The Commonplace
Home Dashboard Papers Evidence Syntheses Digests 🎲
← Papers

In South Africa, AI adoption depresses development in the short term as routine-task substitution and skill mismatches bite, but contributes positively in the long run once structural adjustments occur; crucially, unemployment among highly educated workers persistently undermines development while aggregate unemployment conceals these distributional effects.

Artificial Intelligence, Disaggregated Unemployment, And Sustainable Development In South Africa: Does The Routine-Biased Technological Change Hypothesis Explain The Dynamics?
Felix Aberu, Tersia Botha · May 29, 2026 · Acta Academiae Beregsasiensis Economics
openalex correlational medium evidence 7/10 relevance DOI Source PDF
Using ARDL cointegration on South African annual data (2003–2024), the study finds AI adoption harms sustainable development in the short run via routine-task substitution and adjustment costs but aids development in the long run through productivity and innovation spillovers, with skill-disaggregated unemployment showing divergent effects.

This study examines the relationship between artificial intelligence (AI) adoption and disaggregated unemployment in determining sustainable development in South Africa, grounded in the Routine-Biased Technological Change (RBTC) framework. Using annual time-series data from 2003–2024 and the Autoregressive Distributed Lag (ARDL) modelling approach, the study simultaneously estimates short- and long-run coefficients to capture full developmental dynamics. Empirical results confirm a long-run equilibrium relationship among AI adoption, skill-disaggregated unemployment, and sustainable development. In the short run, AI adoption negatively impacts development due to adjustment costs from routine-task substitution, labour market rigidities, and skill mismatches. In the long run, AI contributes positively and significantly through productivity gains and innovation spillovers once structural adjustments are completed. Regarding skill-specific unemployment, highly educated workers' unemployment consistently impedes development across both horizons, reflecting the productivity costs of underutilised advanced human capital. Unemployment among less-educated workers shows a positive long-run relationship, suggesting structural labour reallocation effects consistent with RBTC predictions. Total unemployment proves statistically insignificant, confirming that aggregate measures mask critical distributional differences across skill groups, and thus validating the core theoretical proposition that skill-biased technological change operates through heterogeneous channels that remain invisible at the aggregate level. The findings are particularly relevant for policymakers in emerging economies facing the dual challenge of accelerating technological transformation and persistent structural unemployment. The study recommends integrated policy frameworks combining AI development incentives, labour market reform, and education strategies to ensure that technological progress translates into inclusive and sustainable development outcomes.

Summary

Main Finding

Using annual South African data (2003–2024) and an ARDL approach, the paper finds a long-run equilibrium relationship among AI adoption, skill-disaggregated unemployment, and sustainable development. Short run: AI adoption depresses sustainable development (adjustment costs from routine-task substitution, labour-market rigidities, skill mismatch). Long run: AI contributes positively and significantly via productivity and innovation spillovers after structural adjustment. High-skilled unemployment persistently impedes development in both horizons; low-skilled (less-educated) unemployment shows a positive long-run association (consistent with RBTC-driven labour reallocation). Aggregate unemployment is statistically insignificant, indicating important distributional heterogeneity masked at the aggregate level.

Key Points

  • Theoretical framing: Routine-Biased Technological Change (RBTC) — AI substitutes routine tasks, complements non-routine cognitive/interactive tasks, producing heterogeneous labour-market effects.
  • Short-run vs long-run dynamics:
    • Short run: negative net effect of AI on sustainable development due to displacement costs, market rigidity and skill mismatch.
    • Long run: positive net effect from productivity growth and innovation spillovers once adjustment completes.
  • Skill-disaggregated unemployment matters:
    • High-skilled unemployment lowers sustainable development in both short and long run — underutilised advanced human capital generates productivity costs.
    • Low-skilled unemployment shows a positive long-run relation, interpreted as structural labour reallocation consistent with RBTC.
    • Total (aggregate) unemployment is not significant — aggregate metrics conceal crucial heterogeneity.
  • Policy concern: “Premature Automation Trap” — rapid automation in labour-intensive sectors before adequate institutional, income and workforce readiness can worsen structural unemployment and hinder inclusive development.
  • Policy recommendation summary: integrated framework combining AI development incentives, labour-market reform, and education/skills upgrading to ensure inclusive and sustainable outcomes.

Data & Methods

  • Data: Annual time-series (South Africa), 2003–2024. Key series include an AI adoption indicator, unemployment disaggregated by education/skill groups (highly educated vs less-educated), and a composite/indicator measure of sustainable development (SDG-relevant outcome).
  • Econometric approach:
    • Autoregressive Distributed Lag (ARDL) modelling to simultaneously estimate short- and long-run coefficients and test for cointegration (bounds testing / long-run equilibrium).
    • The ARDL results show cointegration among AI adoption, skill-disaggregated unemployment, and sustainable development.
    • The paper situates findings within RBTC theory and references related VECM/empirical literature showing substitution effects for routine jobs.
  • Identification caveats noted by authors: adjustment costs, labour-market rigidity, skill mismatches and the potential need to consider sectoral heterogeneity and microdata for causal claims.

Implications for AI Economics

  • Empirical modelling: studies of AI and labour should disaggregate unemployment by skill/education — aggregate unemployment conceals distributional channels crucial to welfare and development outcomes.
  • Dynamics matter: separate short-run (adjustment) from long-run (productivity/innovation) effects; policy responses must address both horizons.
  • Policy design:
    • Combine AI diffusion incentives with active labour-market policies (retraining, mobility support), labour-market reforms to reduce rigidity, and education-system reforms to close skill mismatches.
    • Guard against premature automation in labour-intensive sectors; sequence automation with institutional and workforce readiness to avoid long-term exclusion.
  • Research directions for AI economics:
    • Use sectoral and worker-level microdata to trace reallocation and wage dynamics across skills and occupations.
    • Model interaction effects between AI adoption, institutional conditions (labour regulation, social protection), and human-capital investment.
    • Employ causal identification strategies (natural experiments, firm-level adoption shocks) to isolate displacement vs complementarity channels.
    • Incorporate multidimensional development outcomes (not just employment) when assessing AI’s welfare impact.
  • Relevance: findings are especially salient for emerging economies facing rapid AI diffusion alongside persistent structural unemployment and inequality; policies must be integrative to convert technological progress into inclusive, sustainable development.

Assessment

Paper Typecorrelational Evidence Strengthmedium — The ARDL/cointegration approach is appropriate for detecting long-run associations and short-run dynamics in time-series data, but it cannot fully rule out endogeneity, omitted variables, measurement error in AI adoption, or reverse causality; the small number of annual observations (2003–2024) and single-country focus reduce causal confidence relative to experiments or strong quasi-experiments. Methods Rigormedium — Using ARDL and error-correction modelling with stationarity and bounds-testing shows methodological competence for time-series analysis, but the approach is constrained by limited sample size, potential model specification sensitivity (lag selection, omitted controls), and reliance on aggregate proxies for AI adoption and skill-specific unemployment without external instruments or robustness to alternative causal strategies. SampleCountry-level annual time-series for South Africa covering 2003–2024 (approx. 22 observations) including measures of AI adoption (proxy), sustainable development indicator(s), and unemployment rates disaggregated by skill/education groups (highly educated vs less-educated), plus control variables used in the ARDL specifications. Themeslabor_markets productivity IdentificationARDL (Autoregressive Distributed Lag) bounds-testing and error-correction specification to estimate short- and long-run cointegrating relationships between AI adoption, skill-disaggregated unemployment, and a sustainable development outcome; identification relies on lag structure, stationarity/cointegration diagnostics, and control variables rather than exogenous variation or instruments. GeneralizabilitySingle-country (South Africa) macro time-series limits external validity to other economies, especially advanced economies, Annual data and short time span (≈22 observations) reduce power and may miss faster dynamics, AI adoption likely proxied (e.g., patents, investment, ICT indicators), so measurement error may bias estimates and limits applicability where different proxies are valid, Aggregate national-level analysis masks within-country heterogeneity (industries, regions, firm sizes), Institutional and labor-market features (rigidities, education systems) that shape results may not hold in other emerging or developed contexts

Claims (9)

ClaimDirectionConfidenceOutcomeDetails
The study uses annual time-series data from 2003–2024 and the Autoregressive Distributed Lag (ARDL) modelling approach to estimate short- and long-run coefficients. Other mixed high methodological approach / estimation of short- and long-run relationships
n=22
0.3
There is a long-run equilibrium (cointegrating) relationship among AI adoption, skill-disaggregated unemployment, and sustainable development in South Africa. Fiscal And Macroeconomic mixed high sustainable development (long-run cointegration with AI adoption and skill-disaggregated unemployment)
n=22
0.3
In the short run, AI adoption negatively impacts sustainable development due to adjustment costs from routine-task substitution, labour market rigidities, and skill mismatches. Fiscal And Macroeconomic negative high sustainable development (short-run effect of AI adoption)
n=22
0.3
In the long run, AI adoption contributes positively and significantly to sustainable development through productivity gains and innovation spillovers after structural adjustments are completed. Fiscal And Macroeconomic positive high sustainable development (long-run effect of AI adoption)
n=22
0.3
Unemployment among highly educated workers consistently impedes sustainable development across both short- and long-run horizons. Fiscal And Macroeconomic negative high sustainable development (effect of highly educated workers' unemployment)
n=22
0.3
Unemployment among less-educated workers shows a positive long-run relationship with sustainable development, interpreted as reflecting structural labour reallocation effects consistent with RBTC. Fiscal And Macroeconomic positive high sustainable development (effect of less-educated workers' unemployment)
n=22
0.3
Total (aggregate) unemployment is statistically insignificant in explaining sustainable development, indicating aggregate measures mask critical distributional differences across skill groups. Fiscal And Macroeconomic null_result high sustainable development (effect of total unemployment)
n=22
0.3
The empirical findings validate the core theoretical proposition of Routine-Biased Technological Change that skill-biased technological change operates through heterogeneous channels invisible at the aggregate level. Governance And Regulation positive high support for RBTC theoretical proposition (heterogeneous channels of skill-biased technological change)
n=22
0.3
Policymakers in emerging economies should adopt integrated policy frameworks combining AI development incentives, labour market reform, and education strategies to ensure technological progress translates into inclusive and sustainable development. Governance And Regulation positive high policy recommendation aimed at improving inclusive and sustainable development outcomes
n=22
0.05

Notes