China’s AI pilot-zone policy lifted provincial 'new-quality productivity' between 2012 and 2022, driven largely by industry upgrading and talent clustering; gains were strongest in the eastern provinces and in regions with higher openness and marketization.

Main Finding

The New-Generation Artificial Intelligence Innovation and Development Pilot Zone policy significantly increased "new-quality productivity" in Chinese provinces. Using 2012–2022 panel data for 30 provinces and a multi-period difference-in-differences design, the authors estimate an average uplift of about 2.84% in new-quality productivity in pilot regions (robust at conventional significance levels). The effect is dynamic and accumulates over time, and operates in part through industrial-structure upgrading and talent agglomeration. Effects are heterogeneous across regions and institutional contexts (stronger in the east and in provinces with higher openness and marketization).

Key Points

• Definition and measurement:
  • The paper operationalizes "new-quality productivity" (NP) with a four-dimension index: laborers, means of labor, objects of labor, and factor-combination efficiency. Indicators are combined using the CRITIC entropy-weighting method.
  • Factor-combination efficiency includes TFP via Stochastic Frontier Analysis (SFA), investment-output ratios, and institutional efficiency (marketization index).
• Causal design:
  • Treats pilot-zone designations (2019–2021, adjusted for late approvals) as a quasi-natural experiment and applies a multi-date DID (MDID) with province and year fixed effects.
• Main quantitative result:
  • Treat×Post coefficient ≈ 0.0284 (column with controls) → ~2.84% higher NP in treated provinces.
  • Effect is statistically significant and grows over time (parallel-trends test passed; post-treatment coefficients positive and increasing).
• Mechanisms:
  • Mediation tests point to two transmission channels: (1) industrial structure upgrading (measured as tertiary-sector share) and (2) talent aggregation (location-quotient of full-time-equivalent R&D personnel).
• Robustness:
  • Parallel-trends diagnostic satisfied.
  • Placebo tests (500 random draws) produce coefficients centered at zero.
  • Additional robustness checks: alternative core measures (AI patent counts, AI firms), lagging dependent variable, and excluding overlapping contemporaneous policies — results remain supportive.
• Heterogeneity:
  • Strongest policy effects in the eastern region, next strongest in the western region; effects in central and northeast not statistically significant.
  • Provinces with higher openness and greater marketization experience larger gains from the pilot policy.

Data & Methods

• Sample:
  • Panel data for 30 Chinese provinces (2012–2022), 330 observations after aggregations/adjustments.
  • Data sources: China Statistical Yearbook, China Industrial Statistical Yearbook, provincial yearbooks, EPS database. Missing values filled by linear interpolation where needed.
• Treatment assignment:
  • Province treated in year when one or more cities within it were approved as AI pilot zones (Ministry of Science & Technology batches: 2019, 2020, 2021). Approvals in Sept+ are shifted to the following year to allow for implementation lags.
• Outcome (NP) construction:
  • Composite index combining tertiary indicators across four primary dimensions (labor, means, objects, factor-combination efficiency). Examples: average wage, higher-education share, software revenue, broadband ports per capita, robot density, patent counts, renewable energy share, forest coverage, environmental investment, TFP (SFA), marketization index.
  • Weights determined via CRITIC entropy method (objective weighting that accounts for contrast and correlation).
• Econometric specification:
  • Multi-period DID: NP_it = β0 + β1(Treat_i × Post_t) + β2 X_it + λ_i + μ_t + ε_it.
  • Mediator models and causal steps follow standard mediation regressions (M_it on Treat×Post; NP_it on Treat×Post and M_it).
  • Controls: education spending share, financial development (loans/GDP), industrialization (industry value-added/GDP), log per-capita highway length, urbanization rate.
• Inference and checks:
  • Province and year fixed effects; robust standard errors reported.
  • Parallel trends via event-study-style coefficients; placebo via random treatment reassignment (500 draws); alternative core explanatory variables and lag specifications for endogeneity checks.

Implications for AI Economics

• Policy-level AI interventions can raise macro-level productivity beyond firm-level automation gains:
  • The pilot-zone experiment provides empirical evidence that targeted AI policy packages (infrastructure, institutional supports, "AI+" integration) can produce measurable gains in composite productivity metrics that capture qualitative aspects of productive capacity.
• Importance of channels: industry upgrading and talent matter:
  • AI policies appear to work primarily by shifting industrial composition up the value chain and by concentrating R&D/talent. Evaluations and policy design should therefore explicitly target industry linkages and talent attraction/retention.
• Institutional complementarities and regional targeting:
  • Heterogeneous returns imply that local institutional conditions (marketization, openness) and regional endowments mediate policy effectiveness. Policymakers should combine AI pilot programs with reforms that improve market mechanisms, openness, and absorptive capacity to maximize returns.
• Measurement guidance for researchers:
  • The four-dimension NP index (including factor-combination efficiency and TFP via SFA) is a useful template for capturing qualitative productivity change from AI. Objective weighting (CRITIC) helps balance indicator informational content.
• Evaluation best practices:
  • Multi-period DID with event-study checks, placebo reassignments, alternative treatment proxies (AI patents, AI firms), and mediating-path tests provide a robust assessment framework for policy impacts in AI economics.

Caveats and limitations (noted by implication): - Aggregation at the provincial level may mask intra-provincial heterogeneity and city-level dynamics. - The constructed NP index, while comprehensive, depends on indicator choices and weighting; alternative constructions may yield different magnitudes. - Although multiple robustness checks were implemented, residual confounding from unobserved simultaneous policies or local shocks cannot be fully ruled out.

Policy takeaways (concise): - Scale up targeted AI pilot programs but pair them with measures to promote talent clustering and industry upgrading. - Strengthen institutions (marketization, openness) and infrastructure to amplify policy effects. - Use comprehensive indexes and multi-period causal designs to evaluate AI policy impacts and inform iterative policy design.