Countries that boost AI readiness tend to see measurable gains in e-government: within-country increases in the Government AI Readiness Index are linked to higher E‑Government Development, and the post‑pandemic period shows a notable acceleration in digital public administration—though national trajectories differ sharply.

Main Finding

Within-country improvements in AI capability — especially government AI readiness — are positively and robustly associated with higher levels of e-government development. Panel estimates show a significant within-country effect (FE) of the Government AI Readiness Index (GAIRI) on the UN E‑Government Development Index (EGDI) (FE coefficient = 0.17, p < 0.001). Cross-sectional (RE) estimates produce larger between-country correlations (AIVS: 2.55, p < 0.001; GAIRI: 0.35, p < 0.001), but correlated‑RE (Mundlak) specifications indicate the between‑country component is statistically insignificant — implying dynamic, policy‑driven progress matters more than static structural advantages. Time effects show notable post‑pandemic increases in EGDI (e.g., +7.02 in 2020 and +8.10 in 2022 versus baseline).

Key Points

• Hypotheses tested:
  • H1 (Within): Supported — within‑country increases in AI readiness associate with EGDI gains.
  • H2 (Between): Not supported in correlated‑RE — persistent cross‑country AI advantages do not statistically explain EGDI once country heterogeneity is accounted for.
  • H3 (Robustness): Supported — results hold under corrections for heteroskedasticity, serial correlation, and cross‑sectional dependence.
• Indicators used:
  • AI Vibrancy Score (AIVS; Stanford Global AI Vibrancy Index) and Government AI Readiness Index (GAIRI; Oxford Insights).
  • E-Government Development Index (EGDI; UNDESA).
• Heterogeneity:
  • Large country‑specific effects; examples of positive deviations: Denmark, Estonia, Korea. Examples of persistent negative deviations: India, South Africa.
• Timing:
  • Marked acceleration of digital public administration during the post‑pandemic period (significant positive year effects).
• Robustness:
  • Models checked with FE, RE, Mundlak (correlated RE); diagnostics (Hausman, Breusch‑Pagan, Breusch‑Godfrey/Wooldridge, Pesaran CD); standard errors clustered and Driscoll–Kraay applied.

Data & Methods

• Two complementary panel datasets:
  • AIVS → EGDI: 36 countries, three biennial points (2018, 2020, 2022) → 108 observations.
  • GAIRI → EGDI: 170 countries, three points (2020, 2022, 2024) → 510 observations.
• Variables:
  • Dependent: EGDI (UN E‑Government Knowledgebase).
  • Independents: AIVS (Stanford Global AI Vibrancy Tool) and GAIRI (Oxford Insights Government AI Readiness Index).
• Preprocessing:
  • Addressed non‑normality: Box–Cox transformation on AIVS; Yeo–Johnson on EGDI.
• Econometric strategy:
  • Estimated fixed effects (FE), random effects (RE), and Mundlak correlated random‑effects to decompose within vs between effects.
  • Country and time effects included to capture unobserved heterogeneity and time shocks.
  • Diagnostics: Hausman test to choose FE vs RE; Breusch‑Pagan LM (panel effects), Breusch‑Godfrey/Wooldridge (serial correlation), Pesaran CD (cross‑sectional dependence), Breusch‑Pagan (heteroskedasticity).
  • Robust inference: cluster‑robust and Driscoll–Kraay standard errors.
• Main numerical results (highlights):
  • FE estimate (GAIRI → EGDI): 0.17 (p < 0.001).
  • RE estimates: AIVS → EGDI = 2.55 (p < 0.001); GAIRI → EGDI = 0.35 (p < 0.001).
  • Mundlak: between‑country AI components become statistically insignificant; within‑country effects remain significant.

Implications for AI Economics

• Policy efficacy and returns:
  • The finding that within‑country, time‑varying improvements in AI readiness matter for e‑government implies that policy interventions (training, governance, data platforms) can yield measurable public‑sector digital returns within short/medium horizons. This supports analyses of public investment in AI as productive capital in the public sector.
• Complementarities and sequencing:
  • Results underscore complementarities (human capital, governance, data infrastructure). Economic models of AI adoption should incorporate complementarities between technical AI assets and institutional capacity to capture realized returns.
• Distributional and development concerns:
  • Lack of a robust between‑country AI advantage once correlated factors are controlled suggests that structural endowments alone do not lock countries into better e‑government outcomes; targeted reforms can change trajectories. However, observed country heterogeneity warns that benefits may be unequally distributed, implying the need to model distributional impacts and transition costs in AI diffusion scenarios.
• Measurement and policy evaluation:
  • Distinction between AI vibrancy (research/market vibrancy) and government readiness (policy, governance, infrastructure) matters for predicting public‑sector outcomes. Empirical economic work should treat readiness indices as separate inputs when estimating social returns to AI.
• Research priorities for AI economics:
  • Move beyond correlations toward causal identification (e.g., policy experiments, instrumental variables, synthetic controls) to estimate marginal social returns to government AI investments.
  • Quantify sectoral and labor market spillovers from public‑sector AI adoption (productivity, employment composition, service quality).
  • Incorporate dynamic, heterogeneous treatment effects (which countries/sectors benefit fastest; which face risks) into growth and distributional models.
• Cautionary notes:
  • Ethical, accountability, and governance capacities are integral to realizing benefits; economic models should include regulatory and institutional frictions as constraints on diffusion and welfare gains.
  • Short sample horizons and indicator limitations mean practitioners and modelers should interpret magnitudes cautiously and continue to incorporate updated indices as they evolve.

Limitations noted by authors (relevant for researchers): relatively few time points (biennial EGDI), potential measurement limitations of indices, and residual endogeneity concerns — all of which motivate deeper causal and micro‑level followups.