Back to Dashboard
Ecosystem Monitoring / 8.3

8.3 Assessment of Environmental Impacts

2026 Governance Status: Measurement tooling operationalized; end-to-end lifecycle assessment still open

Original Problem in the Paper

The paper frames AI environmental impact as spanning the entire AI life cycle, including training and inference. It argues that accurate end-to-end understanding is needed for policy incentives and penalties that reduce environmental costs. The open problems it identifies include tracking energy use and carbon emissions across dynamic system instances; accounting for energy sources; developing energy ratings for model-task combinations; using tools such as CodeCarbon for real-time carbon estimates; comparing compute costs across devices and cloud deployments; assessing raw-resource costs such as rare-earth mining/refining and data-center water use; and predicting the environmental costs of constructing and maintaining data centers.

July 2026 Update & Trajectory

Energy and carbon measurement is more operational than when the paper was written. CodeCarbon tracks emissions from local or own-hardware computing; EcoLogits estimates GenAI inference impacts for supported remote API/client workflows; the ML CO2 Impact calculator estimates experiment emissions from hardware, runtime, provider, region, carbon intensity, and offsets; NIST's GenAI Profile recommends documenting anticipated environmental impacts and measuring or estimating energy and water for training, fine-tuning, and deployment; IEA provides Energy and AI modelling/datasets plus related 2026 content; and OECD.AI hosts compute-environment policy and data work.

But the end-to-end assessment problem remains open. NIST states that there is currently no agreed method to estimate environmental impacts from generative AI, and EcoLogits' own methodology is scoped to GenAI inference tasks, with explicit exclusions such as training, networking, end-user devices, and end-of-life. The sources reviewed here support gaps around training/inference variability, water, embodied and semiconductor/resource impacts, data-center construction and maintenance, proprietary-provider opacity, and inconsistent method scope. A task-specific AI Energy Score leaderboard and 1-to-5-star labels exist for a limited set of models and tasks, but the sources reviewed here do not establish a universal, regulator-recognized July 2026 energy label for arbitrary model-task-provider combinations.

Deployed / Operationalized

  • CodeCarbon is an open-source Python library with Python API and CLI workflows for tracking CO2 emissions from code running on local or own hardware; its docs point users to EcoLogits for remote GenAI API calls.
  • EcoLogits is an open-source suite for estimating the environmental footprint of GenAI models at inference; its Python library tracks impacts of AI requests made through official client SDKs, and its methodology currently implements LLM inference and video-generation methodologies.
  • The ML CO2 Impact calculator estimates carbon emissions from hardware type, runtime, cloud provider, region of compute, carbon intensity, and offsets, while warning that datacenter PUE is not included unless the user applies it separately.
  • NIST's GenAI Profile identifies environmental impacts as a GenAI risk and suggests documenting anticipated impacts and measuring or estimating energy and water consumption for training, fine-tuning, and deployment.
  • IEA's Energy and AI report provides global and regional modelling and datasets for AI/data-center electricity demand and analyses implications for energy security, emissions, innovation, and affordability; the IEA report page also links related content including a 2026 Key Questions report, an Energy & AI Observatory data explorer, and an Energy and AI dataset.
  • OECD.AI has an AI Compute and the Environment policy/data area, including regional cloud compute availability and AI-compute investment indicators, with OECD caveats that methods for measuring AI are still evolving.
  • Hugging Face announced AI Energy Score Ratings in February 2025, including a public leaderboard ranking 166 AI models across 10 common AI tasks and a 1-to-5-star rating system for model-task energy use.

New Tractable Vectors

  • Logging per-experiment emissions for many local or own-hardware ML workflows using CodeCarbon.
  • Estimating supported GenAI inference-request impacts using EcoLogits, while keeping its scope limits explicit.
  • Estimating experiment-level carbon using hardware, runtime, provider/region, carbon-intensity, offset, and user-supplied PUE assumptions through ML CO2-style calculators.
  • Including energy and water estimates in AI impact assessments, model/system cards, and procurement questionnaires, consistent with NIST suggested actions.
  • Running regional scenario analysis for AI/data-center electricity demand using IEA global/regional modelling and related energy datasets; precise local forecasts still require site, grid, and provider-specific data.
  • Comparing model/task efficiency in open or telemetry-providing contexts using AI Energy Score-style task benchmarks and CodeCarbon-style measurements.
  • Building operational energy/water dashboards for large deployments from hardware utilization, PUE, carbon-intensity, and water-use-effectiveness inputs, while treating embodied-carbon, supply-chain, construction, and end-of-life impacts as unresolved lifecycle-accounting inputs.

Key Open Questions

  • End-to-end lifecycle accounting that consistently includes training, inference, chip fabrication, raw material extraction/refining, data-center construction and maintenance, cooling water, and end-of-life impacts.
  • Consistent methodology boundaries: tools differ in whether they include PUE, networking, storage, data-center cooling, training, end-user devices, embodied impacts, and end-of-life.
  • Energy-source and deployment-location accounting: the source paper and Nature commentary both emphasize that energy sources or location-specific energy mix can materially change carbon-impact estimates.
  • Proprietary-provider opacity: public sources emphasize lack of open data about proprietary-model operation and energy use; detailed claims about query volume, routing, caching, distillation, or per-request energy need separate provider-disclosure or audit evidence.
  • Operational versus lifecycle dashboards: operational energy/water telemetry is becoming more feasible, but full lifecycle dashboards still need embodied-carbon, supply-chain, construction, and end-of-life inputs that remain difficult to measure.
  • Auditable green claims: NIST specifically calls for verifying carbon capture/offset programs and addressing green-washing; REC or offset claims should be tied to claim-specific audit standards and disclosures rather than treated as self-validating.

Evidence & Primary Sources

  • The source paper, *Open Problems in Technical AI Governance*, Section 8.3, states that AI environmental impact extends across the entire AI life cycle, including training and inference; that end-to-end understanding is important for incentives and penalties; and that open problems include tracking energy/carbon across dynamic instances, accounting for energy sources, model-task energy ratings, CodeCarbon-style estimates, device/cloud cost comparisons, raw-resource costs such as rare-earth minerals and data-center water, and predictions for data-center construction and maintenance. (arXiv version read 7 July 2026): https://arxiv.org/pdf/2407.14981
  • CodeCarbon docs describe CodeCarbon as a lightweight, open-source/free Python library for tracking emissions from code running on local or own hardware, provide Python and CLI examples, and direct users to EcoLogits for remote GenAI API calls. (Published 29 June 2026; accessed 7 July 2026): https://docs.codecarbon.io/
  • EcoLogits describes itself as an open-source suite for estimating the environmental footprint of generative AI models at inference; its Python library tracks environmental impacts of AI requests made through official client SDKs, and the project also provides a calculator and API. (Accessed 7 July 2026): https://ecologits.ai/latest/
  • EcoLogits methodology says currently implemented methodologies include LLM inference and video generation; it uses life-cycle-assessment principles for GenAI inference requests across global warming potential, mineral/resource depletion, primary energy, and water consumption, but excludes training, networking, end-user devices, and end-of-life and notes lack of open data and transparent information from AI/cloud providers, hardware manufacturers, and databases. (Accessed 7 July 2026): https://ecologits.ai/latest/methodology/
  • ML CO2 Impact calculator estimates emissions from hardware type, hours used, cloud provider, region of compute, carbon intensity, and offsets, and explicitly states that datacenter PUE is not included unless the user finds the PUE and multiplies the emitted carbon by that number. (Accessed 7 July 2026): https://mlco2.github.io/impact/
  • NIST AI 600-1, *Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile*, identifies Environmental Impacts as a GenAI risk from high compute resource utilization in training or operating GAI models; says impacts vary by pre-training, fine-tuning, inference, modality, hardware, and task/application; states that there is currently no agreed method to estimate environmental impacts from GAI; and includes suggested action MS-2.12 to document anticipated environmental impacts and measure or estimate energy and water for training, fine-tuning, and deployment, plus MS-2.12-004 on verifying carbon capture/offset effectiveness and addressing green-washing. (July 2024): https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf
  • IEA's *Energy and AI* report page states that the report is based on new global and regional modelling and datasets; includes projections for AI electricity consumption and energy sources; analyses energy security, emissions, innovation, and affordability; was published 10 April 2025; and links related content including a 16 April 2026 *Key Questions on Energy and AI* report, the Energy & AI Observatory data explorer, and an Energy and AI dataset. (Accessed 7 July 2026): https://www.iea.org/reports/energy-and-ai
  • OECD.AI's AI Compute and the Environment overview says policymaking requires understanding AI computing capacities and environmental impact, describes work to measure and benchmark domestic AI computing capacity by country and region, and says the expert group is working with AI computing players to understand AI compute energy consumption. (Copyright 2026; accessed 7 July 2026): https://oecd.ai/en/site/compute-climate
  • OECD.AI's compute-climate data page lists data areas including cloud compute availability by region and worldwide VC investments in AI compute by country, while warning that data sources and methods may differ and that methods for measuring AI are still evolving. (Copyright 2026; accessed 7 July 2026): https://oecd.ai/en/site/compute-climate/data
  • Hugging Face announced AI Energy Score Ratings on 11 February 2025, describing a public leaderboard ranking 166 AI models across 10 common AI tasks and a 1-to-5-star rating system for task-specific model energy use, with plans to expand model/task coverage and encourage adoption. https://huggingface.co/blog/sasha/announcing-ai-energy-score
  • The Nature comment “Light bulbs have energy ratings — so why can’t AI chatbots?” proposes an Energy Star-inspired AI rating approach; describes benchmarking energy for 1,000 queries across ten common AI tasks using CodeCarbon; states that its initial results exclude model storage, networking, and data-center cooling overheads and therefore are lower bounds; says carbon emissions depend on deployment location and energy mix; and identifies proprietary-model impenetrability as the biggest challenge. (21 August 2024): https://www.nature.com/articles/d41586-024-02680-3