Rethinking assessment in the age of generative AI: A systematic literature review of African higher education

Synthesis: This PRISMA 2020 systematic review asks how educators in African higher education are redesigning assessment in response to Generative AI, and synthesizes ten empirical studies published between 2023 and 2025, mainly from Nigeria and South Africa. Its central finding is a gap: students use generative AI far more than their lecturers do, and institutional guidance lags behind both. Faculty attitudes are mixed rather than hostile, shifting away from AI Detection and prohibition toward redesign. Three strategies recur: AI-resilient tasks demanding contextual reasoning, transparent use of AI as a learning partner, and a shift from product-oriented to Process-Oriented Assessment. Only 5 of 26 surveyed South African institutions had publicly available AI policies. The review frames AI in Africa as a catalyst for pedagogical change rather than a threat to ban, while warning that uneven infrastructure and thin governance constrain what is achievable. Its evidence base is small and ChatGPT-centric, so its recommendations are directions rather than settled effects.

Key Findings

1. Students use generative AI substantially more than educators do, a gap reported across Nigeria, South Africa, and the multi-country study that erodes confidence in assessment validity.
2. Faculty attitudes are mixed, not prohibitive. Acceptance of ChatGPT was generally low among 102 surveyed lecturers, yet greater exposure predicted significantly more favorable perceptions (p = 0.001).
3. Institutional guidance is thin. Only 5 of 26 surveyed higher education institutions in South Africa had publicly available AI policy documents.
4. Redesign, not prohibition, is the most recommended response, with tasks rebuilt to require contextual reasoning generic AI answers cannot supply.
5. Process-oriented assessment is the most consistently recommended strategy, through reflective portfolios, oral examinations, or multi-stage assignments that reward reasoning over final products.
6. The evidence base is narrow: ten studies from 2023 to 2025, mainly Nigeria and South Africa, with small samples, single-institution case studies, and self-reported perceptions.

How the review was done

The review follows PRISMA 2020 and uses the PICo framework to frame its question: how are educators in African higher education redesigning assessment in response to generative AI? Database searches identified 185 records and a further 12 through Google Scholar, institutional repositories, backward reference-list searching, and targeted manual searches, yielding 197 records. Following removal of 32 duplicates, 165 unique records remained for title and abstract screening, where 141 records were excluded. The remaining 24 full-text articles were independently assessed by two reviewers, and fourteen were excluded, leaving 10 studies. Given the methodological diversity of the included studies, statistical meta-analysis was not appropriate; an inductive narrative thematic synthesis was used.

What the studies found about educators and students

A knowledge gap runs through the evidence. Tunde-Awe (2024) surveyed 43 language educators and 505 pre-service teachers and found students using ChatGPT extensively for academic work while educators showed comparatively low familiarity, raising questions about traditional assessment validity. Opesemowo et al. (2024) reported generally low attitudinal acceptance among 102 lecturers, with experience dividing responses. Maistry and Singh (2025), drawing on 29 academics at a South African public university, found optimism and caution together: participants saw value in AI for assessment design and Feedback, while worrying about integrity, over-reliance, bias, and absent policy. Student-teacher studies were equally uneven: Makwara et al. (2024) found limited awareness and use among 12 student teachers, whereas Jere et al. (2024) found 11 pre-service teachers weighing potential and limitation.

Three strategies for redesigning assessment

Three complementary strategies emerged. The first is AI-resilient task design: Stack (2023) showed that questions can be rebuilt around contextual reasoning and authentic application of knowledge, and that educators can use generative AI itself to test draft questions and find prompts requiring deeper engagement than generic answers provide. The second is transparent integration of AI as a learning partner, where students evaluate AI-generated outputs critically and justify their own decisions. The third, and the most consistently recommended, is the shift to process-oriented assessment, rewarding reasoning, reflection, and justification rather than final products, reinforced by Tarisayi (2024) and Abubakar et al. (2024) with their emphasis on human oversight and rigorous standards. Across formats, the common move is from content reproduction to demonstrable evidence of learning.

Governance, access, and the African context

Chaka et al. (2024) analyzed AI policies across 26 South African universities and found that only five institutions had publicly available generative AI policies during the review period, with most adopting advisory guidelines or general academic honesty codes. Where policies existed, they required students to disclose AI assistance rather than banning it outright, connecting to AI Use and Disclosure Statements and Academic Integrity. The discussion also insists on equitable access: unequal digital infrastructure and uneven connectivity shape who benefits, so AI literacy should become an explicit learning outcome, and AI-responsive assessment in Africa should reflect local realities rather than replicate models developed elsewhere.

What this means for practice

• Instructors: move marks from final answers to the process, requiring students to document iterations and explain how AI output was used, verified, or rejected.
• Instructors: rebuild assignments around contextual reasoning and justification, and pilot draft questions against generative AI to find prompts generic answers cannot satisfy.
• Assessment designers: since AI-generated text defeats detection-based approaches, carry assessment validity through task design rather than detection scores.
• Administrators: publish explicit acceptable-use guidance with disclosure requirements, since only 5 of 26 surveyed institutions had public policy, and fund faculty development.
• Institutional leaders and policymakers: treat AI literacy as an explicit learning outcome and invest in connectivity so AI-enhanced assessment does not widen inequalities.

Limitations

• Ten studies, predominantly small samples, single-institution case studies, and self-reported perceptions, describing reported practice rather than measured learning gains.
• Coverage is concentrated in Nigeria and South Africa, leaving large regions of the continent underrepresented.
• The literature focuses almost exclusively on ChatGPT; coding assistants such as GitHub Copilot were not empirically examined.
• Few studies evaluated whether redesigned assessment improves outcomes; the thematic synthesis also reflects the reviewers' interpretation.

Citation

Antwi-Boampong, A., Nketia, M. O., Loglo, F. S., Frimpong, N., & Osei-Fosuaa, L. (2026). Rethinking assessment in the age of generative AI: A systematic literature review of African higher education. Social Sciences & Humanities Open, 14, 103332.