Modeling finds AI development is net-polluting: under business-as-usual, modest (ICT-like) AI raises 2100 warming by ~0.1°C while an Industrial-Revolution-scale AI adds ~0.8°C, with climate costs offsetting roughly 20–25% of AI’s economic gains; stronger abatement sharply cuts those losses and in the model supports higher optimal AI investment.

Main Finding

The authors build an integrated assessment model (DICE-AI) that embeds AI as an endogenous sector and find that, under plausible parameterizations, AI development is net polluting. Under current (low) abatement, an ICT-like AI trajectory raises 2100 warming by ~0.1°C and an IR-like (Industrial Revolution–scale) trajectory raises it by ~0.8°C. Climate policy is complementary to AI development: stringent mitigation preserves the returns to AI and increases optimal AI investment. Climate damages and abatement costs, however, materially offset AI’s economic gains unless effective mitigation is implemented.

Key Points

• DICE-AI distinguishes three channels of AI’s climate impact:
  • Computing-related (direct energy use and embodied emissions of AI infrastructure).
  • Immediate application effects (short-term emission reductions if AI is used for mitigation).
  • Economy-wide / system-level effects (AI-driven structural change and growth that alter aggregate emissions).
• Quantitative outcomes (baseline BAU / low-abatement scenario):
  • Global annual emissions in 2100: baseline 61.8 GtCO2 → ICT-like 76.2 GtCO2; IR-like 111.1 GtCO2.
  • Decomposition (2100): ICT-like — computing-related ≈ 9.9 GtCO2, economy-wide ≈ +12.8 GtCO2; IR-like — computing-related ≈ 40.4 GtCO2, economy-wide ≈ +21.2 GtCO2.
  • Warming: ICT-like +0.1°C by 2100; IR-like +0.8°C (relative to the DICE baseline path reported in the paper).
• Welfare and costs:
  • Permanent consumption-equivalent losses under BAU (no adaptation): ICT-like ≈ 2.73%; IR-like ≈ 25.73%.
  • Allowing AI-powered adaptation lowers losses: ICT ≈ 1.59%; IR ≈ 23.24%.
  • Optimal carbon pricing reduces residual consumption losses to: ICT ≈ 0.58%; IR ≈ 2.47%.
  • Combined mitigation + AI adaptation leads to even smaller realized losses (e.g., ICT ≈ 0.28%, IR ≈ 0.99%).
  • Under BAU, climate costs offset roughly 20% (ICT) and 25% (IR) of AI’s economic gains; under optimal carbon pricing these offsets fall to a few percent (e.g., ICT 3.6%, IR 1.1%).
• Social cost of carbon (SCC) and abatement:
  • ICT-like AI (BAU) tends to lower SCC relative to DICE because AI-powered adaptation reduces vulnerability (example SCC comparisons: 2025/2050/2100 — ICT-BAU: 61/126/376 $/tCO2 vs DICE-BAU: 84/166/492 $/tCO2).
  • IR-like AI (BAU) greatly increases SCC (examples: 2050 ≈ $425/tCO2; 2100 ≈ $4,107/tCO2 in the paper), necessitating much stronger mitigation.
• Optimal AI investment response to policy:
  • Under BAU, AI investment rates can decline because climate damages reduce expected returns. Imposing optimal carbon pricing or a 2°C constraint reverses this decline and raises optimal AI investment—i.e., mitigation is complementary to sustained AI investment.
  • Example (abstract figures): under low-abatement by 2100 the optimal AI investment rates are ~3.3% (ICT) and 5.1% (IR); with a 2°C constraint they rise to ~3.7% (ICT) and 12.7% (IR).
• Sensitivity:
  • The damage–output elasticity (how damages scale with output/temperature) is the dominant parameter driving outcomes.
  • To overturn the conclusion that AI is net polluting would require unrealistically large immediate-application abatement effects (carbon-intensity reductions by ~28% for ICT-like and ~90% for IR-like by 2050), implying reversal is unlikely in realistic scenarios.

Data & Methods

• Model: Extended DICE-2023 (DICE-AI) with AI as an explicit intermediate production sector.
  • AI sector has higher capital-embodied carbon intensity (captures energy-hungry data centers and lifecycle infrastructure emissions).
  • Immediate application channel reduces aggregate carbon intensity (if AI used for mitigation).
  • AI capital stock enters the damage function to capture AI-enabled adaptation (reduced vulnerability).
  • Two growth channels for AI: higher TFP growth (innovation) and higher capital income share (automation).
• Calibration:
  • Two stylized AI prospects calibrated from literature: ICT-like (modest, historical-ICT analog) and IR-like (transformative, industrial-revolution analog).
  • Parameter ranges from surveys of empirical and modeling studies used for Monte Carlo runs (2,000 draws over key uncertainty parameters).
• Welfare metric: permanent-equivalent-variation (PEV) measured as percent changes in consumption.
• Simplifying assumptions / scope limits:
  • No explicit energy sector; energy intensity effects are embedded via carbon intensity of capital.
  • Abstracts from detailed AI technology pathways, sectoral spillovers, distributional outcomes.
  • Focuses on global aggregates and welfare-maximizing planner perspectives.

Implications for AI Economics

• Externalities matter: AI investment decisions and evaluations should internalize climate externalities beyond direct computing emissions—especially the economy-wide emissions consequences of AI-driven growth and automation.
• Policy complementarity: Well-designed climate policy (carbon pricing or binding temperature targets) can increase the social returns to AI by protecting future returns from climate damages—so policymakers should not view mitigation and AI development as irreconcilable goals.
• Priority actions for AI sector and firms:
  • Rapid decarbonization of AI infrastructure (data centers, embodied hardware) is necessary but not sufficient; system-level emissions from economic expansion driven by AI must be considered.
  • Investment in AI applications that genuinely reduce aggregate carbon intensity is valuable, but the bar is high to fully offset system-level effects—targeted R&D and deployment in high-impact mitigation domains are required.
• Research and policy priorities:
  • Better empirical measurement of AI’s system-level effects on economic structure, energy demand, and long-run growth is critical (these drive most model uncertainty).
  • Improved estimates of AI-enabled adaptation’s effectiveness and its interaction with mitigation policies.
  • Integrating distributional impacts and sectoral detail into models to inform targeted regulation and incentives (e.g., energy-efficiency standards for AI hardware, carbon pricing revenues used to accelerate clean tech).
• Caution on optimistic narratives: Even if AI generates large economic gains, these gains can amplify emissions quickly and impose large climate costs absent strong mitigation. Economists and policymakers should weigh growth-promoting AI policies alongside robust climate policy to preserve net welfare gains.