Artificial Intelligence Research in Africa and the Importance of Recall Concept

INTRODUCTION

Artificial Intelligence (AI) is reshaping global innovation ecosystems, offering transformative potential across sectors such as healthcare, agriculture, education and infrastructure. In Africa, AI presents unique opportunities to address systemic challenges and support inclusive development (Trabelsi, 2024). Examples range from improving healthcare accessibility in remote areas to advancing precision agriculture in regions affected by adverse climate change conditions to facilitating the improvement and development in infrastructural facilities and systems However, the continent’s AI research landscape remains underexplored, particularly in terms of its scientometric characteristics, funding dynamics and collaboration networks (AUDA-NEPAD, 2021).

The current article leverages scientometric analyses to explore the current state and trajectory of AI research and application across Africa. By analysing scientometric patterns in academic research, collaborative networks and funding trends, we aim to highlight Africa's emerging role in AI innovation but also to identify potential areas for growth, collaboration and investment to drive the Continent's technological future forward.

A key methodological concern addressed in this article is the challenge of recall rate in scientometric investigations. Recall-the proportion of relevant publications successfully retrieved by a search strategy-is often underemphasized in AI-related bibliometric studies (Zhang et al., 2008). This oversight can lead to incomplete datasets, biased interpretations, and ultimately, flawed policy recommendations. In fast-evolving fields like AI, where terminology is diverse and rapidly changing, low recall can significantly distort assessments of national research performance, funding impact and dynamics collaboration. By foregrounding recall as a critical metric, this article seeks to improve the methodological robustness of African AI research mapping and offer more accurate insights for evidence-based policy and strategic planning. The article also makes a substantive contribution in the methodological frameworks utilised in scientometric investigations in fast growing research fields like the artificial intelligence one.

The literature review elaborates on the scientometric approaches utilised for investigating issues of artificial intelligence and the methodology section focuses on the approach we utilise for the current investigation. The findings section describes our findings, and the section discussion elaborates on the findings and develops recommendations.

LITERATURE REVIEW

Scientometric approaches are increasingly employed to analyze the evolution and impact of Artificial Intelligence (AI) research both globally and within Africa. These methods offer valuable insights into publication trends, collaboration networks and thematic developments. Common techniques include keyword searches, bibliometric mapping, co-citation analysis and bibliographic coupling.

Keyword Searches involve using specific terms or phrases to locate relevant articles in databases. Their effectiveness is governed by two key metrics: precision (the proportion of retrieved documents that are relevant) and recall (the proportion of relevant documents that are retrieved). As Buckland (1994) notes, precision and recall are often tension-improving one may reduce the other. In fast-evolving fields like AI, where terminology proliferates rapidly, recall becomes especially critical for comprehensive analysis.

Discipline Categorization is a foundational step that determines how publications are grouped and analysed by subject area. Major bibliometric databases like Web of Science (WoS) and Scopus, use structured classification systems to assign publications to disciplines. For example, in the WoS each journal is assigned to one or more of ~250 subject categories (e.g., “Computer Science: Artificial Intelligence”; “public health” etc.). Scopus also uses a journal classification system called the All Science Journal Classification (ASJC) system. ASJC includes ~330 subject areas grouped into 27 broad disciplines (e.g., “Engineering”, “Computer Science” etc.,). Analyses based on categories reflecting journal-level discipline boundaries may not capture article-level interdisciplinarity or articles published in multidisciplinary journals.

Bibliometric Mapping provides spatial representations of how disciplines, fields, specialties and individual documents are related (Small, 1999). Tools such as VOSviewer and CiteSpace are commonly used to visualize these relationships.

Co-citation Analysis measures how frequently two documents are cited together, revealing intellectual linkages (Small, 1973). Bibliographic Coupling identifies connections between documents that cite common references (Martyn, 1964).

Full-Text and Abstract Mining refer to using natural language processing or machine learning to analyse full texts or abstracts for relevance (Yau et al., 2014).

Despite the utility of these methods, few studies have applied them rigorously to African AI research. Foundational works by (Pouris et al., 2014; Sooryamoorthy, 2021) provide broad assessments of African research systems but do not delve deeply into AI-specific dynamics. More recent studies have begun to fill this gap, though many rely on limited keyword strategies that constrain their scope and recall.

A number of studies use limited keyword strategies to achieve their objectives. Examples include the following: (Njei et al., 2023) investigated AI applications in African healthcare using PubMed, Scopus, and Web of Science. Their search terms included clusters related to AI (“Artificial Intelligence,” “Machine Learning,” “Deep Learning”), technical processes (“Feature extraction,” “Classification,” “Detection”), and performance metrics (“Precision,” “Recall,” “F-score”). Despite this structured approach, only 26 relevant articles were identified across 31 countries. The authors also performed backward citation analysis but did not employ iterative keyword expansion or subject category filters, limiting the comprehensiveness of their dataset. (Kondo et al., 2023) examined general AI research in Africa from 2013 to 2022 using Scopus. Their keyword strategy was broad but shallow, relying primarily on “Artificial Intelligence,” “Machine Learning,” “Deep Learning,” and “Natural Language Processing.” This yielded 1,644 publications-higher than Njei et al., but still modest compared to high-recall studies. The absence of term expansion and journal filters likely contributed to under coverage of emerging subfields and interdisciplinary work.

(Lianou et al., 2022; Pandey et al., 2021) conducted global AI reviews using short keyword strings, often restricted to “AI,” “ML,” or “DL.” These studies, while methodologically sound in other respects, suffer from low recall and are less suitable for regional or sectoral benchmarking.

• In contrast with the above investigations (Liu et al., 2021) developed a multi-stage scientometric approach to maximize recall while maintaining conceptual precision. Their methodology involved: Benchmarking Core Records using a combination of keyword and journal filters.
• Keyword Extraction and Frequency Analysis from “Author Keywords” and “Keywords Plus.”
• Co-occurrence Analysis and Manual Refinement to validate term relevance.
• Augmentation via Subject Category Search using the Web of Science category “Computer Science: Artificial Intelligence.”

The approach is described in more detail in the Methodology section

Liu et al., compared their results with those of (Gao et al., 2019; Zhou et al., 2020; WIPO, 2021). Their approach retrieved 100% of Gao’s records and 78% of Zhou’s, while excluding over 50% of WIPO’s broader but less precise dataset. This demonstrated the superiority of high-recall strategies in capturing the full spectrum of AI research.

In the African context, where AI research is emerging across diverse sectors and institutions, high-recall methodologies are essential to ensure visibility of underrepresented actors and themes. Studies with limited keyword sets risk producing skewed or incomplete findings, which can misinform policy decisions, funding allocations and strategic planning. As Liu et al., argue, the use of hundreds of validated keywords, journal filters and subject categories is not merely a technical enhancement but a methodological imperative.

To summarise the reliance on short or generic search strings in many scientometric studies risks underestimating the true scale and diversity of AI research. This has serious implications particularly for scientifically small countries and expensive scientific disciplines. Incomplete data may lead to skewed assessments of national research capacity, affecting funding and strategic planning. Furthermore, low-recall studies may overlook emerging areas or underrepresented institutions, distorting investment priorities. Finally, it should be emphasised that publishing studies with flawed methodologies in peer-reviewed journals undermines the integrity of scientometric research.

METHODOLOGY

This investigation adopts and adapts the high-recall scientometric methodology developed by (Liu et al., 2021), which was specifically designed to overcome the limitations of narrow keyword-based searches in fast-evolving fields such as Artificial Intelligence (AI). Traditional bibliometric studies often rely on a small set of generic terms-such as “Artificial Intelligence” or “Machine Learning”-which can lead to significant under-coverage of relevant literature. Liu et al. proposed a multi-stage, systematic approach to maximize recall while maintaining conceptual precision, thereby improving the robustness of scientometric analyses and policy-relevant insights.

The first step involved identifying a set of core AI publications to serve as benchmarking records. This was achieved by combining two filters:

• A core keyword search using “Artificial Intelligence” in the title, abstract, and keywords.
• A journal filter, selecting publications from 19 leading AI-focused journals indexed in the Web of Science (WoS) and Scopus databases.

These benchmarking records were assumed to represent high-confidence AI literature and formed the basis for subsequent keyword expansion.

The second step involved keyword extraction and frequency analysis. Using the benchmarking records, Liu et al., extracted two types of metadata:

• Author Keywords: Terms provided by authors to describe their work.
• Keywords Plus: Terms algorithmically generated by WoS based on the titles of cited references.

The frequency of each keyword was calculated, and high-frequency terms were subjected to semantic validation. This involved cross-referencing definitions and usage contexts from authoritative online sources (e.g., Wikipedia, ACM Digital Library, IEEE Xplore) to ensure that each term was genuinely relevant to AI and not a generic statistical or computational concept.

The next step involved co-occurrence analysis and manual refinement. To refine the keyword set, Liu et al., conducted co-occurrence analysis using tools such as VOSviewer and CiteSpace. This helped identify clusters of related terms and eliminate outliers or ambiguous keywords. Manual inspection was also employed to validate the inclusion of terms based on their conceptual relevance to AI. For example, terms like “neural network,” “reinforcement learning” and “natural language processing” were retained, while more generic terms like “logistic regression” or “fuzzy logic” were excluded unless they appeared in explicitly AI-related contexts.

The final step involved augmenting the refined keyword set with a subject category filter. Specifically, Liu et al., used the WoS category “Computer Science: Artificial Intelligence” to capture additional relevant records that may not have been retrieved through keyword searches alone. This hybrid strategy-combining keyword expansion with subject classification-ensured both breadth and depth in the dataset.

To validate their methodology, Liu et al., compared their results with three prominent studies:

• Gao et al., (2019): Focused on AI patenting and publication trends.
• Zhou et al., (2020): Used a narrower keyword set to map global AI research.
• WIPO (2021): Employed a broader search strategy that included also statistical and mathematical terms.

Liu et al., successfully retrieved 100% of Gao’s records and 78% of Zhou’s, demonstrating strong coverage. However, they found that 54% of WIPO’s results included documents that did not meet their conceptual definition of AI-highlighting the trade-off between recall and precision. Their critique emphasized the importance of excluding overly generic terms to avoid inflating AI publication counts with unrelated literature.

For the current investigation, we replicate Liu et al., methodology to assess the scale and scope of AI research across the African continent. We applied the refined keyword set, journal filters, and subject category search to the WoS and Scopus databases, focusing on the period 2014-2023. The full list of search terms, journal titles, and classification criteria is provided in Appendices 1 and 2. This approach ensures high recall and conceptual precision, enabling a more accurate mapping of Africa’s AI research landscape and informing evidence-based policy recommendations.

AI Research in Africa

Our investigation identifies that African researchers were involved in 45 937 AI publications indexed in the WoS during the period 2014-2023. The drop during 2023 can be argued that it is the result of the pandemic of COVID 2019- a post pandemic correction. During the COVID-19 years (2020-2022), there was an explosive surge in publications, especially in health and biomedical fields. Researchers rushed to publish findings on vaccines, treatments, and public health strategies. In 2023 the publishing ecosystem began to normalize returning to pre-pandemic growth rates (Whitaker et al., 2025).

Figure 1 shows the annual number of AI publications on the Continent. AI publications increased from 1659 during 2014 to 6756 during 2023. The Main contributing countries were Egypt with 11920 publications during the period; South Africa with 7221 publications; Tunisia 6987 and Morocco 6013 publications.

Click to enlarge

We also identified the type of documents produced during the period from the top two countries- Egypt and South Africa. Egypt’s publications were 76% articles and 19.3% proceeding papers. South Africa’s publications were 65.3% articles and 26.7% proceeding papers. A recent publication (Pouris, 2025) it was estimated that “In the total SA publications, the average proceedings is 12%.” (p 6). Furthermore, it was argued that “NRF supports traveling to conferences. This is probably one of the few ways that a researcher can benefit from NRF funding. Based on this assumption it can be argued that energy researchers are supported by NRF to participate in international conferences, and this affects their outputs in proceedings”. (p 12). As NRF is also the main funder of AI research it can be argued that this is the reason that AI research produces twice the number of proceedings 26.7% as the average of the country (12%).

The focus of the AI publications was Computer science 27 219 publications and engineering 15 189 publications. The two disciplines cover 92.3% of the total number of AI publications.

Table 1 shows that Northern Africa countries produce most publications in the Continent (68%). Southern Africa follows and the region with the least number of AI publications is Central Africa.

Figure 2 shows the AI publications per region. It is interesting to notice that Northern African countries were well advanced in the AI field in the beginning of the under-examination period. Northern African countries produced 5885 AI publications during 2022 just before their decline to 4405 publications during 2023.

Click to enlarge

Tables 2 to 6 identify the main funders of the articles in the five African regions. Identification of the main funders as they are identified in the publications by the authors show that the National Natural Science Foundation of China (NSFC) is the main funder. NSFC is top funder in North African countries (556 listings) and in Eastern (144) and Western Africa countries (180). It is second in the Southern Africa countries (207) where the National Research Foundation of South Africa appears to be the top funder with approximately 900 listings (with name permutations).

King Saud University (555) appears as second funder in Norther African countries with the same number of listings as NSFC. The UK Research Innovation appears third in Southern African countries.

Overall, it is apparent that National Natural Science Foundation of China is the main funder of AI in Africa. Arab organisations appear to dominate Northern Africa countries research. It should be noted while China has a presence across all regions, NRF provides support only within South Africa.

The recent political changes in the USA should be expected to affect negatively the support by USA organisations.

Table 7 shows the collaboration matrix of the five African regions among themselves, and Table 8 shows the collaboration with non-African countries. It becomes apparent that collaboration with foreign countries dominates the collaboration domain.

FINDING AND DISCUSSION

This article aims to provide a realistic picture of AI research in the African Continent using an appropriate methodology. It is identified that Africa produced 45 937 AI publications indexed in the WoS during the period 2014-2023.

It should be mentioned that in a similar investigation (Kondo et al., 2023) 1644 publications identified in SCOPUS during the decade 2013-2022. The more detailed approach in searching for AI publications from Africa identified just below 28 times more publications than the search with a limited number of search terms. It is apparent that limited attention to recall issues can severely underestimate the publications in the field and provide misleading conclusions. It is interesting to note that the set of key words identified 33 179 African AI publications. The category indexed in the WoS as Artificial Intelligence identified 19 145 African publications. Finally, the search for relevant publications in the journals identified as journals focusing on AI identified only 462 relevant publications. It is apparent that key words searching cover a wider spectrum of articles than the other approaches. Furthermore, it is emphasised that a relatively recent study comparing the databases (with shortcomings) found 27 million documents in Scopus and 23 million in Web of Science, with an overlap of 18 million (Visser, et al., 2020). Hence, Scopus was expected to cover more AI documents than WoS. Such challenges in recall rate will also lead to wrong conclusions (e.g. which country is more productive in the field and where it is the searched for specialization).

Our findings show that AI publications increased from 1659 during 2014 to 6756 during 2023. The Main contributing countries were Egypt with 11920 publications and South Africa with 7221 publications during the period.

Regional analysis identifies that the Northern African countries dominate the AI publications domain. Identification of the main funders, as they are identified in the publications by the authors, show that the National Natural Science Foundation of China (NSFC) is the main funder followed by King Saud University (555) which appears as second funder in Northern African countries with the same number of listings as NSFC. South African publications appear to be supported by the National Research Foundation.

Comparisons of the collaboration patterns of the regions among themselves and with non-African countries identifies that African countries collaborate mainly with the rest of the World and less among themselves. This finding confirms those by (Pouris et al., 2015). It was argued that African researchers are attracted by research resources and funding which is almost always abroad.

Identification of the type of publications produced indicates that the field of AI is affected by the way the National Research Foundation supports research and the field of AI produces twice as many proceedings that the average scientific discipline in the country.

CONCLUSION

The analysis demonstrates that AI research in Africa is significantly underestimated when limited search strategies are used, leading to inaccurate assessments of productivity and specialization. A comprehensive search approach captures a far larger body of work-up to 28 times more-highlighting the importance of addressing recall issues in bibliometric studies. AI publications across the continent have grown substantially over the past decade, with Northern Africa dominating output and Egypt and South Africa emerging as leading contributors. Funding patterns show a strong influence from international agencies, particularly the National Natural Science Foundation of China and King Saud University. Collaboration trends reveal that African researchers partner more frequently with international collaborators than with other African countries, reinforcing earlier findings and reflecting the pull of external research resources. Additionally, the prominence of conference proceedings in AI output indicates the impact of national funding frameworks on publication types within the field.