In the rapidly evolving world of technology, the sustainability of artificial intelligence (AI) systems is an important issue. These systems promise to revolutionize everything from energy management to medical diagnosis. But while we reap the benefits of this advanced technology, we also need to think about its impact on our environment.
First of all, we must not forget that our society has actively embraced AI technology. Issues of ethics and sustainable development are now arising with regard to artificial intelligence systems. How does the training of AI algorithms affect our environment, and what are the financial and environmental costs? What must be taken into account to achieve sustainable AI?
AI – a choice of society
Our society has decided to use AI technology. Artificial intelligence is not a natural phenomenon, but a technology that is consciously developed and used by humans. We can continue to develop and use this technology, but we do not have to. We should therefore choose carefully and decide that AI systems are used in a way that protects the values of our society, e.g. privacy, dignity, fairness, justice, sustainability. Artificial intelligence designed to advance our society must not lead us to sacrifice our values to technology.
Aimee van Wynsberghe (2021). Sustainable AI: AI for sustainability and the sustainability of AI. AI and Ethics, 1(3), 213–218. https://doi.org/10.1007/s43681-021-00043-6
AI Ethics
These thoughts have led to an ethical approach to AI technology. But AI is constantly evolving, and so the focus of the ethical approach has evolved with the technology.
Initially, AI ethics asked what an artificial intelligence could do. This led to fanciful scenarios of robot uprisings.
Then AI ethics addressed the practical problems of machine learning: the black box algorithm and the problem of accountability, as well as the rise of facial and emotion recognition systems that violate civil rights.
Now it is necessary to address the ecological consequences of AI technology. The new AI ethics now seeks to sensitize scientists, policymakers, AI developers, and the general public to the sustainability and environmental impact of AI.
AI ethics distinguishes two areas: the area of sustainability with the help of AI and the area of sustainability of AI.
The area of sustainability through AI is already somewhat more developed with the well-known non-profit organization „AI4Good“. Here, the goal is to achieve the United Nations Sustainable Development Goals (SDGs) with the help of AI applications and machine learning. Please also read the articles: AI Technologies: Working for Sustainability and Artificial Intelligence: Too hungry and too thirsty
The field – the sustainability of AI – is still quite young and requires detailed research. This includes the question: What is sustainable AI and how can it be linked to sustainable development?
What is sustainable development?
The United Nations Division for Sustainable Development Goals defines sustainable development as „development that meets the needs of the present without compromising the ability of future generations to meet their own needs“. Sustainable development therefore requires that the world’s resources be distributed fairly from one generation to the next without compromising their own innovative development. Sustainable development should balance the needs of the environment, the economy and society.
The Division for Sustainable Development Goals (DSDG) of the United Nations Department of Economic and Social Affairs (UNDESA) serves as the SGD Secretariat and provides concrete support and capacity building for the Sustainable Development Goals and related areas such as water, energy, climate, oceans, urbanization, transport, science and technology. The Sustainable Development Goals Division plays a key role in assessing the United Nations‘ system-wide implementation of the 2030 Agenda and in advocacy and public awareness on the Sustainable Development Goals.
About | Department of Economic and Social Affairs (un.org)
What is sustainable AI?
What does sustainable AI mean in this context? On the one hand, this topic covers the area of sustainability with the help of artificial intelligence; on the other hand, the area of sustainability of artificial intelligence itself needs to be discussed. It is about how artificial intelligence can be developed in a way that preserves environmental resources for us and future generations, while at the same time being compatible with the economic models for societies and the social values that are fundamental to a particular society.
Sustainable AI supports the environmental and social sustainability of AI products throughout the life cycle of an AI technology. It focuses on the entire socio-technical system of AI technology, not just AI applications.
Thus, on the one hand, sustainable AI is concerned with the hardware, the methods used to train the AI, the data processing, and the application of the AI. At the same time, there is also a need to address issues of sustainable development throughout the entire life cycle of an AI technology.
The impact of AI training on the environment
The environmental impact of AI training (and fine-tuning) is at the heart of AI sustainability. According to a scientific study, approximately 626,000 pounds (284,000 kilograms) of carbon dioxide are emitted when training a single deep learning model for natural language processing. This is roughly equivalent to the carbon emitted by five cars over their entire lifetime, or the carbon emitted by an average of 57 people for one year.
Strubell, E., Ganesh, A., & McCallum, A. (2019). Energy and Policy Considerations for Deep Learning in NLP. https://doi.org/10.48550/ARXIV.1906.02243.
Other studies have shown that Google’s AlphaGo Zero research generated about 96 tons of CO2 in 40 days of training, which is equivalent to 1000 hours of flying or the carbon footprint of 23 American households.
For a single, complete GPT-3 training session, scientists calculated a CO₂ impact of 552 tons of CO2 equivalent.
CO2-Ausstoß von ChatGPT und Co: Klimakiller künstliche Intelligenz? – Golem.de
The manufacture, use, and disposal of a 14-inch MacBook Pro generates 271 kilograms of carbon dioxide equivalent. A passenger flying from New York to San Francisco and back consumes between 500 kilograms and one ton of CO2, depending on the airline. The average emissions of all cars registered in Germany in 2022 were about 1.1 tons per year per vehicle over a mileage of 10,000 kilometers. (Kraftfahrt-Bundesamt – Monatliche Neuzulassungen – Neuzulassungsbarometer im Dezember 2022 (kba.de)).
Florian Zandt,( 22. Februar 2023). CO2-AUSSTOSS VON CHATGPT UND CO. Klimakiller künstliche Intelligenz? https://glm.io/171908.
The researchers also emphasize that fine-tuning – also known as „re-purposing“ or „refining“ – an AI model is more expensive than training a model from scratch.
Thus, training AI algorithms has both financial and environmental costs. The financial costs come from the hardware and power consumption or computing time in the cloud. The environmental cost is calculated as a CO2 footprint.
It is possible to generate some of the required energy from renewable sources or to offset it with carbon credits. However, this is currently only done in a few places. Unfortunately, renewable energy is not available in large quantities on a consistent and reliable basis.
Therefore, when studying the sustainability of AI technology, scientists are focusing on sustainable data sources, power supply, and infrastructure to measure and reduce the carbon footprint when training and/or fine-tuning an algorithm.
Carbon Tracker
This raises the question of how to measure the sustainability of the development and use of AI technology. This can be done, for example, by measuring the carbon footprint, the computing power used to train algorithms, etc.
Scientists have already developed techniques for tracking carbon emissions. There is an emissions calculator for machine learning and an experiment impact tracker.
Jens Gröger. Digitaler CO2-Fußabdruck Datensammlung zur Abschätzung von Herstellungsaufwand, Energieverbrauch und Nutzung digitaler Endgeräte und Dienste. Öko-Institut e.V. Berlin, 14. Juli 2020. Im Auftrag des Bund für Umwelt und Naturschutz Deutschland e.V. (BUND). Digitaler CO2-Fußabdruck (oeko.de)
Towards sustainable AI
When we look at AI technology, it is important to consider the impact that it will have on the environment. There are three important considerations to keep in mind on the path to sustainable AI technology.
First, we do not yet know very much about this technology. So far, we can think of the use of artificial intelligence as a kind of social experiment in which we still have a lot to learn.
We also need AI technology expert groups in governments. They should actively network with companies and public institutions and encourage them to report on the CO2 emissions generated by the training and optimization of AI systems. It is also important to support small and medium-sized enterprises that actively promote sustainable AI concepts.
Furthermore, as part of its regulatory options for AI, the European Commission should, for example, create a „proportionality framework“ to assess whether training or fine-tuning an AI model for a particular task is proportionate to its carbon footprint and overall environmental impact.
After all, with around 600 million people in the world without access to modern electricity, does it make sense to train AI models to beat the world champ at Go (AlphaGo) instead of bringing electricity to those homes?
Please find the German Translation of this article here: Künstliche Intelligenz: Fit für Nachhaltigkeit?