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 web-search and reasoning AI models deployed widely in 2025 impose substantially higher cumulative environmental costs than earlier AI generations, while transparency about those impacts has declined. Current environmental governance—focused on facilities and training—does not adequately capture model-level and inference-phase impacts. The paper proposes model-level transparency, user choice rights, and international coordination, and offers specific legislative amendments (notably to the EU AI Act, Consumer Rights Directive, and Digital Services Act) to operationalize these reforms.

Key Points

• Empirical evidence shows generative search/reasoning models (which proliferated in 2025) have much larger cumulative environmental impacts than prior AI approaches, largely driven by inference at scale.
• Transparency about AI environmental impacts has decreased even as deployments accelerate, creating an information gap for regulators, users, and researchers.
• Current regulatory regimes across eleven jurisdictions are predominantly:
  • Facility-level rather than model-level,
  • Focused on training emissions rather than inference consumption,
  • Lacking AI-specific energy-disclosure rules except in the EU.
• These features limit regulators’ ability to monitor and mitigate the full environmental externalities of modern AI systems.
• Policy proposal (three-pronged):
• Mandatory model-level transparency covering inference energy consumption, standardized benchmarks, and disclosure of compute locations.
• User rights to opt out of nonessential generative-AI integration and to choose environmentally optimized models.
• International coordination to prevent regulatory arbitrage and ensure consistent standards.
• Concrete legislative recommendations include amendments to the EU AI Act and consumer/digital legislation (Consumer Rights Directive, Digital Services Act) that could be adapted by other jurisdictions.

Data & Methods

• Evidence synthesis: Collation of empirical studies and operational data documenting energy and emissions profiles of generative web-search and reasoning models versus prior AI approaches (focus on 2025-era model families and deployment patterns).
• Regulatory mapping: Comparative analysis of environmental governance across eleven jurisdictions to identify where rules apply (facility vs model), which lifecycle phases are regulated (training vs inference), and the presence/absence of AI-specific energy disclosure requirements.
• Policy design: Development of a three-part policy package and drafting of candidate legislative amendments tailored to existing EU instruments as templates for wider adoption.
• (Implicit) Methodological strengths: cross-disciplinary approach combining engineering-level emissions accounting and legal/policy analysis.
• (Implicit) Limitations: jurisdictional sample size (eleven) and reliance on available empirical/operational data, which the paper notes is increasingly patchy due to declining transparency.

Implications for AI Economics

• Externality internalization: Model-level disclosure and user-choice rights would help internalize negative environmental externalities, shifting costs into firms’ deployment and pricing decisions.
• Incentives for efficiency: Mandatory inference benchmarks and public reporting create market and regulatory incentives to optimize models for energy efficiency (e.g., model compression, routing, edge/off-peak inference).
• Product differentiation and competition: Environmental-performance labeling and user opt-outs could create demand for “eco-optimized” models, opening niche markets and influencing competition among providers.
• Compliance and transaction costs: Firms face additional reporting and compliance costs; small providers may be disproportionately affected unless rules are proportionate.
• Geographic investment and regulatory arbitrage: Without international coordination, providers may relocate compute or obscure compute locations to avoid stricter regimes; harmonized rules reduce these distortions.
• Consumer surplus and adoption dynamics: User rights to opt out may slow default integration of generative AI into digital services, moderating adoption speed and associated network effects.
• Policy as innovation driver: Clear, model-level regulation can direct R&D investments toward energy-efficient architectures and inference techniques, altering the technological frontier in AI economics.
• Market design implications: Requirements to disclose compute locations and inference consumption could enable carbon-pricing mechanisms or procurement rules that favor low-carbon compute supply.

If you want, I can (a) extract likely contents of the proposed legislative amendments in more detail, (b) outline metrics and benchmark designs for model-level inference disclosure, or (c) summarize policy trade-offs and implementation challenges.