AI adoption raises CO2 emissions on average, but good institutions and strong digital networks blunt the trade-off; the emissions impact is largest in energy-inefficient and AI-advanced countries.

Main Finding

Using a panel of 104 countries (2000–2023) and dynamic/instrumental estimators, the paper finds that AI adoption is, on average, associated with higher CO2 emissions. However, this positive (emissions‑increasing) effect is substantially mitigated — and can be reversed — in countries with high governance quality (GQI) and advanced digital infrastructure (DII). The environmental impact of AI is context‑dependent: it is weaker in energy‑efficient and early‑AI‑diffusion settings and stronger in energy‑inefficient and AI‑advanced contexts.

Key Points

• Aggregate effect: AI adoption currently correlates with higher national CO2 emissions, consistent with economy‑level rebound and infrastructure energy costs outweighing micro‑level efficiency gains in many contexts.
• Moderation by governance: Strong institutional quality (rule of law, government effectiveness, corruption control, etc.) significantly reduces the emissions intensity of AI; weak governance exacerbates AI’s emissions impact.
• Moderation by digital infrastructure: Robust digital ecosystems amplify AI’s potential to reduce emissions (through better data flows, scalability of green solutions, innovation spillovers), whereas underdeveloped DI can leave AI as a net emissions contributor.
• Heterogeneity:
  • Energy‑efficient countries show a weaker (less positive) AI→CO2 effect.
  • Energy‑inefficient countries show a stronger positive effect.
  • Low‑AI (early diffusion) countries exhibit weaker environmental impacts from AI than high‑AI (advanced diffusion) countries, where aggregate energy demands of AI deployment become more pronounced.
• Theoretical framing: integrates endogenous growth theory (innovation and rebound effects), institutional theory (governance shaping technological outcomes), and technological ecosystem theory (role of digital infrastructure).

Data & Methods

• Sample: Balanced panel of 104 countries spanning 2000–2023.
• Main variables: CO2 emissions (dependent); measures of AI adoption/expansion (country‑level AI indicator); governance quality (GQI) and digital infrastructure (DII) as moderators. (Paper constructs/uses cross‑country indices for GQI and DII.)
• Estimation strategies:
  • Two‑step system GMM (to address dynamics and endogeneity in panel context).
  • Two‑stage least squares (2SLS) as robustness for endogeneity concerns.
• Empirical design features:
  • Interaction terms between AI and GQI/DII to test moderation hypotheses.
  • Subsample/heterogeneity analyses by energy efficiency and by low‑AI vs high‑AI stages.
  • Robustness checks across estimators and sample splits (details in paper).
• Controls: standard macroeconomic covariates and country fixed‑effects/dynamic specifications to account for confounders and persistence (paper reports controlling for key drivers; see full text for exact control set).

Implications for AI Economics

• Research implications:
  • Models of AI’s environmental impact must include institutional and infrastructural moderators (not just technological parameters) to avoid misleading aggregate conclusions.
  • Future micro→macro research should quantify rebound effects and scale‑up energy costs (data center buildout, training/deployment energy) alongside firm‑level efficiency gains.
  • Cross‑country heterogeneity is critical: results from China/OECD are not universally generalizable; comparative work across governance and DI regimes is needed.
  • Better measurement: improved, comparable country‑level AI adoption/usage metrics and lifecycle accounting for AI infrastructure emissions would strengthen causal inference.
• Policy implications:
  • Governance first: strengthen institutions (transparency, enforcement, carbon pricing, emissions markets) to steer AI adoption toward decarbonization.
  • Digital infrastructure policy: invest in resilient, energy‑efficient digital infrastructure (green data centers, grid decarbonization) so DI acts as an enabler rather than an emissions driver.
  • Complementary measures: pair AI diffusion with energy‑efficiency standards, incentives for renewable‑powered computing, and regulation of high‑emissions AI uses to limit rebound effects.
  • Priority sequencing: countries with weak governance and poor DI should prioritize institutional and infrastructure upgrades before scaling energy‑intensive AI deployments if climate goals are central.
• Practical takeaway: AI is not an automatic climate win. Its net effect on CO2 depends on governance and digital ecosystem readiness — important considerations for policymakers, investors, and researchers evaluating AI as part of climate strategies.

Reference: Acharjee, P. & Neogi, D. (2026). "Artificial Intelligence: A Blessing or a Curse for Climate Action (SDG 13)? The Moderating Roles of Governance Quality and Digital Infrastructure." Journal of Risk Analysis and Crisis Response, 16(1), 63–86. DOI: https://doi.org/10.54560/jracr.v16i1.750