The background material compiles the knowledge base for AI recommendations and supports education providers and teachers in developing AI competence and utilizing AI locally.

The background material covers various themes and should be used as a whole.

Sustainable development means ensuring the well-being of people within the limits of Earth's ecological capacity.  The most important concept in sustainable development is interdependence, which means that the three pillars of sustainable development, namely social, economic and ecological sustainability, are inextricably linked. For example, social sustainability cannot be achieved without also ensuring economic and environmental sustainability.

Digitalisation may be useful in strengthening sustainable development. This means utilising technology for the development of environmentally friendly measures. Sustainable business operations make supply chains transparent, highlight life cycle impacts and integrate resource efficiency into operations, for example. On the other hand, the environmental impacts of digitalisation must be recognised, including the materials needed for the manufacture of devices and the energy consumed by their use. This section focuses particularly on the environmental aspects of AI.

Sustainability can be measured by describing the handprint and footprint of activities.  The concept of footprint is used to describe the negative environmental impacts of human activity. The handprint, on the other hand, refers to the positive impacts of an activity, which support sustainable development. The carbon handprint describes the positive climate impact of products and services. 

Back to top

The rapid increase in the use of AI has posed challenges to environmental sustainability and will continue to do so in the future. Data centres and all digital devices are dependent on critical minerals and rare earth elements, which are often extracted unsustainably. Data centres also use a great deal of water and consume vast amounts of electricity. In addition to this, they produce electrical and electronic waste, which often contains hazardous substances.

Data centres consumed an estimated 460 TWh (terawatt hours) of energy in 2022. The International Energy Agency estimates that the global electricity consumption of data centres, cryptocurrencies and AI will be between 650 and 1,050 TWh in 2026. The range is broad because consumption depends on the pace of deployment, efficiency improvements and AI and cryptocurrency trends. The increase compared to 2022 is roughly equivalent to the annual electricity consumption of Sweden in the lowest consumption forecast and of Germany in the highest forecast. At the same time, the AI industry is expected to grow exponentially and consume at least ten times the amount of energy by 2026 compared to 2023 [1].

The energy consumption and carbon dioxide emissions of the use of machine learning models remained an unresearched area for a long time. The situation changed in 2019 when a group of researchers studied and quantified the energy consumption of AI models used in natural language processing (NLP). In other words, only in recent years have researchers begun to study how much energy is spent on the training and utilisation of large language models, for example. The use of AI increased sharply when generative AI was made available to everyone via the internet on various digital devices. As a result, most of the total energy consumption caused by AI comes from the actual use of AI.

Info box 2 contains more detailed information on the energy consumption caused by the use of machine learning models, meaning that the figures do not include the electricity consumed in the training phase of AI models, for example. For more information, please see the study [2]. 

The energy consumption of AI was examined in a study carried out in 2024, which involved researchers examining 88 different machine learning models using 10 tasks of five different modalities. They were divided into the following categories: a) Text classification b) Image classification c) Text to text d) Image to text e) Text to image

According to the study, the different types of tasks mentioned above require very different amounts of energy. Studies show that the classification of text and images requires considerably less energy than generating text and images. Text classification requires the least energy from the tasks examined. From a computer technology point of view, this is only natural. To concretise energy consumption, the study used the charging of a smartphone, which requires an average of 22 Wh of energy, as reference. According to the study, a thousand prompts to generate text required roughly the same amount of energy on average as charging a smartphone battery two times. The content and length of the generated text have an effect on energy consumption, as does the efficiency of the AI model. Generating an image requires assigning a colour to all the pixels that make up the image. Text, on the other hand, consists of lines of characters (letters). This is why generating an image is much more energy-intensive than generating text. The efficiency of AI models varies, leading to large variation in energy consumption. Generating ten images with the least efficient image generator consumed the same amount of energy as charging a smartphone from empty to full 5.2 times. 

The study also examined the impact of the size of the AI system measured by the number of parameters on energy consumption. The widely used large image-generating AI models can consume up to 6,800 times more energy than the least energy-intensive text generator. The most energy-intensive image generator (stable diffusion-xl-base-1) required approximately 5.36 kWh of energy to generate 1,000 images. To concretise this, the researchers used a freely available online converter: this amount of energy corresponds to 6.6 km of driving on an average combustion engine car in their country (national average 8.7 km). The relationship between energy and CO2e is affected by the different ways in which energy is generated in different places around the world, which means that the variation in the ratio can be significant. According to one estimate, by August 2023 AI had been used to generate a total of 15.5 billion images, of which 80% have been generated using applications based on Stable Diffusion. Assuming that this trend continues, image generation will have a significant ecological footprint.

According to research, using multi-purpose AI applications for more limited tasks is more energy-intensive than using task-specific AI applications developed for a specific purpose. Because of this, researchers see the current trend of using generative AI for all kinds of tasks as a negative development. As examples, researchers mention the increased use of generative AI for information searching and online navigation. They point out that the provided benefits should be weighed more consciously against the increase in energy consumption and emissions. This phenomenon is widely known among AI developers, and some companies are investing in more narrowly focused AI applications.

Millions of users use large image-generating AI applications every day, causing a significant ecological footprint. Based on an analysis by the IEA, the unnecessary generation of images should be limited. The generation of video and music were not included in this analysis, and the energy consumption of the generation of multimedia content remains an unexplored area as of yet. Every user of AI can affect the footprint of AI through their own discretion and behaviour. For example, according to the IEA’s analysis, conventional information search using Open AI’s ChatGPT application requires approximately 10 times more energy than conventional information search using Google's search engine [1]. 

We can expect the share of AI of the global energy consumption of ICT technology to grow as more and more companies, governments and organisations use AI to increase the efficiency and productivity of their operations and teams.

Training AI systems or machine learning models also requires a great deal of energy. The largest IT companies have calculations of how the total energy consumption of AI is divided between training and use. AI providers can significantly reduce energy consumption during the training of AI through best practices. Researchers refer to these methods collectively as 4M, which stands for model, machine, mechanisation and map optimisation. These methods focus only on operating costs, meaning the energy costs of using AI hardware and data centres. They do not go into complex life cycle analysis, which also includes emissions from raw material production and the manufacture of all the components needed for hardware and data centres. [2], [3]
Back to top

In order to be sustainable, the overall handprint of AI should exceed its overall footprint, meaning its carbon dioxide emissions and other negative environmental impacts. As of yet, there are no studies proving with sufficient certainty that the benefits of AI will outweigh its negative impacts. Expectations are high that AI can be used to achieve social well-being, such as time savings in different jobs, but this should be done without placing an excessive strain on the environment. 

AI can be used to improve the efficiency of different processes and tasks in manufacturing and education, for example. In education, both principals and teachers benefit from specialised AI systems and generative AI. This may apply to everything ranging from the use of learning analytics to various translation applications.  

Technological development and research are prerequisites for achieving the ambitious climate targets set now and in the future.  Research report [5] provides information on what a smart wind farm of the future could look like. The proposed model is based on human-cyber interaction that transcends disciplines: this type of interaction is key whether we are talking about technology, medicine, finance or education.  

Another common feature across disciplines is that AI systems can prevent damage that can occur before people notice it. In a wind farm that is partly controlled and operated by AI, the AI can detect damage to e.g. wind turbines before humans, in which case the damage can be repaired at a significantly lower cost than before. In education, learning analytics or AI could similarly detect a learner’s difficulties in a specific subject, such as mathematics, at an early stage, allowing their knowledge gaps to be addressed sooner. From the learner's point of view, this could be significant not only in terms of learning outcomes, but well-being as well if they can at least partly avoid the anxiety that would be caused by not receiving support and constantly lagging behind in mathematics. Mathematical anxiety is a major global problem at both the individual and societal level.  

In order to be sustainable, the overall handprint of AI should exceed its overall footprint, meaning its carbon dioxide emissions and other negative environmental impacts. As of yet, there are no studies proving with sufficient certainty that the benefits of AI will outweigh its negative impacts. Expectations are high that AI can be used to achieve social well-being, such as time savings in different jobs, but this should be done without placing an excessive strain on the environment.  
AI can be used to improve the efficiency of different processes and tasks in manufacturing and education, for example. In education, both principals and teachers benefit from specialised AI systems and generative AI. This may apply to everything ranging from the use of learning analytics to various translation applications.  

Technological development and research are prerequisites for achieving the ambitious climate targets set now and in the future.  Research report [5] provides information on what a smart wind farm of the future could look like. The proposed model is based on human-cyber interaction that transcends disciplines: this type of interaction is key whether we are talking about technology, medicine, finance or education.  

Another common feature across disciplines is that AI systems can prevent damage that can occur before people notice it. In a wind farm that is partly controlled and operated by AI, the AI can detect damage to e.g. wind turbines before humans, in which case the damage can be repaired at a significantly lower cost than before. In education, learning analytics or AI could similarly detect a learner’s difficulties in a specific subject, such as mathematics, at an early stage, allowing their knowledge gaps to be addressed sooner. From the learner's point of view, this could be significant not only in terms of learning outcomes, but well-being as well if they can at least partly avoid the anxiety that would be caused by not receiving support and constantly lagging behind in mathematics. Mathematical anxiety is a major global problem at both the individual and societal level.  

Back to top

The utilisation of generative AI technologies offers both promising opportunities and significant challenges in education. It is also important to take into account factors such as social sustainability, which includes the well-being of both learners and teaching staff.

Interactive simulations and games that utilise AI can help learners understand sustainability challenges and test different solutions in practice. For example, climate change simulations created with the help of AI can visualise the long-term impacts of human activity on the environment. They can provide concrete examples and experiences for learners. Such simulations can be created quite easily on existing platforms, such as community-based online games.  

Language models can help collect and analyse large sets of environmental data, which can then be used to generate visual, informative and interactive presentations. Such visualisations can be valuable tools in teaching, helping to illustrate and outline complex and systemic environmental issues and promote learning. For example, educational institutions can collect meteorological data and data on their own electricity consumption to illustrate energy consumption.  

The capacity of language models to simplify and scale text to different skill levels is also important in terms of ecological sustainability. Coupled with translation applications, they can significantly improve access to information on the different themes of ecological sustainability and eliminate the need to acquire additional material for different languages and skill levels, for example. This is useful in groups with students from different language and cultural backgrounds. Tailoring and differentiating educational resources with language models can also help learners understand and internalise the goals of ecological sustainability and get different learners interested in the circular economy, for example.

The generation and use of AI data requires a great deal of energy and non-renewable natural resources globally, so the use of AI must always be justified and planned. Learners should be taught not to use AI to generate additional images for the sake of experimentation or generative AI for ordinary information searches, for example. AI can also be used in a wide range of ways to teach and learn about sustainability. It is also important to take into account the impacts of the use of AI on well-being. Learning should take place in varied pedagogically appropriate environments, avoiding harmful dependence on digital applications. At its best, a learning environment utilising AI can enrich teaching and learning also within the boundaries of sustainability. 

Back to top

Back to top
Updated