Survey of 1,080 Americans finds users get less blame for AI energy use than appliances

The shift went to developers (28.69 points for AI versus 17.72 for appliances) and policy-makers (18.92 versus 13.34). Vendors were treated about the same in both cases (25.76 versus 27.15, not a significant difference). The authors read this as responsibility spreading across many actors when computing happens in remote data centers.

When asked what they could personally do to save energy with generative AI, 44.9% of coded answers were about using it less, abstaining or switching to non-AI tools. Only 5.7% named choosing a provider, model or output type that uses less energy. The authors note that inference accounts for roughly 60% of AI's energy demand, so how people use these tools matters.

Data came from an online survey on the Prolific platform. Of 1,098 respondents, 18 were dropped for saying they had not answered seriously, leaving 1,080 aged 18 to 89 (mean age 45.94). Respondents were paid £1.24 for a median of nine minutes. Authors Michael Klesel and Uwe Messer framed the work with construal-level theory, which holds that things felt as psychologically distant are thought of more abstractly.

Awareness was tested with two drag-and-drop ranking tasks, each covering eight hours of use. The AI tasks were image generation, image captioning, speech recognition, question answering and text classification, ordered using Luccioni et al.'s per-task energy estimates (477 Wh for image generation versus 0.5 Wh for text classification). The household tasks ran from a space heater (10.6 kWh) down to a cable modem (0.048 kWh), using US Department of Energy figures. Correlations with the correct order were 0.37 for AI and 0.46 for household tasks.

Responsibility was measured with a modified dollar-metric scale: 100 points split among policy-makers, vendors, users, engineers and no one, asked separately for AI and for appliances. Differences for policy-makers, users, developers and 'no one' stayed significant after Bonferroni correction. The authors interpret the move away from users as consistent with a cloud-based service feeling remote.

Open-text answers on personal energy-saving strategies were coded into seven categories using a theory-derived framework and an LLM ensemble. Logistic regressions showed that men were more likely to suggest modality choice (B = 0.79), and that younger and more proficient respondents suggested model choice. People who used AI less often were more likely to say abstain. The authors call prompt optimization and model selection valuable but note most respondents did not think of them.