Special Issue “Towards the Sustainability of AI; Multi-Disciplinary Approaches to Investigate the Hidden Costs of AI”

Artificial Intelligence (AI) applications, i.e., applications of machine learning, deep learning, and other related technologies, are increasing at a rapid pace in our personal and professional lives. AI is used, for example, to predict fraud in the banking sector, benefitting both customer and company; as a decision support tool in healthcare, enhancing the efficiency of healthcare institutions; or to predict and mitigate natural disasters, protecting individuals in the surrounding area. The responsible use of AI may be of great benefit for humanity, from monitoring and repairing climate change destruction to uncovering new forms of disease and their respective treatments. Yet, despite the success and efficiency that AI promises, there are growing societal and ethical concerns that need to be addressed to prevent the design, development, and use of AI from creating new and exacerbating existing social and environmental injustices.

To date, the field of AI ethics has focused on uncovering and raising awareness of a host of issues in relation to AI including the potential loss of jobs, the quality of jobs available, privacy concerns about data collection, the embedding of cultural stereotypes and prejudices into AI models when using historical data to train these models, and the lack of transparency of decision rules generated by AI models, to name a few. While all of these are pressing issues, the issue of the sustainability of AI remains underexplored. Very recently, researchers have begun to uncover the environmental risks related to the materiality of AI [1,2,3,4]. These first publications focus strongly on energy consumption and on greenhouse gas emissions produced by training, tuning, and using AI systems. However, the environmental impact of AI does not stop there; the hardware used to run algorithms requires an industrial infrastructure of mining of natural resources (e.g., gold, tungsten), assembly of technical elements, cooling of technical infrastructures, and (electronic) waste disposal. Along this line of production, maintenance, and obsolescence, environmental risks arise that remain obscured and unquantified. To exacerbate these issues, AI is not only itself materially instantiated, it also changes the material conditions of the context in which it is employed, bringing efficiency gains, but potentially also creating rebound and ripple effects [5]. If AI is truly to succeed in making our world a better, more sustainable place, its full environmental and social impact must be uncovered. This is the objective of research on Sustainable AI.

Sustainable AI is a term that has recently garnered more and more attention, yet the understanding of what it means is still in development. In recent years, initiatives and articles predominantly addressed AI for sustainability, with few conferences [6,7], articles [2,8], and books [9,10] being dedicated to the sustainability of AI. It is this observation that has inspired the founding of the Sustainable AI Lab at the University of Bonn, Germany. Research towards the sustainability of AI is aimed at shining a light on the currently hidden sustainability issues and impacts of AI and, based on this knowledge, providing an ethically sound way forward for academics, policy makers, and industry alike. This requires an exploration of the concept of sustainability itself, its relation to AI, and a clear understanding of where in the AI design, development, implementation, and use process ethical issues can and should be addressed. Thus, Sustainable AI has an enormous task ahead to both raise awareness of possible harms as well as assist in their prevention.

A big step towards this goal is to re-assess our collective conception of AI. Generally, the term ‘AI’ denotes any and all computing technologies that emulate human cognitive abilities. Not only is this general definition ambiguous towards the terms ‘emulate’ and ‘human cognitive abilities’, it also invokes connotations of the digital, the virtual, the mental, and the immaterial [11]. If we are to think about AI in the context of sustainability, we ought to broaden our conception of what AI is and how it is embedded in the material world as a material object. The co-editors of this Special Issue, Aimee van Wynsberghe, Tijs Vandemeulebroucke, Larissa Bolte, and Jamila Nachid, hence urge readers to conceptualise AI as a world object [12]. Indeed, AI “[…] affects the world as a whole and not just a small corner of it” [13] (p. 5). This encompasses, first, the realisation that software cannot and should not be divorced from hardware. Every AI software requires hardware and a surrounding material infrastructure to run [10]. Second, this means realising that hardware does not exist as an isolated entity. Its existence is always dependent on complex global networks of production, supply, and use [8,9]. In the specific case of AI, these global networks are: socio-environmental, connecting labour force and raw materials to produce and distribute the technical components of the hardware running AI; material, providing the technical products and socio-technical relations that together produce AI; and digital, providing computational analysis.

We do not claim here to have exhaustively explored the materiality of AI technologies or to have given a definition that suits all contexts. Instead, we intend our conceptualisation of AI to act as an invitation to shift perspective and to facilitate discussion on the sustainability of AI.

This discussion is a multi-faceted one. For this reason, this collection of papers explores the issue of Sustainable AI from a variety of different angles. The first set of papers deals with diverse, fundamental ethical questions in relation to the sustainability of AI. In the vein of reconceptualising AI, Bolte, Vandemeulebroucke, and van Wynsberghe [14] argue that problems with current AI ethics guidelines are due to a conceptualisation of AI as isolated artefacts, which can be revised by conceptualising sustainability as a property of complex systems. Halsband [15] can be read as an elaboration on this topic, arguing for intergenerational justice as the normative core of the sustainability concept, while Robbins and van Wynsberghe [16] point out the consequences of infrastructural lock-in if the interconnectedness of AI with ecological, social, and economic systems is disregarded. However, AI can also help us achieve our visions for a more sustainable future, as shown by Bartman [17], who argues for the legitimacy of certain forms of AI-powered climate nudging.

The second set of papers focusses on sustainability frameworks for AI in diverse contexts. Continuing the theme of ethical groundwork, Genovesi and Mönig [18] connect sustainability to an ethics of responsibility and investigate how sustainability can be included in an ethical AI certification. Two papers propose frameworks based on the UN Sustainable Development Goals (SDGs): Gupta and Rhyner [19] strive to connect digitalisation with the SDGs, while Sætra [20] develops a framework for more comprehensive corporate reporting on the sustainability impact of AI. To make corporate impacts more visible to consumers, Kindylidi and Cabral [21] assess whether the current EU consumer protection framework is sufficient to promote sharing and substantiation of sustainability information to consumers.

Two of the papers in this collection present methodologies to relevant stakeholders. Ligozat et al. [22] point out the lack of attention on the negative sustainability impacts of AI for Green and introduce Life Cycle Assessment as a useful methodology for anyone developing AI for Green solutions. Raper et al. [23], with SDG 13 “Climate Action” in mind, present the notion of sustainability budgets addressed at software developers.

Getting clearer on how the relation between sustainability and AI is construed in the field, Samuel, Lucivero, and Somavilla [24] survey how stakeholders researching, governing, or working on the environmental impacts of digital technologies utilise different conceptions of ‘environmental sustainability’.

Finally, we close our collection with Bliek [25], who reviews successful sustainability applications of machine learning that use the technique of surrogate-based optimisation and gives recommendations to researchers who work on or apply that technique.

It is evident from this collection that Sustainable AI can only be tackled in an interdisciplinary manner. The question of ‘What is sustainable AI?’ must be approached from conceptual, ethical, political, sociological, empirical, technical, and many more perspectives. Ultimately, in asking this question, we ask how we envision our future with AI and how this vision may be blurred by leaving its material impact out of sight. We ask this question at a crucial time where climate and environmental worries accelerate while enthusiasm for AI runs high. The co-editors hence urge researchers to come together and work on the interrelations between these two grand developments of our time to find a sustainable path forward. With this collection of papers, we take a necessary step in this direction.