Chinese manufacturing firms have steadily increased the coordination between digital transformation and green efficiency since 2014, though progress is uneven across industries and regions; government R&D subsidies, tougher environmental regulation, and higher human capital are linked to stronger digital–green coupling.

Main Finding

The paper quantifies firm-level synergy between corporate digitalization (including AI, big data, cloud, blockchain, and digital applications) and green efficiency for Chinese A‑share manufacturing firms (2014–2023). Using text‑based digitalization measures and a super‑efficiency SBM DEA for green performance, it finds that digitalization, green efficiency, and their coupling coordination all increased substantially over the sample. The mean digitalization index rose from 0.142 → 0.387 (annual growth ~10.5%), green efficiency from 0.524 → 0.681, and the coupling coordination degree from 0.412 → 0.586 (moving the sample from “verge of disorder” toward “barely/primary coordinated”). Key positive drivers of higher coordination are government R&D subsidy intensity, environmental regulation intensity, and firm human capital.

Key Points

• Sample: 2,683 listed manufacturing firms, 21,328 firm‑year observations (2014–2023).
• Digitalization measure:
  • Constructed from MD&A annual report text using an 87‑keyword dictionary across five dimensions: AI, big data, cloud computing, blockchain, and digital applications.
  • Base indicator = ln(1 + total keyword frequency) per firm-year; combined via entropy weights into a composite index U1 ∈ [0,1].
• Green efficiency measure:
  • Super‑efficiency SBM DEA model with undesirable outputs.
  • Inputs: labor (employees), capital (net fixed assets), energy (converted to standard coal).
  • Desirable output: operating revenue; Undesirable: industrial SO2 and CO2 emissions.
  • Composite index U2 ∈ [0,1] (ρ capped at 1).
• Coupling coordination:
  • Coupling degree C = 2√(U1 U2)/(U1 + U2)^2; composite index T = 0.5·U1 + 0.5·U2; coordination D = √(C × T). D is graded into 10 coordination levels.
• Drivers:
  • Panel Tobit with firm & year fixed effects (D ∈ [0,1] bounded dependent variable). Core regressors: R&D subsidy intensity (RDS), environmental regulation intensity (ENV), human capital share (HC). Controls: firm size, age, leverage, ownership.
  • RDS, ENV, HC show significant positive effects on D.
• Heterogeneity:
  • Industry: high‑tech manufacturing (computer/communications/electronics) shows much higher digitalization and higher coordination; heavy industries lag in digitalization though green efficiency gaps are smaller.
  • Region: eastern region leads central and western regions; regional gaps narrowed but persisted.
• Implementation details:
  • Data sources: CSMAR, CNRDS, Wind, annual report texts.
  • Preprocessing: PDF scraping, MD&A extraction, jieba segmentation, winsorize (1%/99%), deflation to 2014 prices, range standardization, entropy weighting.
  • Tools: Python (text processing), Stata 17, MaxDEA Ultra (SBM). 80/20 train/validation split used for reproducibility.

Data & Methods

• Data:
  • Firm‑level financials, patents, annual report MD&A text, and pollutant emissions for Chinese listed manufacturing firms.
  • Exclusions: ST firms, financial/insurance firms, missing key variables.
• Digitalization index construction:
  • 87 keywords → 5 subdimensions.
  • Frequency counts per MD&A (negation and third‑party mentions excluded).
  • Base score = ln(1 + sum of frequencies across dimensions); nondimensionalized and entropy‑weighted into U1.
• Green efficiency:
  • Super‑efficiency SBM DEA with undesirable outputs (allows ranking of efficient units and accounts for emissions).
  • Inputs and outputs standardized and used to compute ρ; U2 = min(ρ,1).
• Coupling coordination modeling:
  • Standard physics‑inspired coupling coordination formulas (C, T, D) with equal weights for subsystems.
  • D graded into 10 categories from extreme disorder to superior coordination.
• Econometrics:
  • Panel Tobit to account for bounded D, with firm and year fixed effects, clustered standard errors.
  • Main explanatory variables: RDS (subsidy/operating revenue), ENV (regional environmental governance investment/GDP), HC (share of employees with bachelor+).
  • Controls: ln(total assets), firm age, leverage, ownership attributes.

Implications for AI Economics

• Measurement: The paper demonstrates a scalable approach to quantify firm‑level digital/AI adoption using annual report MD&A text (keyword dictionary + ln(1+freq) + entropy weighting). This is a practical proxy for research that needs firm‑level AI/digitalization indicators when direct usage/installation data are unavailable.
• Policy levers that reinforce AI→green synergies:
  • R&D subsidies (public funding) and environmental regulation both strengthen coordination between digitalization and green efficiency—suggesting complementary roles of innovation funding and well‑designed regulation to realize environmentally beneficial impacts of AI/digital investments.
  • Human capital matters: investment in skilled labor amplifies the coordination effect, indicating complementarities between AI adoption and workforce skills.
• Targeting and heterogeneity:
  • Heavy industries lag in digitalization despite similar green efficiency baselines; targeted policy (digital adoption incentives, upskilling, data infrastructure) for these sectors could accelerate green transitions using digital/AI tech.
  • Regional disparities imply place‑based policies: eastern regions lead, so central/western regions may need catch‑up support (data infrastructure, subsidies, training).
• Methodological contribution for AI economics:
  • Integrating text‑based AI adoption proxies with operational research measures (DEA with undesirable outputs) and system‑level coupling metrics provides a replicable framework to study AI’s environmental externalities and co‑evolution with firm performance.
• Cautions for interpretation and future research directions:
  • Keyword frequency is an adoption/proximity proxy—not a direct measure of AI deployment intensity or effectiveness; future work could combine procurement, patent, or investment data for validation.
  • Endogeneity and causal identification: the paper documents associations and temporal patterns, but causal pathways (e.g., digitalization → green efficiency vs. reverse or common drivers) deserve instruments or quasi‑experimental designs.
  • Emissions and efficiency measures depend on data quality and conversion assumptions (e.g., CO2 estimation); robustness checks and alternative specifications would strengthen causal claims.

Overall, the paper provides a practical, data‑driven pipeline for measuring firm‑level digital/AI adoption and linking it quantitatively to green performance, with actionable policy implications for using subsidies, regulation, and human capital development to amplify positive AI→green outcomes in manufacturing.