While there is a general agreement on the environmental impact of GenAI, and a growing number of tools that try to bring awareness to this issue, the impact of such feedback on GenAI users remains unclear. To understand this impact, we developed a technology probe, EcoChat, as a Chrome extension for ChatGPT, that provides real-time data on inference energy and carbon emission use. We deployed the EcoChat probe in a four-week in-the-wild study with 52 participants supplemented with 18 interviews. Although usage patterns did not change significantly, participants reported increased awareness of GenAI inference’s environmental impact, and EcoChat encouraged them to reflect upon what constitutes ‘reasonable use’ in relation to the perceived value of their GenAI use. Interestingly however, participants also had difficulties in recalling specific data values, emission metaphors or even orders of magnitude of this impact. Accordingly, we identified a need for (a) more critical and transparent engagement with such estimated and uncertain data values; (b) more social and normative comparisons in eco-feedback interfaces to help end-users contextualize the data; as well as the opportunity for such interfaces to (c) provoke end users’ reflection and eventually support awareness, agency, and even contribute to systemic changes. Together, these findings unpack end user’s expectations and limitations to GenAI environmental impact, inform the design of future user-centered eco-feedback tools, and ultimately advance HCI debates on AI sustainability.
Ren et al. (Mon,) studied this question.