, systems-level approach is essential to ensure responsible, equitable, and sustainable AI integration in African higher education. Future research should prioritize standardized measurement of AI use, multi-institutional sampling, and longitudinal designs to better capture adoption trajectories and policy effects across African higher education systems ([52]; [18]; [59]; [40]).

The implications of these findings extend beyond Africa. As AI adoption accelerates globally, the challenges observed in African higher education, particularly around governance, equity, and infrastructure, offer important lessons for other resource-constrained settings undergoing digital transformation.

Importantly, the very high heterogeneity (I² = 97.4%) indicates that AI adoption is highly context-dependent, reinforcing the need to interpret pooled estimates as descriptive summaries rather than precise continental metrics.

Table 4. Policy implications and strategic recommendations for AI integration in African higher education.