Generative web-search and reasoning models deployed in 2025 impose substantially higher cumulative environmental costs even as reporting declines; existing rules—focused on facilities and training—miss model-level and inference emissions. The paper recommends mandatory model-level energy disclosures, consumer opt-outs, and coordinated international reforms, with concrete amendments proposed for EU laws to operationalize these changes.

Main Finding

Generative AI—especially large reasoning models and generative web-search systems—is driving rapidly rising environmental impacts, but legal and regulatory frameworks largely fail to capture or constrain these costs. The paper documents stark inference-level energy differentials (reasoning models can use tens to hundreds of times more energy than non-reasoning models), shows that policy mostly regulates facilities and training (not model-level inference), and proposes concrete policy interventions (mandatory model-level transparency including inference, user rights to non‑generative/green options, and international coordination) to close the transparency and governance gap.

Key Points

• Empirical evidence
  • Reasoning models introduced in 2025 substantially expand the inference-energy frontier: AI Energy Score leaderboards show average reasoning models use ~30× more energy than non-reasoning models; some reasoning instances consume 150–700× more energy than base models.
  • Generative web-search can raise cumulative energy use even if per-query Wh falls slightly: Google’s reported median Gemini prompt (2025) ~0.24 Wh vs ~0.3 Wh per search (2009) but global query volumes rose from ~792B/year (2009) to ~5T/year (2025), implying a large cumulative increase.
  • Published (but proxy-based) estimates: GPT‑5 ~18 Wh per ~1,000-token prompt; ChatGPT-scale traffic (billions of requests/day) implies electricity use comparable to that of hundreds of thousands–millions of households annually—highlighting the potential scale.
• Regulatory mapping (11 jurisdictions + international)
  • Regulatory axes identified: voluntary vs mandatory disclosure; facility-level vs model-level focus; high vs low environmental ambition.
  • Most regulation is facility-centered (data‑centre reporting) and/or training-focused. Few jurisdictions require or operationalize model-level inference disclosures.
  • EU is the most advanced in model-related transparency but with important gaps: the AI Act emphasizes training compute/energy for certain categories but largely omits inference; the General-Purpose AI Code of Practice adds inference guidance but is voluntary; confidentiality rules and trade-secret carve-outs limit public access.
  • Ireland’s CRU decision (Dec 2025) is an example of administrative regulation that links grid connection to additional renewable procurement and on-site resilience—a potentially exportable model for greener data-centre siting.
  • US federal posture is largely deregulatory; state laws (e.g., California) introduce various reporting rules but generally omit model-level energy disclosures (though SB 253 expands corporate Scope 1–3 disclosure obligations).
• Policy proposals (three-pronged)
  • Mandatory model-level transparency requirements capturing: inference energy per representative prompt or per-token, cumulative usage/volumes (or standardized benchmarks), and geographic/compute-location disclosure (to assess grid impacts and additionality).
  • User rights: opt-out from unnecessary generative-AI mediation and the right to choose environmentally optimized (green) models or non-generative alternatives; consumer-rights amendments proposed (EU AI Act, Consumer Rights Directive, DSA as templates).
  • International coordination: harmonized reporting standards to avoid regulatory arbitrage and leakage; standardized benchmarks and sanctions to make disclosures comparable and enforceable.
• Broader observations
  • A growing “transparency gap”: AI usage rises while legally mandated, comparable environmental disclosures decline.
  • Jevons-style rebound effects are plausible: lower marginal costs and easier access increase usage and total energy consumption despite efficiency gains.
  • Data gaps and confidentiality regimes make robust public audit and policy calibration difficult.

Data & Methods

• Empirical inputs
  • AI Energy Score project leaderboards (comparative inference-energy benchmarking across models and tasks; Feb & Dec 2025 used to show expansion of energy-efficiency spread after reasoning models introduced).
  • Public corporate disclosures and blog posts (e.g., Google 2009 and 2025 energy/per-query statements; company sustainability reports).
  • Third-party estimates and proxies (e.g., GPT‑5 per‑prompt energy, ChatGPT request volumes) — authors emphasize these are indicative rather than definitive.
  • Case studies (Web search and reasoning) constructed from the above data and academic literature on AI lifecycle emissions and inference vs training shares.
• Legal-methods
  • Comparative legal analysis across 11 jurisdictions and the international framework, mapping existing statutes, directives, administrative orders, and industry codes along the three regulatory axes (voluntary/mandatory; facility/model; ambition level).
  • Review of EU legislative texts (AI Act, EED, CSRD, CSDDD), national administrative decisions (Ireland CRU), and US federal/state statutes and executive policy.
• Limitations noted by authors
  • Many energy/use estimates rely on third-party proxies or vendor disclosures that are incomplete and non-standardized.
  • Confidentiality and trade-secret exemptions limit replicability and public audit.
  • Heterogeneity across models, modalities, and deployment contexts complicates straight comparisons; lifecycle analyses still evolving.

Implications for AI Economics

• Cost structure and marginal costs
  • Higher inference energy for reasoning/generative models raises the true marginal cost per query/prompt compared to prior generation non-generative systems. Firms that internalize energy costs (e.g., via carbon pricing or renewable procurement) will face higher operating costs, which can alter pricing strategies, product design (prompt length limits, sampling), and margin calculus.
• Investment and location choices
  • Model-level and compute-location disclosures will affect where firms place workloads. Transparency that links compute to grid mix and additionality can shift investment toward regions with cleaner electricity or toward on-site renewables and storage; policies like Ireland’s CRU conditions can materially affect data‑centre siting economics.
• Competition and market structure
  • If only large incumbents can absorb higher inference energy costs, generative reasoning may entrench dominant platforms. Conversely, mandatory model-level disclosure and user-rights to choose green/non-generative options can lower informational barriers and foster competition by enabling green differentiation (green models, compressed/edge models).
• Innovation incentives and model design
  • Explicit inference-focused regulation and disclosure creates incentives for research into energy-efficient architectures (model distillation, retrieval-augmented systems, prompt optimization, hardware-software co-design) and alternative business models (edge inference, pay-per-token pricing that internalizes energy).
• Externalities, regulation, and pricing
  • Transparent model-level metrics are a precondition for properly pricing the environmental externality (via carbon taxes, sectoral levies, or tradeable allowances). Without comparable disclosure, regulators and markets cannot effectively internalize social costs.
• Demand-side effects and consumer welfare
  • User rights to opt out of generative mediation and to select green models could reduce exposure to high-energy models, shifting aggregate demand toward lower-energy alternatives and altering welfare calculations from productivity gains vs environmental costs.
• Macroeconomic and distributional effects
  • Large-scale deployment of energy-intensive AI could affect electricity markets (peak loads, grid resilience), requiring investments in generation, storage, and demand management. Localized impacts (water, land use, pollution from embodied emissions) create distributional conflicts—potentially raising social costs and prompting location-specific regulation.
• Policy design consequences
  • Model-level, inference-focused disclosure is economically meaningful: facility-level metrics mask heterogeneity and prevent accurate cost allocation or benchmarking. International harmonization of metrics and benchmarks reduces regulatory arbitrage and enables cross-border comparability for investors, buyers, and regulators.
• Risk of rebound (Jevons) and market equilibrium
  • Efficiency gains that reduce per-use cost may increase total usage; absent binding constraints or pricing of externalities, aggregate emissions can rise despite per-query efficiency—an equilibrium consideration for welfare-maximizing policy.

Overall, the paper argues that accurate inference-level measurement and disclosure are essential inputs for economic policy to internalize environmental externalities, steer innovation toward energy-efficient technologies, and design market rules that prevent concentration and regulatory arbitrage.