China’s AI pilot zones raise firms’ critical digital patenting, particularly outside manufacturing and high‑tech hubs; the policy effect is strongest in fiercely competitive industries.

Main Finding

AI pilot zone policies (National New‑Generation AI Innovation Development Zones) causally and robustly increase firms’ critical digital technology innovation. Using Chinese A‑share listed firms (2014–2024) and a staggered difference‑in‑differences design, the paper finds a positive and statistically significant treatment effect (baseline coefficient ≈ 0.108 on the firm‑year critical digital patent count), roughly a 12% increase relative to the sample mean KDTI (mean ≈ 0.87). The result passes parallel‑trend tests and multiple placebo checks.

Key Points

• Policy intervention: National New‑Generation AI Innovation Development Zones rolled out in three batches (2019, 2020, 2021); treated firms are those located in approved pilot cities.
• Main outcome (KDTI): annual count of firm‑level critical digital technology patents across seven categories (AI, high‑end chips, quantum information, IoT, blockchain, industrial internet, metaverse) from CNRDS.
• Baseline effect: AIPZ → +0.108 KDTI (firm‑year), significant at 1% with firm and year fixed effects and controls.
• Robustness: event‑study (parallel trends) shows no pre‑trend; three placebo tests (individual, unconstrained mixed, constrained mixed) show the baseline estimate is not a spurious result.
• Heterogeneity (effects larger and statistically different at ~10% level):
  • Industry competition: stronger effect in high‑competition industries (coef ≈ 0.120) than low‑competition (≈ 0.067).
  • Industry type: stronger for non‑manufacturing firms (≈ 0.166) than manufacturing (≈ 0.059).
  • Technology intensity: stronger for non‑high‑tech firms (≈ 0.143) than high‑tech firms (≈ 0.071).
• Proposed transmission channels: regional resource agglomeration (shared platforms, talent programs, fiscal/tax/R&D subsidies), upgraded innovation infrastructure and institutional support, and enhanced industry‑university‑research collaboration.
• Sample restrictions and processing: excludes ST/*ST/PT and financial firms; winsorizes continuous vars at 1%/99%; uses city‑ and firm‑level controls.

Data & Methods

• Sample: Balanced panel of A‑share listed firms on Shanghai and Shenzhen exchanges, 2014–2024 (≈ 29,600 firm‑year observations after cleaning).
• Treatment assignment: firm treated if its prefectural city is approved as an AI pilot in the year of approval or later (three approval batches).
• Dependent variable: firm annual count of critical digital patents (CNRDS Critical Digital Technology Patent Database), matched at firm and subsidiary level.
• Econometric strategy:
  • Staggered difference‑in‑differences (DID) with firm fixed effects (λi) and year fixed effects (μt).
  • Controls: firm‑level (age, log assets, leverage, ROA, top3 shareholding, CEO‑chair duality, board size) and city‑level (SciTech spending share, government intervention, fiscal pressure).
  • Dynamic event‑study specification to test parallel trends.
  • Placebo tests: (1) randomly assign treated firms, (2) unconstrained mixed placebo, (3) constrained mixed placebo (preserve batch structure but randomize assignments/time).
• Heterogeneity tests: split samples by industry HHI (competition), CSRC manufacturing vs non‑manufacturing, and high‑tech vs non‑high‑tech (based on National Bureau of Statistics classification).
• Additional data: city characteristics from statistical yearbooks; non‑listed firm data (Qichacha) were collected for supply‑chain spillover analysis (not detailed in the excerpt).

Implications for AI Economics

• Micro evidence that AI pilot zone policies can accelerate firm‑level development of foundational digital technologies (patentable inventions) — supporting the view that area‑based AI policy can produce measurable R&D returns at the firm level.
• Heterogeneous effects highlight policy complementarities and limits:
  • Policies seem particularly effective where firms face weaker internal capabilities or tighter market pressures (non‑high‑tech, non‑manufacturing, high‑competition industries). Thus pilots can help narrow capacity gaps and stimulate catch‑up innovation.
  • Smaller incremental gains for established high‑tech and manufacturing firms suggest diminishing marginal returns of public pilot support where baseline R&D is already high; different instruments (e.g., financing support, capital‑intensive infrastructure) may be needed there.
• Policy design recommendations:
  • Combine pilot zones with targeted financing instruments (to address manufacturing firms’ financing frictions) and with measures that encourage uptake among high‑tech incumbents (to raise marginality).
  • Maintain and expand industry‑university‑research platforms and shared computing/talent resources, since agglomeration and institutional upgrades are plausible channels.
  • Leverage competition dynamics: in more competitive sectors, policy subsidies and shared infrastructure can spur faster private R&D response.
• For researchers: using patent counts in narrowly defined “critical digital” categories is a useful micro‑level metric; future work should track long‑run commercialization, productivity effects, and spillovers to non‑listed firms and regions. Limitations to note include reliance on listed firms and patent‑based measures (which capture only one dimension of innovation).