Big-data adoption improves manufacturers' carbon-efficiency in China by spurring green innovation and tightening internal controls; gains are largest among non-state-owned, high-tech and less-concentrated firms.

Main Finding

Big data technology application (BDTA) significantly improves carbon emission efficiency (CEE) of Chinese manufacturing firms (2010–2023). The effect operates at least partly through two micro channels: increasing firms’ green innovation and improving internal control quality. The result is robust to multiple sensitivity and endogeneity checks and is stronger for non-state-owned firms, high‑tech firms, and firms in less concentrated markets.

Key Points

• Primary result: Firm-level BDTA → higher CEE (more economic output per unit of CO2).
• Mechanisms:
  • Green innovation (measured by green patent applications): BDTA raises the quantity/efficiency of green innovation, which increases CEE.
  • Internal control quality (DIB internal control index): BDTA improves dynamic monitoring and governance, reducing information frictions and agency costs, which enhances CEE.
• Heterogeneity: Effects are larger for non-state-owned enterprises, high-tech enterprises, and enterprises operating in low market-concentration industries.
• Robustness: Authors report stability of results across robustness tests and address endogeneity concerns (details summarized in the paper).

Data & Methods

• Sample: Shanghai & Shenzhen A‑share listed manufacturing firms, 2010–2023; final sample 2,490 firms, 21,886 firm‑year observations after exclusions and 1% winsorization of continuous variables.
• Dependent variable (CEE): operating revenue divided by firm CO2. Firm CO2 is not directly observed — it is estimated by allocating industry CO2 to firms proportional to each firm’s share of industry operating costs. Industry CO2 is derived from energy consumption and standard emission coefficients (coal, coke, crude oil, gasoline, kerosene, diesel, fuel oil, natural gas).
• Key independent variable (BDTA): log(1 + frequency) of big‑data‑related keywords in the MD&A text of annual reports. Keywords constructed by the authors from technical features and application scenarios and extracted using Python (jieba).
• Mechanism variables:
  • Green innovation (GI): log(1 + number of green patent applications).
  • Internal control quality (ICQ): DIB internal control & risk management index divided by 100.
• Controls: firm size, firm age, leverage (asset–liability ratio), return on assets, operating income growth, top1 share (ownership concentration), board size, etc.
• Econometric approach: panel two‑way fixed effects models (firm and year), clustered standard errors at firm-year level. Mediation/stepwise regressions used to test channels. Authors report robustness checks and endogeneity tests (instrumental or other approaches described in full paper).

Implications for AI Economics

• Operational value of AI/big‑data at firm level: The paper provides empirical evidence that investments in data technologies (AI/IoT/big‑data stacks) can raise carbon productivity — an important negative externality channel where digital adoption reduces emissions intensity per unit of output.
• Mechanisms relevant for economic modeling:
  • Innovation channel: BDTA accelerates green R&D and lowers innovation costs — models of productivity should incorporate how data-driven capabilities shift the direction and productivity of R&D toward lower‑carbon technologies.
  • Governance channel: Improved monitoring and decision systems reduce agency costs and reallocate resources toward cleaner capital — models linking corporate governance, technology adoption, and environmental outcomes are warranted.
• Policy design:
  • Targeted promotion of BDTA (subsidies, tax incentives, public data infrastructure) could be an effective complement to direct carbon policies, especially for non-SOEs and high‑tech firms.
  • Supporting firms’ capability to translate data into green patents and stronger internal controls may increase the emissions‑efficiency payoffs of digitalization.
  • Caution: heterogeneity implies one-size-fits-all digitalization policies may be suboptimal; design should consider ownership, technological capacity, and market structure.
• Measurement and research practice:
  • The study demonstrates practical use of NLP/text‑based proxies for firm digital adoption — useful for empirical AI economics when direct adoption metrics are unavailable.
  • But reliance on disclosure‑based proxies and industry‑allocated emissions highlights measurement limits: future work should seek firm‑level emissions data (sensors, regulatory disclosures) and richer measures of AI/ML intensity.
• Directions for future research in AI economics:
  • Causal identification (e.g., natural experiments, policy shocks) to better quantify causal impacts of BDTA on emissions.
  • Disaggregating BDTA into specific AI/ML methods or applications (predictive maintenance, process control, energy management) to estimate heterogeneous technology effects.
  • General equilibrium and labor‑market implications: how BDTA-driven decarbonization interacts with employment, capital reallocation, and international competitiveness in manufacturing.

Limitations to note (from the paper): firm CO2 is indirectly estimated from industry totals and operating‑cost shares; BDTA is proxied by disclosure frequency which may reflect reporting behavior as well as real adoption. These caveats suggest caution in extrapolating magnitudes but not the qualitative mechanisms.