Public spending helps lower unemployment on average, but its power fades as automation and AI accelerate; in periods and countries with faster technological change, fiscal policy shows markedly weaker employment effects.

Main Finding

The paper finds that the unemployment-reducing (employment-enhancing) impact of public expenditures weakens as technological transformation accelerates. Using a dynamic panel threshold model for 38 countries (2000–2021) with R&D intensity as the threshold variable, the authors estimate an R&D threshold of 1.02% of GDP; public spending’s effect on unemployment differs across regimes defined by this threshold, and becomes less effective in economies with higher R&D intensity (i.e., faster technological change).

Key Points

• Research question: Does the effectiveness of public expenditures in lowering unemployment depend on the pace of technological transformation?
• Theory: Builds on SBTC and task-based technological change frameworks and on a model (Acemoglu/Restrepo; Ebeke & Eklou) showing fiscal policy’s employment effect depends on automation pace and substitution between capital and labor.
• Main empirical result: R&D/GDP threshold = 1.02%. Below this threshold the estimated coefficient on public expenditure is positive but statistically insignificant (β1 = 0.169), while above the threshold the authors report that the employment-enhancing effect of public spending weakens (i.e., public spending is less able to reduce unemployment when technological change is more advanced).
• Diagnostics: The panel exhibits slope heterogeneity and cross-sectional dependence, motivating the use of second-generation/dynamic panel methods.
• Policy context discussed: literature on fiscal policy, job-policies (subsidies, UBI), and redistribution as responses to technology-driven labor-market disruption.

Data & Methods

• Sample: 38 countries (2000–2021). Countries include a mix of advanced and emerging economies (e.g., Belgium, Brazil, Canada, France, Germany, Korea, Mexico, South Africa, Turkey, US, UK, etc.).
• Key variables:
  • Dependent: unemployment rate (Unemp).
  • Main regressor: public expenditures as % of GDP (Govexp).
  • Threshold variable (proxy for technological transformation): R&D spending as % of GDP (RD).
  • Controls: trade openness (trade), GDP growth (growth), inflation (inf), inequality (P90/P10), crisis dummy for 2008–2009.
• Estimation approach:
  • Dynamic panel threshold model (Kremer et al., 2013) to allow regime-dependent effects and to include lagged dependent variable.
  • First-differences/forward orthogonal deviations used to address unit roots and eliminate fixed effects (Arellano & Bover).
  • Tests reported for slope heterogeneity (Pesaran & Yamagata) and cross-sectional dependence (Pesaran CD), motivating second-generation estimators.
• Key estimated threshold: RD = 1.02% of GDP. Below/above this threshold, the effect of govexp on unemployment differs.

Implications for AI Economics

• Fiscal policy effectiveness is endogenous to technological context: as automation/AI adoption rises (proxied here by R&D intensity), traditional public spending becomes less potent at lowering unemployment. This implies that conventional fiscal stimulus or broad public expenditure increases may have smaller employment multipliers in advanced-technology regimes.
• Composition matters: the finding points to the need for reorienting public spending toward measures that are complementary to labor in an AI/automation era — e.g., human-capital investments (training, reskilling), job subsidies targeted at tasks less automatable, support for job creation in non-automatable activities, and redistribution to address displacement.
• Threshold thinking: policy prescriptions may need to be country-specific depending on where an economy sits relative to technology/adoption thresholds (here proxied by R&D/GDP). High–R&D economies may require different fiscal mixes than low–R&D economies.
• Research agenda suggestions for AI economists:
  • Disaggregate public spending by functional category (education, active labor market programs, capital-intensive infrastructure) to identify which spending types retain employment effectiveness under high automation.
  • Use direct measures of automation/AI adoption (task measures, industrial robot density, AI patenting, firm-level AI use) rather than R&D/GDP to refine thresholds.
  • Explore heterogeneous impacts across skill groups and occupations (since AI/automation is task-biased) and analyze distributional consequences.
  • Causal identification: employ quasi-experimental designs or instrumental variables to address potential remaining endogeneity between technology adoption, public spending, and labor outcomes.
  • Microdata and firm-level studies to trace how fiscal incentives translate into task allocation (human vs. machine) and hiring decisions in the presence of AI.

Reference (paper summarized) Beşirli, Ö. B., & Seçilmiş, İ. E. (2025). The effect of public expenditures on unemployment: Does technology matter? Journal of Mehmet Akif Ersoy University Economics and Administrative Sciences Faculty, 12(4), 1564–1577. https://doi.org/10.30798/makuiibf.1700034

If you’d like, I can (a) extract and summarize the paper’s tables/coefficients in more detail, (b) draft policy recommendations tailored to high-R&D vs low-R&D countries, or (c) propose an empirical extension using direct AI/robot adoption measures.