Chinese listed firms that adopt AI exhibit higher innovation efficiency, driven by better product‑level productivity, data use and ESG scores; the uplift is largest in central and western state‑owned, asset‑ and labor‑intensive firms.

Main Finding

AI application by Chinese A-share firms (Shanghai & Shenzhen, 2012–2023) significantly improves enterprise innovation efficiency. This effect operates mainly via three mediating channels—new-quality productivity, data-factor utilization, and ESG performance—and is moderated (weakened or strengthened) by market segmentation, fiscal support, and human-capital structure. The effect is stronger for central & western firms, state-owned enterprises, and asset- or labor-intensive firms. Results are robust to alternative innovation measures, lagged AI, and winsorization.

Key Points

• Primary result: Firm-level AI application → higher innovation efficiency (measured as patents per unit R&D). Coefficients positive and statistically significant across multiple specifications.
• Mediators (all found significant):
  • New-quality productive forces (NQPF): AI → higher NQPF → higher innovation efficiency.
  • Data factor utilization (Data): AI → greater use of data-related factors → higher innovation efficiency.
  • ESG performance (ESG): AI → better ESG metrics → higher innovation efficiency.
• Moderators:
  • Market segmentation (Seg): interaction with AI is negative → greater market segmentation weakens AI’s positive effect.
  • Fiscal support (GOV, government subsidies): positive interaction → subsidies amplify AI’s benefit for innovation efficiency.
  • Human capital structure (Human, share with bachelor+): positive interaction → better human capital amplifies AI’s benefit.
• Heterogeneity: Effects are stronger for firms that are state-owned, located in central/western China, and that are asset-intensive or labor-intensive.
• Robustness: Findings hold when using a weighted-patent innovation measure (InnoEff2), lagged AI, and 2% winsorization.

Data & Methods

• Sample: Shanghai & Shenzhen A‑share listed firms, 2012–2023. Exclusions: ST/*ST/PT firms, financial sector, firms with extensive missing data.
• Main sample sizes reported: up to ~25,701 observations for some variables; benchmark regressions use ~13,684–22,297 firm-year observations depending on controls.
• Dependent variable:
  • InnoEff1 = Patent1 / ln(1 + R&D expenditure), where Patent1 = ln(1 + total patent applications (invention + utility model + design)).
  • Robustness check: InnoEff2 uses weighted patent counts (3:2:1) in numerator.
• Key independent variable:
  • AI: firm-level AI application intensity measured via textual analysis of annual reports using a machine-learning-built AI dictionary (following Li Y. et al. methodology).
• Mediators:
  • NQPF: index of new-quality productive forces (labor, objects, means) via entropy method.
  • Data: frequency count of data-related keywords in annual reports (negatives excluded).
  • ESG: firm ESG score from Shanghai Huazheng Index Information Service.
• Moderators:
  • Seg: province-level market segmentation index (deviation from provincial marketization indices).
  • GOV: government subsidies (fiscal support intensity).
  • Human: share of employees with bachelor’s degree or above.
• Controls: firm size, ROA, leverage, liquidity, age, board size, executive pay, Top5 ownership, PE; macro controls: GDP, industry structure, regional financial development.
• Econometric approach:
  • Two-way fixed-effects panel regressions (firm and year fixed effects).
  • Mediation analysis via stepwise regressions (AI→mediator, mediator→InnoEff controlling for AI).
  • Moderation assessed via interaction terms (AI × moderator).
  • Robustness: alternative dependent variable, one-period lagged AI, winsorization (2%).
• Diagnostics: correlation analysis shows AI positively correlated with innovation efficiency; VIFs reported <2 (no serious multicollinearity).

Implications for AI Economics

• Multi-channel value capture: AI increases innovation efficiency not only by technological substitution but by enhancing organizational productive characteristics (new-quality productivity), better leveraging data as a production factor, and improving ESG outcomes—highlighting that AI’s economic value is realized across technical, organizational, and reputational dimensions.
• Role of public policy and institutions:
  • Fiscal support (government subsidies) complements AI investments—policy can amplify firm-level R&D productivity gains from AI.
  • Market segmentation is a barrier: fragmented markets blunt AI’s enabling effects; reducing barriers to market integration could improve aggregate returns to AI adoption.
• Human capital matters: returns to AI are larger where higher-skilled labor is available, underscoring complementarities between AI and workforce upskilling—implication for education and training policy and firm HR strategy.
• Heterogeneous adoption returns: regional, ownership, and factor-intensity heterogeneity suggests targeted policy and firm strategy—e.g., state-owned and central/western firms may have higher marginal gains from AI; asset- and labor-intensive firms can particularly benefit.
• Measurement and evaluation: the study demonstrates feasibility of measuring firm-level AI application via textual analysis, but also underscores that AI’s effects should be evaluated through multiple outcome channels (patent efficiency, data use, ESG).
• Research and policy gaps: causal identification relies on panel fixed effects and lagged variables but may still face endogeneity (reverse causality, omitted variables). Future AI economics work should seek stronger quasi-experimental designs (IVs, policy shocks, randomized pilots) to pin down causality and quantify welfare gains.

Limitations to note for interpreting results: reliance on textual-proxy measure of AI; patent-based innovation-efficiency metrics may not capture all innovation forms; potential endogeneity not fully addressed beyond lags and fixed effects. Future work could extend to broader firm populations, alternative innovation outputs (product market outcomes), and causal identification strategies.