Chinese listed firms that disclose and invest more in AI show higher green innovation efficiency, with financing strength amplifying the gains and the largest effects in state-owned, large and manufacturing firms.

Main Finding

Enterprise investment in artificial intelligence (AI) — measured both as (a) annual-report AI keyword disclosure (lnAI_freq) and (b) AI investment intensity (AI_invest, investment/total assets) — is positively associated with firm-level green innovation efficiency (two DEA-based outcomes: green R&D efficiency and green achievement-transformation efficiency). The effect is economically meaningful and robust to multiple checks; financing capacity (higher leverage) amplifies the effect, and the positive effect is stronger for state-owned firms, large firms, and manufacturing firms.

Key Points

• Sample and outcomes

  • China A‑share listed firms, 2007–2023; final sample: 26,779 firm‑year observations (financial firms and ST firms excluded; winsorized).
  • Two dependent variables (DEA scores, bounded [0,1]): green_rd (green technology R&D efficiency) and green_trans (green achievement transformation efficiency).

• Core explanatory variables

  • lnAI_freq: log(1 + AI keyword frequency in annual reports) — captures disclosure/attention to AI.
  • AI_invest: ratio of total AI investment to total assets — captures actual resource input.

• Baseline quantitative results (random-effects panel Tobit)

  • lnAI_freq coefficient ≈ 0.042 (p < 0.01) for green_rd (similar for green_trans).
    • Interpreted: a one‑unit increase in the logged AI word frequency is associated with a ≈0.042 increase in the DEA efficiency score. Relative to the sample mean (~0.54), this is roughly a 7.8% increase.
  • AI_invest coefficient ≈ 1.080 (p < 0.01) for green_rd (similar for green_trans).
    • Because AI_invest is small in level (mean ≈0.005, SD ≈0.009), a one standard-deviation increase in AI_invest implies an approximate 0.0097 absolute increase in the DEA score — roughly 1.8% of the sample mean (the paper’s abstract reports a ~4% improvement; see Data & Methods notes below on scale).

• Control variables

  • Firm size and age positively related to green innovation efficiency.
  • Leverage and high revenue growth negatively associated with the DEA measures (growth associated with prioritizing scale).
  • Board size negative; CEO-chair duality positive; top1 (ownership concentration) negative.

• Moderation and heterogeneity

  • Asset‑liability ratio (leverage) positively moderates AI’s effect on green innovation — firms with greater financing capacity can better translate AI investment into green innovation efficiency.
  • Stronger effects observed in state‑owned enterprises, larger firms, and manufacturing sector firms.

• Robustness and causal-strengthening checks

  • Authors report robustness to high-dimensional fixed-effects specifications and to a difference‑in‑differences (DID) design exploiting the National New‑Generation Artificial Intelligence Innovation and Development Pilot Zones policy shock; the DID passes a parallel‑trends test per the paper.

Data & Methods

• Dependent variables

  • Two-stage green innovation efficiency measures obtained via Data Envelopment Analysis (DEA): green R&D efficiency (green_rd) and green achievement transformation efficiency (green_trans), both bounded (0,1).

• Explanatory measures

  • Text mining of firms’ annual reports to count AI-related keywords; lnAI_freq = log(1 + count).
  • Financial statement–based AI_invest = AI-related investment / total assets.

• Controls and moderators

  • Controls: Size (ln total assets), Age, Leverage (total liabilities/total assets), Growth (revenue growth), BoardSize, Dual (CEO/chair duality), Top1 (largest shareholder’s share).
  • Moderators (centered): Size, Leverage, Top1.

• Main estimation strategy

  • Random-effects panel Tobit model (accounts for censoring at 0 and 1 in DEA scores).
  • Moderation tested via interaction terms (AI × moderator).
  • Robustness: high-dimensional fixed-effects (reghdfe) with firm and year or industry and year fixed effects; clustering at firm level.
  • Policy-based DID: exploitation of pilot zone designation for new‑generation AI innovation as exogenous shock; parallel‑trend checks performed.

• Data processing

  • Data sources: CSMAR (financials/governance), Guotai’an patent database, annual report texts.
  • Exclusions: financial firms, ST firms; winsorization at 1st/99th percentiles; missing data removed → 26,779 observations.

• Note on reported magnitudes

  • The paper’s abstract reports a one‑standard‑deviation increase in AI_invest improves green innovation efficiency by ~4%; the tables show AI_invest SD ≈ 0.009 and coefficient ≈ 1.080, which implies an absolute change ≈ 0.0097 (~1.8% of the mean ≈ 0.54). This is a minor internal inconsistency in percentage interpretation; the raw coefficient and SD are reported above for transparency.

Implications for AI Economics

• AI as productive (and green‑augmenting) capital: The paper provides micro‑level evidence that firm AI investment — both signaling (disclosure) and actual resource input — raises the efficiency of converting R&D and innovations into green outputs, supporting models that treat digital/AI capital as augmenting total factor productivity with environmental benefits.
• Importance of measurement: Combining textual disclosure metrics with financial investment intensity is useful for capturing both attention/signal and actual resource deployment; future AI economics work should replicate or extend this dual-measure approach.
• Financing constraints and complementarities: The positive moderation by leverage highlights that AI’s green benefits depend on complementary financial capacity. Models of AI adoption should incorporate interactions between digital capital and firm financing/frictions.
• Heterogeneous impacts: Stronger effects in SOEs, large firms, and manufacturing suggest distributional consequences of AI-driven green gains — policy and market dynamics could widen or narrow firm‑level green productivity gaps depending on access to capital and organizational scale.
• Methodological guidance: For bounded efficiency outcomes, Tobit (or other censored models) and high‑dimensional fixed effects help address bias; policy DID designs (pilot zones) can support causal claims if parallel trends hold.
• Policy relevance: Supports targeted policies that (a) incentivize AI investment for green innovation, (b) improve SME and private‑firm access to finance to realize AI’s green potential, and (c) promote meaningful AI disclosure to signal capabilities and attract green collaboration/investment.

Limitations to note (and useful directions for future research) - Single-country (China) listed‑firm sample limits external validity; private and small unlisted firms not observed. - AI keyword frequency is an imperfect proxy for AI capability/adoption; richer measures (models used, compute, AI talent, measurable AI outputs) would strengthen causal interpretation. - Potential endogeneity: firms that are already more green‑efficient might choose to invest/disclose AI more; the DID helps but more quasi‑experimental designs or instrumental variables would further bolster causal claims. - Mechanisms: micro‑level processes (e.g., which AI uses reduce emissions/waste vs. accelerate green productization) require firm‑case or process‑level data.

If you want, I can: - Extract the exact list of AI keywords used for lnAI_freq (if provided in an appendix), or - Recompute percentage effects under alternative normalization (e.g., using median or different SD), or - Produce a one‑page slide summary suitable for policy audiences.