Countries with higher measured human capital move out of low-productivity states much faster—median exit times 28 versus 79 years—so AI is likely to widen or narrow the 'intelligence divide' depending on whether it reduces learning and implementation costs rather than only augmenting existing skills.

Main Finding

Measured human capital explains only a modest share of cross-country productivity levels but strongly marks different regimes of productivity mobility. Using distribution-dynamics (finite-state Markov) methods on 1,097 five-year country transitions, the authors show that capital intensity and measured human capital move more readily than total factor productivity (TFP). Small differences in five-year mobility compound into decades of expected persistence in low-productivity states: countries with higher measured human capital are much likelier to exit the bottom of the productivity distribution. The paper reframes the “intelligence divide” as a conversion problem—AI will promote catch-up only if it raises the capacity to convert cheaper/codified knowledge into productivity mobility (i.e., raises the adoption/implementation rate μ), not merely if it is widely available.

Key Points

• Two complementary perspectives reconciled:
  • Development accounting (levels): measured schooling is a limited direct explanation of productivity gaps.
  • Diffusion/absorptive-capacity (mobility): human capital increases the rate at which frontier knowledge is adopted and implemented.
• Empirical asymmetry: capital intensity and measured human capital are more mobile than relative TFP. Technologies can diffuse without producing productivity convergence.
• Quantitative mobility results (five-year transitions / implied mean exit times from lowest state):
  • Capital intensity: mean exit time ≈ 16.1 years.
  • Measured human capital: mean exit time ≈ 31.5 years.
  • TFP: mean exit time ≈ 44.0 years.
  • Fixed within-year median split on human capital: exit times ≈ 79.1 years (below median) vs 28.0 years (above median).
  • Data-selected cutoff (0.678 of U.S. human-capital level): 61.9 years (below) vs 25.4 years (above).
• The result is robust to alternative state definitions, frontiers, weighting, continuous specifications, residualized human-capital terms, and out-of-sample prediction tests.
• Interpretation: measured human capital is an informative marker of broader capability (including management, institutions, infrastructure, finance) that converts inputs and knowledge into sustained productivity gains. It is not presented as a proven causal mechanism.

Data & Methods

• Sample: 1,097 five-year country-period transitions (postwar sample; details in paper).
• Key empirical tools:
  • Caselli–Coleman-style development accounting to decompose output per worker into capital, measured human capital, and TFP (levels exercise).
  • Finite-state Markov transition (distribution-dynamics) framework in the Quah tradition to estimate transition matrices for relative productivity, capital intensity, and measured human capital.
  • Fixed-rank (1975 quartiles), within-year median splits, and data-selected thresholds to classify states; smooth/continuous human-capital terms tested too.
  • Statistical tests: block-permutation tests to compare transition matrices; out-of-sample held-out prediction for robustness.
  • Supplementary regressions of continuous five-year productivity growth on human capital with flexible controls (initial productivity, income, capital intensity) and alternative productivity measures.
• Key modeling idea: simple reduced-form dynamic equation Δa_t = μ(c_t, F_t)(1 − a_t) − ξ_t, where a_t is relative productivity, c_t represents productive capability (including measured human capital), F_t is cost of implementation, μ(·) is adoption/implementation rate; ∂μ/∂c > 0 is central.
• Design: descriptive and reduced-form; authors emphasize these are not causal estimates of schooling’s effect on growth.

Implications for AI Economics

• The intelligence divide: the critical margin is the capacity to convert accessible/cheap AI-based intelligence into productivity mobility (raising μ), not merely access/adoption. AI can:
  • Be frontier-expanding (raise global frontier) without promoting catch-up.
  • Augment skilled workers and thus raise returns where complementary capabilities already exist (risk of divergence).
  • Reduce costs of diagnosis, adaptation, validation, and implementation—only this diffusion-enhancing channel is likely to promote convergence.
• Policy and investment priorities if convergence is desired:
  • Build absorptive capacity, not just access: improve education quality (skills relevant to implementation), management, firm capabilities, data infrastructure, maintenance/repair networks, finance, and public administration.
  • Support AI tools designed to lower adaptation/validation costs in low-capability contexts (e.g., automated diagnostics, low-code/no-code adaptation, localized model fine-tuning, robust small-data methods).
  • Promote complementary institutions: regulatory frameworks, standards, and platforms that help translate AI outputs into reliable production processes.
  • Target interventions where AI fits a concrete bottleneck (analogous to mobile-phone impact on fishermen), rather than seeking to transplant frontier practices unchanged.
• Measurement and evaluation:
  • Monitor productivity mobility (transition probabilities, exit times) in addition to adoption and diffusion metrics.
  • Disaggregate sectoral and firm-level responses: diffusion effects may be highly heterogeneous across sectors and firm types.
• Research directions:
  • Micro and causal evidence on whether AI lowers implementation/learning costs in low-capability settings (RCTs, firm-level adoption studies).
  • Sectoral studies on which AI applications are most diffusion-enhancing versus frontier-augmenting.
  • Investigation of how specific complementarities (management practices, finance, infrastructure) interact with AI to change μ.
• Cautionary note: empirical results are descriptive and reduced-form. The association between measured schooling and mobility is robust and informative, but not a proven causal pathway; measured human capital proxies for a broader capability bundle.

Summary takeaway: AI’s distributional impact hinges on whether it primarily augments existing skill-rich environments or lowers the costs of learning and implementation for economies with weaker capabilities. Closing the intelligence divide requires policy and investment that raise the conversion capacity to turn cheaper intelligence into sustained productivity mobility.