Environmental Impact of AI Training

Regulatory approaches

The instruments that touch this topic do so through three distinct modalities, none of which sets a binding emissions or energy ceiling. The first is mandatory documentation confined to a single actor class: under the EU AI Act, providers of general-purpose AI (GPAI) models must compile technical documentation that includes the "known or estimated energy consumption" of the model, and where consumption is unknown may estimate it from the computational resources used (EU AI Act Art. 53(1)(a) read with Annex XI; AI Office GPAI Model Documentation Form, July 2025). This is a transparency duty, not a performance standard, and legal analysis notes it interlocks with broader levers such as the Energy Efficiency Directive's data-centre reporting and corporate sustainability reporting rather than standing alone ¹. The second modality is delegated standard-setting: Art. 40(2) directs the Commission to request standardisation deliverables "on reporting and documentation processes to improve AI systems' resource performance," including reducing a high-risk system's energy and resource consumption over its lifecycle and the "energy-efficient development of general-purpose AI models" (EU AI Act Art. 40(2)) — an aim that maps onto the literature's case that compute-efficiency be treated as a first-class, reportable metric ². The third is purely voluntary: Art. 95(2)(b) tasks the AI Office and Member States with facilitating codes of conduct on "assessing and minimising the impact of AI systems on environmental sustainability, including … energy-efficient programming and techniques for the efficient design, training and use of AI" (EU AI Act Art. 95(2)(b)). Outside AI-specific law, France regulates the same footprint through general digital-environment rules: the REEN Act (Loi n° 2021-1485) underpins the ARCEP–ADEME digital-footprint observatory created in December 2024 (ARCEP, Dec. 2024). The composite pattern is procedural and disclosure-led rather than substantive. At the international soft-law level, UNESCO's Recommendation on the Ethics of Artificial Intelligence (2021) adds a further assessment-led layer: under its 'Environment and Ecosystems' policy area (para 84), Member States and business enterprises should assess the direct and indirect environmental impact across the AI life cycle, including its carbon footprint and energy consumption (UNESCO Recommendation on the Ethics of AI 2021, para 84).

Key fault lines

The contested questions are less about whether AI consumes energy and water than about who must measure what, and how. A first fault line is the lifecycle boundary: training emissions are a one-time capital cost while inference recurs across a model's commercial life, and there is no settled rule for allocating training impact across served queries — a problem sharpened because frontier providers rarely disclose lifetime inference volumes, and because deployment energy per 1,000 inferences can dwarf per-task systems ³. Full-lifecycle accounting compounds this: BLOOM's training emitted ~24.7 tCO2e from dynamic power but ~50.5 tCO2e once manufacturing and idle consumption were counted ⁴, echoing evidence that embodied hardware carbon can rival operational emissions ⁵. A second concerns accounting method: the GHG Protocol mandates dual location-based and market-based Scope 2 reporting, which can yield divergent figures for the same data centre depending on power-purchase agreements and renewable-energy certificates, so a provider can appear near-zero-carbon by one method and materially emitting by the other (GHG Protocol Scope 2 Guidance 2015). Water has no equivalent disclosure norm at all, even though training a model such as GPT-3 can evaporate millions of litres ⁶. A third, jurisdictional, fault line is scope: critics argue the EU framework is a "missed opportunity" because mandatory measurement reaches GPAI models alone, leaves high-risk and ordinary systems unaddressed, treats water and minerals as residual "other resources," and relies on voluntary codes whose precedents have underperformed (Heinrich Böll Stiftung 2024; EU AI Act Art. 40(2), Art. 95). Underlying these is a structural dispute — editorial synthesis — over whether AI-specific instruments are the right vehicle at all, or whether grid-decarbonisation and general data-centre energy law (e.g., France's REEN regime) better address an impact that is fundamentally about electricity and water, not algorithms.

Trajectory / what's changing

The governance picture is shifting from voluntary aspiration toward operational measurement, though still without binding limits. The pivotal recent step is the EU GPAI Code of Practice, whose final text the AI Office published on 10 July 2025; its transparency chapter operationalises the Annex XI energy-documentation duty through a Model Documentation Form, with obligations for models placed on the market from 2 August 2025 and a transition window to 2 August 2027 for pre-existing models (EU AI Act Art. 53; AI Office GPAI Code of Practice, July 2025). This push toward standardised metrics responds to a decade of method papers arguing footprints can be made reproducible from runtime, hardware and grid location ⁷ and to projections that AI servers could draw 85–134 TWh/year by 2027 absent disclosure ⁸. Building on this, the European Commission ran a targeted consultation on measuring the energy consumption and emissions of AI models from 7 April to 1 June 2026, explicitly to design a measurement framework for the Act's energy objectives and a possible AI energy-and-emissions label spanning training and inference (European Commission, Apr.–June 2026). At national level, France's ARCEP–ADEME observatory (created December 2024 under the REEN Act) is extending verified digital-footprint reporting toward AI-specific lifecycle stages (ARCEP, Dec. 2024). In the United States, the Artificial Intelligence Environmental Impacts Act (S. 3732, 118th Congress, introduced 1 February 2024) would direct an EPA study, a NIST stakeholder consortium, and a voluntary reporting system — but it remains a measurement-and-study bill, not yet enacted, and imposes no caps (Congress.gov, S. 3732). The common direction across these dated developments is toward standardised metrics and labels rather than enforceable thresholds. Reinforcing this turn toward standardised taxonomies, the NIST AI RMF Generative AI Profile (NIST AI 600-1) names Environmental Impacts as one of twelve GenAI risk categories, lending a US risk-management vocabulary to the measurement push even though it, too, sets no threshold.

Historical primacy & cross-jurisdiction tension

First addressed by OECD AI Principles (Recommendation) on 2019-05-22 (implicit). Subsequent regimes have either codified, diverged from, or remained silent on this baseline.

• Forum-shoppingNIST AI RMF Generative AI Profile↔Executive Order 14179 — Removing Barriers to American Leadership in AI
• Forum-shoppingUNESCO Recommendation on the Ethics of Artificial Intelligence↔UK Pro-Innovation Approach to AI Regulation (White Paper)

Silent regimes — gap signal

Instruments that do not address Environmental Impact of AI Training — candidates for future policy work.

• Executive Order 14179 — Removing Barriers to American Leadership in AIUS
• UK Pro-Innovation Approach to AI Regulation (White Paper)UK
• Interim Measures for Generative AI Service ManagementCN
• NIST AI Risk Management FrameworkUS
• Bletchley Declaration on AI Safetyglobal
• Seoul Declaration on Safe, Innovative and Inclusive AIglobal
• California SB-1047: Safe and Secure Innovation for Frontier AI Models ActUS
• India Digital Personal Data Protection Act + AI Advisory (MEITY)IN
• Brazil AI Bill (PL 2338/2023)BR
• Anthropic Responsible Scaling Policy (RSP) v2US
• OpenAI Preparedness FrameworkUS
• Google DeepMind Frontier Safety FrameworkUS
• Meta Frontier AI FrameworkUS
• UK-US AI Safety Institute Memorandum of Understandingglobal
• White House Voluntary AI CommitmentsUS
• Singapore Model AI Governance Framework for Generative AISG
• Japan METI AI Guidelines for BusinessJP
• General Data Protection Regulation (GDPR)EU
• EU General-Purpose AI Code of PracticeEU
• OMB Memorandum M-24-10 (Advancing Governance, Innovation, and Risk Management for Agency Use of AI)US
• GSA Generative AI and Specialized Computing Infrastructure Acquisition Resource GuideUS
• DoD Responsible AI Strategy and Implementation PathwayUS
• FedRAMP AI Cloud Procurement GuidanceUS
• DFARS Subpart 252.204 (Safeguarding Covered Defense Information and Cyber Incident Reporting)US
• California SB-53: Transparency in Frontier Artificial Intelligence Act (TFAIA)US
• California SB 243: Companion ChatbotsUS
• California SB 942: AI Transparency ActUS
• Revised Product Liability Directive (Directive (EU) 2024/2853)EU
• Directive (EU) 2024/2831 on improving working conditions in platform workEU
• Provisions on the Administration of Deep Synthesis of Internet Information ServicesCN
• New York RAISE Act: Responsible AI Safety and Education ActUS
• TAKE IT DOWN Act (Tools to Address Known Exploitation by Immobilizing Technological Deepfakes on Websites and Networks Act)US
• Japan AI Promotion Act (Act on the Promotion of Research, Development and Utilization of AI-Related Technologies)JP

Does governance work? — the social-science evidence

What the peer-reviewed social science shows: whether the harm this topic addresses is empirically real, and whether governance of it works. The badge is the epistemic status of the evidence(not the policy debate) — “thin” or “absent” efficacy evidence is itself a finding (the “second silence”). Each epistemic-status label is Policy Window's editorial assessment of the cited evidence base (a structured classification), not a verdict any single source issues.

• Is the harm real?evidence: established

  The resource demands of AI compute are empirically documented at the model level: Strubell et al. (2019) quantified large-NLP training energy/carbon, Luccioni et al. (2023) estimated BLOOM's training at ~24.7 tCO2eq (dynamic power) rising to ~50.5 tCO2eq with manufacturing and deployment, Li et al. (2023) estimated GPT-3-scale training in US datacenters can evaporate on the order of hundreds of thousands of litres of freshwater (their central figure ~700,000 L), and Luccioni, Jernite & Strubell (2024) showed generative inference is markedly more energy-intensive per query than task-specific models; at the macro scale the IEA (2024) and de Vries (2023) document rapidly rising datacenter electricity demand. Honest caveat: absolute estimates vary by up to orders of magnitude with grid carbon intensity, hardware, utilisation and accounting boundaries, and cleanly attributing the AI-specific increment (versus general datacenter and crypto growth) remains genuinely contested — the IEA itself bundles AI with datacenters and crypto — so the existence of the footprint is established while its magnitude and trajectory are not.

  Sources: Strubell, Ganesh & McCallum 2019 (ACL Anthology P19-1355; 'Energy and Policy Considerations for Deep Learning in NLP'); Luccioni, Viguier & Ligozat 2023 (JMLR 24; BLOOM 176B carbon footprint, 24.7/50.5 tCO2eq; arXiv:2211.02001); Li, Yang, Islam & Ren 2023 (arXiv:2304.03271, 'Making AI Less Thirsty', later Comm. ACM 2025); Luccioni, Jernite & Strubell 2024 (ACM FAccT '24, 'Power Hungry Processing', DOI 10.1145/3630106.3658542); de Vries 2023 (Joule 7(10):2191-2194, DOI 10.1016/j.joule.2023.09.004); IEA 2024 (Electricity 2024)

• Does governance work?evidence: absent

  There is no impact evaluation showing that any AI-specific environmental-governance instrument reduces energy, water or carbon use, because every named instrument is voluntary or non-binding and very recent: EU AI Act Art. 95 codes of conduct are explicitly optional with no sanctions, and NIST AI 600-1 and the G7 Hiroshima Code are guidance, not enforceable caps. The closest analogue evaluation literature is divided in a way that disfavours the voluntary form chosen here: rigorous reviews find voluntary environmental programs generally fail to produce significant abatement beyond business-as-usual (Koehler 2007; Morgenstern & Pizer 2007), whereas the one form with credible positive evidence is mandatory disclosure (Downar et al. 2021 found a UK carbon-reporting mandate cut emissions ~8% versus a control group) which the AI instruments do not yet impose, leaving the proposition that AI environmental governance works essentially untested.

  Sources: EU AI Act Art. 95 / Recital 142 (Reg. (EU) 2024/1689); NIST AI 600-1 (2024, GenAI Profile); G7 Hiroshima Process International Code of Conduct (30 Oct 2023); Koehler 2007 (Policy Studies Journal 35(4):689-722); Morgenstern & Pizer (eds.) 2007 (Reality Check, RFF Press); Downar, Ernstberger, Reichelstein, Schwenen & Zaklan 2021 (Review of Accounting Studies 26(3):1137-1175)

Related topics

Topics with similar coverage patterns across the tracked instruments.

• AI-Driven Worker Displacement
• Technological Sovereignty
• Development-Rights Framings
• AI in Education