The European Union Artificial Intelligence Regulation—the EU AI Act, officially known as Regulation (EU) 2024/1689—entered into force on 1 August 2024, and applies progressively. It is the world’s first comprehensive, horizontal AI regulation, based on a risk-based approach: it prohibits certain practices, imposes strict obligations on high-risk systems, and introduces transparency rules for limited-risk systems.
Introduction: The Regulatory Framework and International Context
The development of the Regulation is shaped by the EU’s twin digital and green transitions. As part of the European Green Deal and the digital strategy, the AI Act aims to ensure that AI contributes to sustainable development while minimizing risks to health, safety, fundamental rights, and the environment. The preamble and recitals (eg, Recital 4) emphasize that AI can improve energy efficiency, support biodiversity protection, combat climate change, and optimize resource allocation. At the same time, the Regulation acknowledges the negative environmental impacts of AI, such as high energy consumption, high water use, and electronic waste generation.
‘AI can improve energy efficiency, support biodiversity protection, combat climate change, and optimize resource allocation’
In the international regulatory landscape, the EU AI Act is considered ambitious. The United States follows a sectoral, largely voluntary approach (eg, NIST framework, Biden-era executive order) with fewer mandatory environmental requirements. In China, a state-directed, security-focused regulatory model prevails, treating green AI as a strategic tool, though transparency remains comparatively lower. Canada focuses primarily on government automated decision-making. UNESCO and OECD recommendations include ethical and sustainability principles but are non-binding.
The EU AI Act therefore seeks to position the EU as a global regulatory leader, although critics argue that the environmental dimension was weakened in comparison to the European Parliament’s version (that is, stronger obligations disappeared during trilogue negotiations).
This analysis examines the extent to which the Regulation can be considered sustainable from an environmental perspective. It not only reviews the green provisions but also assesses their applicability and challenges.
Green Provisions in the Regulation
1. Resource Efficiency and Harmonized Standards (Article 40)
Article 40 empowers the European Commission to develop harmonized standards, in cooperation with standardization organizations (eg, CEN, CENELEC), to improve the resource performance of AI systems and general-purpose AI (GPAI) models. The focus is on energy efficiency and the reduction of energy and other resource consumption throughout the entire life cycle (training, inference, deployment). It applies primarily to high-risk AI systems and GPAI models. The standards are voluntary, but compliance creates a presumption of conformity with the Regulation’s requirements. The Commission may request the development of standards covering reporting, documentation, and measurement.
The main challenges are as follows: developing standards is slow, and they must remain technology-neutral, which is difficult in a rapidly evolving field like AI. There are no mandatory minimum thresholds (eg, specific CO₂ quotas). Life cycle assessment (LCA) is complex: measuring the entire chain (hardware manufacturing, data centre, e-waste) is complicated. Compliance can be costly for small and medium-sized enterprises. Critics argue that this is more of an incentive than an enforcement mechanism and does not guarantee actual reductions.
2. Energy Consumption Documentation for GPAI Models (Annex XI and Related Articles)
Providers of GPAI models (eg, foundation models such as GPT-style models) must prepare technical documentation that includes the known or estimated energy consumption of the model. The AI Office may request this information at any time without justification. This requirement applies to large general-purpose models; transitional arrangements exist for earlier models placed on the market before 2 August 2025. Where the energy consumption of the model is unknown, the estimate may be based on information about computational resources used.
This also means that estimates are permitted, and there is no mandatory precise measurement or public carbon footprint disclosure (only to authorities). The inference phase is often omitted, even though it dominates real-world usage. Global supply chains make full transparency difficult. Penalties are high (up to 3 per cent of global annual turnover), as detailed in Article 99, but regulatory enforcement capacity may be insufficient. The provision encourages documentation rather than direct reduction.
3. Risk Management for High-Risk Systems (Article 9)
High-risk AI systems must implement a risk management system that may include environmental risks (though the environment is not an explicit primary category). This requires continuous monitoring and risk mitigation.
However, the fact that the environment is not a standalone risk category presents a challenge. As a result, the approach remains largely procedural, meaning that it may primarily require documentation rather than substantive mitigation. Many AI systems with significant environmental impact do not qualify as high-risk (eg, large generative models), and the broad interpretation may result in inconsistent or weak implementation. There is no mandatory sustainability impact assessment (SIA).
4. Voluntary Codes of Conduct (Article 95)
The AI Office and the Member States encourage the development of voluntary codes of conduct. These codes may address the minimization of the environmental impact of AI, energy-efficient programming, and sustainable design. They apply primarily to non-high-risk AI systems, although their scope can be extended.
Earlier versions proposed by the European Parliament contained stronger obligations, but under the present regulation, participation remains purely voluntary for all stakeholders. The codes are entirely voluntary, with no sanctions for non-compliance, often leading to limited uptake and a significant risk of ‘greenwashing’. The effectiveness of these codes would likely improve if they were combined with certain mandatory elements.
5. Regulatory Sandboxes and Innovation Support
Regulatory sandboxes allow for the testing of AI systems while incorporating considerations of environmental protection, climate change mitigation, and biodiversity. These sandboxes operate in a controlled environment, making them particularly suitable for developing and testing green AI solutions. However, challenges remain: the number of projects is limited, the process is resource-intensive, and the approach is not yet scalable in a systematic way.
Other green elements include recitals that emphasize AI’s positive role in supporting the European Green Deal, exemptions or faster approval pathways for AI systems serving environmental objectives, and transparency obligations (eg, the logging of energy consumption in certain cases).
Overall, these provisions are more procedural and voluntary than mandatory and measurable. Despite the Regulation’s positive intent, there are notable ‘green blind spots’: the lack of mandatory LCA, inference-phase reporting, concrete thresholds, and strong enforcement mechanisms.
Conclusion: Opportunities for Improvement
The environmental sustainability of the AI Act can be rated as moderate: a step forward compared with the absence of regulation, yet it remains insufficiently ambitious given the climate crisis and the explosive growth of AI (especially data centre energy demand).
‘The environmental sustainability of the AI Act can be rated as moderate’
To strengthen the environmental effectiveness of the Regulation, several additional steps should be taken. A mandatory SIA should be introduced for all large-scale or environmentally significant models, including the inference phase. In addition, concrete energy efficiency thresholds and carbon quotas aligned with the EU Emissions Trading System (ETS) should be established.
Full life cycle reporting and public transparency are also necessary. For high-risk systems, either mandatory green codes of conduct should apply, or appropriate incentives—such as tax breaks and penalties—should be introduced. A stronger focus should be placed on both green AI solutions within regulatory sandboxes and enhanced international cooperation. Finally, regular reviews—for example, every two years—should be implemented to ensure that the framework keeps pace with technological developments.
In summary, the AI Act provides an appropriate framework, but its sustainability provisions remain more aspirational than obligatory. To become truly green, it must move beyond voluntarism toward measurable, enforceable mechanisms. Only then can AI truly protect the planet for future generations.
Related articles: