Companies that orient toward AI report lower carbon intensity, and their peers and suppliers show downstream emissions reductions too; the link appears routed through higher-quality green innovation though causality is not fully established.

Main Finding

Firms with a stronger artificial intelligence orientation (AIO) exhibit lower carbon emission intensity, and these decarbonization gains spill outward to both industry peers (horizontal spillovers) and upstream suppliers (vertical spillovers). The effect operates partly through improvements in green innovation quality. The relationship is robust but heterogeneous: the magnitude and allocation of benefits depend on climate risk, industry competition, and AI task exposure.

Key Points

• Definition: AIO = a firm’s strategic commitment to embedding AI into decision-making and long-term development (measured as AI-related language density in MD&A disclosures, filtered for contextual relevance).
• Focal-firm effect: Higher AIO is significantly associated with lower firm-level Scope 1 + Scope 2 emission intensity. Baseline pooled coefficients were large (e.g., −0.4278 / −0.4161) and remain negative and significant with firm and year fixed effects (−0.0397 in the fully specified model).
• Spillovers:
  • Horizontal: Industry peers’ carbon intensity falls when other firms in the industry have stronger AIO (demonstration and competitive pressure channels).
  • Vertical: Upstream suppliers’ emissions intensity also declines when their buyers have stronger AIO (supply-chain transparency, digital coordination and capability diffusion).
• Mechanism: AIO raises the quality and deployability of green innovation (green patenting/innovation quality measures), which makes low-carbon solutions more codifiable, scalable and likely to diffuse across networks.
• Heterogeneity: The strength and distribution of decarbonization benefits vary by contextual factors (firm exposure to climate risk, degree of industry competition, and the types of tasks susceptible to AI). Under some conditions, benefits are retained within the adopter; under others they diffuse more widely.
• Contribution: Moves beyond technology-adoption counts (hardware/robot metrics) to a strategic orientation measure and documents system-level environmental returns to strategic AI commitment.

Data & Methods

• Sample: Chinese A-share listed firms, 2010–2023; final sample 3,046 firms and 15,353 firm-year observations. Exclusions: financial firms, ST/*ST/PT listings, firms with missing data. Data winsorized at 1st/99th percentiles.
• AIO measurement:
  • Text source: MD&A sections of annual reports (scraped from CNINFO and firm websites).
  • Procedure: seed-term retrieval of AI-related sentences, then context-aware semantic filtering using Qwen2.5-9B to retain only sentences indicating firm-level AI strategy/implementation. AIO = frequency of validated AI terms / MD&A word count, transformed as ln(1 + AIO).
• Dependent variable: Firm-level carbon emission intensity based on Scope 1 + Scope 2 emissions standardized by economic output; log-transformed (ln) for normalization.
• Mechanism variable: Green innovation quality proxied from patent data (CNRDS)—paper links AIO to higher-quality green patents/innovation outcomes.
• Controls: Firm size, leverage, ROA, Tobin’s Q, firm age, tangible asset ratio, capital intensity, board size, CEO–chair duality, Top10 shareholding, plus year and firm fixed effects.
• Spillover identification: Horizontal effects evaluated at the industry-peer level; vertical effects evaluated for upstream suppliers (supply-chain relationships and supplier emissions inferred using available supply-chain linkage data and firm disclosures—primary non-financial variables from CSMAR and carbon disclosures).
• Estimation: Panel regressions with firm and year fixed effects; robustness checks reported (including pooled/specification comparisons shown by varying inclusion of controls and fixed effects).

Implications for AI Economics

• Policy and regulation:
  • Encouraging strategic AI adoption (not just hardware purchases) can yield system-level decarbonization beyond adopting firms. Policies that incentivize strategic AI integration and promote digital supply-chain platforms could magnify economy-wide emissions gains.
  • Regulators should account for spillovers when evaluating firm-level AI subsidies or carbon policies—benefits may accrue to non-adopters, altering cost–benefit assessments.
• Corporate strategy:
  • Firms seeking sustainability gains should embed AI as an organizational orientation (strategy, processes, governance) rather than as isolated tools; this increases the likelihood that AI-driven innovations are high-quality, scalable, and adopted across networks.
  • Buyer-driven digital coordination (AI-enabled procurement/monitoring) is a viable pathway for reducing supplier emissions.
• Research and measurement:
  • Using contextualized text measures (LLMs for semantic filtering) offers a promising way to capture strategic orientation toward AI—this complements physical adoption metrics used in prior studies.
  • Future causal work should pursue stronger identification (e.g., exogenous shocks to AI orientation, instrumental variables, or quasi-experiments) to confirm causality and quantify welfare implications.
• Caveats and open questions:
  • Energy footprint and rebound risks of AI remain relevant—this paper shows net reductions in intensity but does not rule out absolute increases in energy use under some scenarios.
  • Generalizability beyond China, and finer-grained mapping of supplier-level transmission mechanisms (e.g., which supplier types benefit most), are important next steps.
• Economic significance:
  • Even after firm fixed effects, the negative association persists, suggesting that strategic AI orientation has measurable, persistent links to carbon efficiency—worth considering in models of technology diffusion, environmental externalities, and industrial policy.