AI’s underlying infrastructure must be environmentally sustainable and respectful of our planet’s limits. At the same time, the discussion about the relationship between the benefits of AI systems and their environmental costs must be grounded in facts and figures. Currently, however, the developers and operators of these systems are not providing the data necessary. This absence of publicly available information hampers the development and enactment of effective policies.

The European Union’s AI Act, which is currently being drafted, could for the first time require companies to measure and disclose information on the environmental impact of certain AI systems. The European Parliament has proposed requiring companies to measure the energy and resource consumption of foundation models and high-risk systems. This requires that data collection methods be integrated into these systems.

Critics often claim that the obligation of measuring the environmental impacts of AI systems is too complicated and places too great a burden on small and medium-sized enterprises in particular – and ultimately hinders innovation. But easy-to-use measurement methods already exist for monitoring energy consumption, CO₂-equivalent emissions, water consumption, the use of minerals for hardware and the generation of electronic waste and thus assessing the sustainability of AI systems.

Without a comprehensive life cycle analysis, we can’t adequately capture the environmental footprint of AI models. Providers of large language models (LLMs), in particular, tend to disclose only the direct energy consumption and emissions for a single training cycle. The result is an incomplete picture. Consider, for example, the training of the BLOOM model. Energy consumption during the training phase corresponds to the emission of around 24.7 tons of CO₂ equivalents. But if you factor in hardware production and operational energy, the emissions value doubles. And this does not yet include the continuous emissions produced during the application of the model. Reliable figures from this inference phase are lacking, but first indicators suggest that emissions could be immense – both in the production of the necessary hardware for the application and during operation. That’s why we need to measure how AI systems impact the climate throughout their entire life cycle – from raw material extraction and CO₂ emissions to pollution and water consumption – so that we can enable informed decisions and targeted policies based on solid knowledge.

Companies can already automatically log and report much of the data needed to assess the sustainability of AI systems, such as operational data from computer systems – i.e., how often calculations are performed and how long these processes take. When such metadata is stored in a spreadsheet, it can be used to generate efficiency metrics. Metrics such as “Power Usage Effectiveness” (PuE), for example, show how much energy a data center uses for computing in relation to its overall energy consumption. This parameter makes it possible to compare the energy efficiency of data centers. By examining power consumption, the energy mix of the data center, the carbon intensity of the energy grid and the percentage of CO₂ the provider is potentially compensating, emissions can in turn be calculated.

During system development and training alone, the data listed in the table below should be recorded to comprehensively assess and compare the energy consumption of AI systems. Similar requirements can be formulated for all other environmental impacts, such as emissions, water consumption, mineral extraction and hardware disposal.

The life cycle approach demonstrates that various stakeholders need to provide accurate measurements. For instance, hardware manufacturers such as Nvidia should disclose environmental data on products that are widely used in the development and application of AI models.

There are already many measurement methods available for assessing the environmental impact during system development and training, material extraction, hardware manufacturing and disposal as well as different ones for tracking carbon. Some hardware manufacturers already report emissions levels for a handful of their products. Other approaches for assessing environmental impact during system deployment still need to be developed – reliable metrics and comparable units of measurement for assessing emissions during the application phase, for example.

To be as accurate as possible about the environmental impact during the deployment phase, we propose that AI providers define various standard usage scenarios prior to market launch.

Procedures and methods for assessing environmental impact during system deployment have not yet been established. Developers of AI systems can record their energy consumption during training. However, under the requirements formulated within the AI Act to document energy consumption, this will most likely not be feasible during inference.

Thus, energy consumption during the application phase and the extent of emissions generated during this phase must be estimated. To that end, we are proposing two basic options, perhaps in combination:

There is no lack of technical means for measuring the environmental impact of AI systems. There is, however, still a lack of political will to make AI more sustainable. This is all the more irresponsible considering that AI is a resource- and energy-intensive technology that is becoming increasingly pervasive in all areas of life. The European Parliament has taken some important steps in the right direction to ensure that AI does not further harm the environment, the climate, people and the planet. Nevertheless, data on environmental impacts of AI systems is indispensable. Clear and comprehensive requirements must be introduced for publicly available reports that include such data. This could make AI systems more environmentally sustainable, while at the same time distributing their risks, harmful consequences and benefits more equitably around the world. If the EU is serious about aligning the use of AI with the common good, then it should put all people, and not just Europeans, at the center of its focus. Whatever form the AI Act eventually takes: People will only be truly protected from the negative consequences of AI systems if their impact on the environment is effectively monitored.