Bear-ing the Burden: Impact of Affective Eco-Feedback on Sustainable LLM Use

The increasing use of large language models (LLMs) has raised growing concerns about the environmental impact of user-driven inference, while end-users often remain unaware of their carbon footprint. This paper investigates whether eco-feedback integrated into LLM interfaces can influence prompting behaviour. We conducted a mixed-method user study (N = 30) using a repetitive text-generation task across three conditions: a standard interface, rational eco-feedback providing quantitative resource metrics, and affective eco-feedback delivered via an embodied bear character. Results show that while both feedback types heightened environmental awareness, only affective eco-feedback significantly reduced excessive prompting compared to the control condition. These findings suggest that continuous, self-attributable signal during interaction can foster more mindful LLM usage, highlighting the potential of human-centered interface design as a complementary approach to sustainable AI.