8.3 Assessment of Environmental Impacts
Original Problem in the Paper
Paper motivation/open problem: AI environmental impact spans the “entire AI life cycle,” including training and inference; accurate end-to-end understanding is needed to set incentives/penalties for reducing environmental costs. Open problems: logistical difficulty tracking energy/carbon across dynamic system instances; current efforts struggle with energy sources; develop energy ratings for model-task combinations; real-time emissions tools such as CodeCarbon; compare compute costs across devices/cloud; assess raw resource costs including rare-earth mining/refining and data-center water; provide end-to-end predictions for constructing/maintaining data centers.
July 2026 Update & Trajectory
Energy/carbon measurement is more operational: CodeCarbon/EcoLogits-style tools, ML CO2 Impact calculators, NIST GenAI environmental suggested actions, IEA Energy & AI modelling/observatory, and OECD compute-climate work. But end-to-end AI environmental assessment remains unsolved: inference at scale, water, grid marginal emissions, embodied carbon, semiconductor/mineral impacts, data-center construction, rebound effects, and provider opacity are not consistently measured or reported. Verified 2026 status for IEA related tools/2026 key questions listing and CodeCarbon docs; no verified universal July 2026 energy-rating label for model-task combinations.
Deployed / Operationalized
- CodeCarbon tracks local-compute emissions via Python/CLI; EcoLogits tracks remote GenAI API call emissions.
- ML CO2 Impact calculator estimates carbon emissions from hardware, runtime, provider, region, carbon intensity, and offsets, with caveat that PUE is excluded unless user supplies it.
- NIST GenAI Profile includes environmental risk and suggested actions to document anticipated impacts and measure/estimate energy and water for training, fine-tuning, and deployment.
- IEA Energy & AI report and Energy & AI Observatory provide global/regional modelling and datasets for data-center/AI electricity demand and climate/energy impacts.
- OECD.AI has AI Compute and Environment/compute-climate policy area and live AI compute data.
New Tractable Vectors
- Per-experiment and per-API-call carbon logging is now straightforward for many ML workflows.
- Regional data-center electricity-demand forecasting is tractable using IEA modelling and grid-demand data.
- Include energy/water estimates in AI impact assessments, model cards/system cards, and procurement questionnaires.
- Compare model/task efficiency across providers and deployment modes if providers expose enough telemetry.
- Use hardware utilization, PUE, carbon-intensity, and water-use-effectiveness data to build lifecycle dashboards for large deployments.
Key Open Questions
- End-to-end lifecycle accounting including chip fabrication, rare-earth/mineral extraction, data-center construction, cooling water, backup generation, and e-waste.
- Marginal versus average grid emissions and time/location-based accounting for dynamic cloud workloads.
- Inference-scale opacity: providers rarely publish query volume, model routing, cache hits, distillation use, or per-request energy.
- Environmental rebound: efficiency gains can lower cost and increase total demand.
- Auditable green claims: offsets/RECs, carbon-free-energy matching, and water-positive commitments require independent verification.
Evidence & Primary Sources
- IEA Energy and AI report uses global/regional modelling and datasets to project AI/data-center electricity consumption and analyze energy security, emissions, innovation, and affordability; report published 10 April 2025 and related Energy & AI Observatory exists. (10 April 2025; related content includes 2026 Key Questions and Observatory): https://www.iea.org/reports/energy-and-ai
- CodeCarbon is an open-source Python library/CLI to track and reduce CO2 emissions from local computing; docs point to EcoLogits for remote GenAI API calls. (Accessed 7 July 2026): https://docs.codecarbon.io/
- ML CO2 Impact calculator estimates emissions from hardware, runtime, cloud provider/region, carbon intensity, and offsets, but notes PUE is not included unless added by user. (Accessed 7 July 2026): https://mlco2.github.io/impact/
- NIST GenAI Profile identifies environmental impacts as a GenAI risk and suggests documenting anticipated impacts and measuring/estimating energy and water for training, fine-tuning, and deployment. (July 2024): https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf
- OECD.AI data/policy catalogue includes AI Compute and the Environment and AI compute live-data areas. (Copyright 2026): https://oecd.ai/en/data
- NIST GenAI Profile warns there is no agreed method to estimate environmental impacts from GAI and that impacts vary by training/inference, modality, hardware, and task. (July 2024): https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf