The training and fine-tuning of Machine Learning (ML) models is costly to the environment in terms of computational power (energy) and ultimately carbon emissions (Bender et al 2021, Schwartz et al 2020, Strubell et al 2019). For example, roughly 15% of Google’s energy consumption in 2021 was dedicated to AI tasks (Patterson et al 2022). With the ongoing increase in Machine Learning (ML) in both model numbers as well as their achieved complexity, its environmental impact is adding up to be significant. While the ML and HCI communities has been identifying such critical environmental issues, there is still a lack of comprehensive methodologies to support system designers and developers in taking a sustainability-by-design approach where they understand, measure, and mitigate such issues. For instance, the few existing tools that calculate carbon emission costs (Anthony et al 2020, Lacoste et al 2019, Lannelongue 2021) largely focus on technical measurements, largely overlooking if and how they are used by practitioners in the field.
The examination of eco-feedback technology’s impact on behaviour change has a long history in Human-Computer Interaction (HCI) and Visualization research. Such research has often focused on domestic settings for instance understanding how people react to smart meter displays or custom energy visualisations (e.g., Costanza et al 2012). Parallel to the state-of-the-art in ML carbon tracking, early such work also was centred around the tools themselves, only later maturing to appreciate the more complex socio-technical issues that define if and how eco-feedback impacts behaviour. For instance, early studies documented a correlation between eco-feedback and feelings of guilt (Prost et al. 2015), yet more recent work has since focused on empowering messages such as the encouragement of collective action (Panagiotidou et al 2023; 2024, Bird and Rogers 2010).
This PhD will draw from research on sustainable HCI, behaviour-change and critical data visualisation to examine how to best communicate ML’s potential environmental effects. Specifically the student will investigate current sustainability-oriented practices within ML teams and strive towards developing a design toolkit for supporting more sustainable-oriented decision-making. The project may include activities such as conducting interviews and observations, running co-design workshops as well as developing visualisations of carbon emissions of machine learning models to act as a tools meant for eco-feedback.