Computational Social Science and African Realities: A Methodological and Theoretical Critique of Big Data, Machine Learning and the Measurement of African Societies

Abstract

Social scientists increasingly measure African societies through data that Africans did not produce for that purpose: satellite imagery, night-time lights, mobile-phone and platform data, social-media text, conflict event records and, most recently, large language models, processed by machine learning into estimates of poverty, population, inequality, conflict and opinion where surveys and censuses are missing or late. Evaluations of these methods focus mainly on predictive accuracy, and the problems they raise are discussed in separate literatures that seldom engage interpretive African social science. This paper offers a methodological and theoretical critique. Its objectives are to integrate recent evidence into a single framework, derive propositions on the conditions of validity and the policy consequences of computational measurement, clarify the relationship between computational and interpretive research, and propose standards for responsible use. The study is an integrative, theory-building review. The population was peer-reviewed journal articles published between 2023 and 2026; a purposive sample of 41 articles was retained from 73 fully screened records after verification against publisher records and screening for retractions, corrections and expressions of concern. Findings were coded against three provisional problems, compared across data sources and synthesised abductively. The analysis yields a coverage–validity–reflexivity framework. Coverage gaps are structured, falling on poorer, rural, less educated and female populations; construct validity weakens as the distance between signal and concept grows; and measurement used for policy changes what it measures. Computational methods therefore ease data scarcity while introducing biases harder to see than those of the surveys they supplement. Seven propositions address structured coverage, inherited labels, proxy distance, invisible bias, reflexivity, measurement authority and complementarity with interpretive research, illustrated with Nigerian cases. The paper recommends coverage statements, disaggregated local validation with reported uncertainty, explicit construct-validity arguments, reflexivity assessments and appeal mechanisms before computational measures are used for allocation, partnership with national statistical offices, African research leadership, and caution in substituting language-model outputs for African respondents.