Among the provisions of Regulation (EU) 2024/1689, the AI Act, that deal with the environment, Article 95 is the only one that addresses all artificial intelligence systems in general terms and, expressly, deployers as well. It imposes no obligations: it provides that the AI Office and the Member States shall facilitate the drawing up of codes of conduct for the voluntary application of specific requirements, including the assessment and minimisation of the impact of AI systems on environmental sustainability (EUR-Lex). In February 2026 the International Telecommunication Union approved Recommendation ITU-T L.1801, devoted to guidelines for assessing the environmental impact of AI systems (ITU). The tools for making a commitment and those for measuring it exist, but they belong to different sources and must be combined with care by anyone intending to declare environmental results.
Article 95: voluntary requirements, objectives and indicators
Article 95 distinguishes two types of code. Paragraph 1 concerns codes intended to foster the voluntary application to systems other than high-risk systems of some or all of the requirements of Chapter III, Section 2. Paragraph 2, by contrast, concerns codes relating to the voluntary application, “including by deployers”, of specific requirements to all AI systems, “on the basis of clear objectives and key performance indicators to measure the achievement of those objectives”. Among the elements listed by way of example, point (b) mentions “assessing and minimising the impact of AI systems on environmental sustainability, including as regards energy-efficient programming and techniques for the efficient design, training and use of AI” (EUR-Lex).
The structure of the provision deserves attention. The legislator links the voluntary commitment to measurability: without indicators, the code does not correspond to the model of Article 95. Recital 165 reiterates this, stating that to be effective codes should be based on clear objectives and on indicators making it possible to measure their achievement. Point (b), moreover, is not limited to the energy consumed in operation: it covers design, training and use, and therefore the entire path of the system.
Article 112(7) completes the picture: by 2 August 2028 and every three years thereafter the Commission shall evaluate the impact and effectiveness of voluntary codes of conduct for systems other than high-risk systems, “including as regards environmental sustainability” (EUR-Lex). A code drawn up today will therefore fall within the Commission’s first evaluation.
Who draws up the codes and what emerges from the Commission’s sources
Paragraph 3 provides that codes of conduct may be drawn up “by individual providers or deployers of AI systems or by organisations representing them or by both”, including with the involvement of any interested stakeholders, including civil society organisations and academia, and that they may cover one or more systems taking into account the similarity of their intended purpose. The Digital Omnibus on AI, Regulation (EU) 2026/1744, replaced paragraph 4: when encouraging and facilitating the drawing up of codes, the AI Office and the Member States now take into account the specific interests and needs of SMEs, including start-ups, and of small mid-cap enterprises (EUR-Lex).
As regards public initiatives, the Commission reports that it has endorsed the code of practice for general-purpose models and that it is facilitating the drawing up of a code on the transparency of AI systems (European Commission). These instruments are linked to other articles of the Regulation. The official Commission pages we consulted show no initiatives devoted to Article 95 codes in the environmental field: for now, the initiative lies with organisations and their associations.
The available metrics: from the data centre to the AI system
Anyone wishing to build indicators already has reference points available, but must choose the level at which to measure. The first level is that of infrastructure. The ISO/IEC 30134 series, maintained by subcommittee ISO/IEC JTC 1/SC 39, defines key performance indicators for data centres: ISO/IEC 30134-2:2026, the second edition published in January 2026 to replace the 2016 edition, governs Power Usage Effectiveness (PUE), that is, the ratio between the energy consumed by the entire facility and that used by the IT equipment (ISO); ISO/IEC 30134-3:2016 defines the Renewable Energy Factor (REF) and specifies its method of calculation and presentation (ISO). The same series includes indicators on the carbon dioxide emissions (Part 8) and water consumption (Part 9) of the data centre (ISO; ISO). On the European reporting obligations applying to data centres and on engineering solutions for reducing their consumption, we refer to our analyses of the energy bill of artificial intelligence and of underwater data centres.
The second level is that of the AI system. Technical report ISO/IEC TR 20226:2025, published in July 2025 by subcommittee ISO/IEC JTC 1/SC 42, provides an overview of the environmental sustainability aspects of AI systems across their lifecycle, including workload, resource use, carbon impact, pollution, waste, transport and location, and of the related potential metrics. Its record specifies that the document does not identify the opportunities through which AI can improve sustainability outcomes (ISO): the report concerns the environmental footprint of AI, not the environmental benefits that AI can deliver. Recommendation ITU-T L.1801, approved on 6 February 2026, in force and freely accessible, provides guidelines for assessing the environmental impact of AI systems (ITU).
The third level is that of the product lifecycle. ISO 14040:2006 describes the principles and framework of life cycle assessment, while ISO 14044:2006 specifies its requirements and guidelines, both maintained by subcommittee ISO/TC 207/SC 5 (ISO; ISO). These are general standards, not designed for AI, but they offer a common language for defining the goal, scope and boundaries of a study.
Then there is the organisational level. ISO/IEC 42001:2023 specifies the requirements for an AI management system, on whose operation and certification we refer to our analysis of the AIMS scheme (ISO); ISO/IEC 42005:2025 provides guidance on the impact assessment of AI systems, and its record cites ethical, social and environmental concerns among the reasons for the standard (ISO). In our view, a code under Article 95 finds in these instruments the natural setting for assigning responsibilities and periodically verifying the indicators; on the relationship between impact assessment and other compliance obligations we refer to our analysis of ISO/IEC 42005.
Estimated values and recorded values
The AI Act itself accepts estimation. For general-purpose models, Annex XI requires documentation of the “known or estimated energy consumption of the model” and, where the consumption is unknown, allows it to be based on the computational resources used (EUR-Lex). The choice is understandable, but for those making voluntary commitments the distinction must be declared. A value estimated from computing capacity is not equivalent to consumption recorded by a meter. A deployer using cloud services, in particular, often has only the data supplied by its provider: an indicator included in a code should be fed by data that the organisation is actually able to obtain, with an indication of the method, the perimeter and the uncertainty.
Critical aspects: voluntary commitments and generic environmental claims
A voluntary commitment becomes risky when it is communicated externally. Directive (EU) 2024/825, which Member States apply from 27 September 2026, has added to the commercial practices considered unfair in all circumstances the making of a generic environmental claim for which the trader is not able to demonstrate recognised excellent environmental performance; among the examples of generic claims, recital 9 cites precisely “energy efficient”. The same Directive lists among the actions that may be misleading a claim related to future environmental performance made without “clear, objective, publicly available and verifiable commitments set out in a detailed and realistic implementation plan that includes measurable and time-bound targets”, regularly verified by an independent third party (EUR-Lex). These rules concern commercial practices directed at consumers; adherence to a code of conduct is not, on its own, sufficient to support an environmental claim made to the public.
The second aspect is the rebound effect. We flag it as our own assessment: a system that is more efficient per individual operation may be used more often, and total consumption may grow even as efficiency improves. For this reason we consider that a credible code should combine intensity indicators with indicators of absolute consumption.
Conclusions
Article 95 offers a regulatory framework: it establishes who may draw up codes, requires measurable objectives and indicators and entrusts the Commission with a periodic evaluation from 2028, but leaves the choice of metrics to organisations. Metrics exist at different levels, from infrastructure to lifecycle, and none of them alone exhausts the problem. Those intending to adhere to or promote a code would be well advised to declare the chosen level of measurement, distinguish recorded values from estimated ones, combine intensity with absolute consumption and link commitments to a management system capable of verifying them, before turning them into external communication.
Author: Valentina Grazia Sapuppo
