Good for the Planet, Bad for Me? Intended and Unintended Consequences of AI Energy Consumption Disclosure

Paper Title

Good for the Planet, Bad for Me? Intended and Unintended Consequences of AI Energy Consumption Disclosure

Publication Info

• Topic area: Sustainable Human-Computer Interaction (HCI) and AI energy consumption transparency.
• Keywords: AI energy consumption, energy disclosure, pro-environmental behavior, nudging, small language models, large language models, user perception, placebo effect, sustainable design, moral licensing.

Background and Problem

• Problem / challenge: The environmental impact of AI, particularly during inference, is substantial but lacks transparency. Users are unable to make informed decisions about energy-efficient AI models due to limited awareness of energy demands.
• Significance: Addressing AI’s energy consumption is critical for sustainability, as inference accounts for 60% of total energy use. Transparency can empower users to make environmentally informed choices, aligning with regulatory demands.
• Motivation and related work: Previous research highlights the environmental cost of AI training and inference but focuses primarily on technical solutions. Energy consumption disclosure (ECD) has proven effective in other domains (e.g., EU energy labels) but lacks empirical validation in AI contexts. This paper aims to fill this gap by exploring the behavioral and perceptual effects of ECD in AI model selection.

Solution

• Proposed approach: Energy Consumption Disclosure (ECD) as a choice architecture intervention to nudge users toward selecting energy-efficient small language models (SLMs) over large language models (LLMs).
• Novelty:
  1. Demonstrates the effectiveness of ECD in significantly increasing pro-environmental behavior (PEB) in AI model choice.
  2. Identifies the moderating role of pro-environmental attitude (PEA) on the effectiveness of ECD.
  3. Reveals unintended consequences of ECD, such as placebo effects reducing user satisfaction and perceived quality.
• Procedure and key techniques:
  • Conducted a one-factorial between-participants online experiment with 365 participants.
  • Participants chose between two AI models (SLM and LLM), with the treatment group receiving energy efficiency scores alongside performance ratings.
  • Both models were powered by the same underlying GPT-4o mini system to isolate psychological effects.
  • Measured outcomes included model choice, intensity of AI usage, perceived satisfaction, and perceived quality.

Results

• Concrete findings:
  • ECD increased the selection of the energy-efficient SLM from 4.7% in the control group to 39.3% in the treatment group (odds ratio = 12.89).
  • Participants with higher PEA were more influenced by ECD (interaction effect significant at p = .018).
  • SLM users reported lower satisfaction (Mdn = 5.33 vs. 6.00, p = .017) and perceived quality (Mdn = 5.00 vs. 5.33, p = .012) compared to LLM users.
• Advantage over baselines:
  • ECD demonstrated a stronger effect on model choice than pre-existing pro-environmental attitudes.
  • Contrary to modest effects typically reported in nudging literature, ECD yielded substantial behavioral changes.
• Experiments / evaluation:
  • Participants planned a 7-day vacation using their selected AI model.
  • Behavioral data (e.g., number of prompts, average tokens per prompt) and perceptual data (e.g., satisfaction, quality) were collected post-task.
  • Statistical analyses included chi-square tests, logistic regression, and Mann–Whitney U tests.
• Limitations and future work:
  • Limited generalizability to high-stakes scenarios, cross-cultural settings, and long-term effects.
  • Aggregated energy labels may not disentangle the effects of performance and energy cues.
  • Placebo effects on perception require further exploration to mitigate negative impacts.
  • Future research should investigate alternative visualizations, longitudinal effects, and broader environmental metrics (e.g., water consumption).

Summary

This study demonstrates that energy consumption disclosure (ECD) is a powerful tool for promoting pro-environmental behavior in AI model selection, increasing the likelihood of choosing energy-efficient small language models (SLMs) by over 12 times. However, it also reveals unintended consequences, such as reduced user satisfaction and perceived quality due to placebo effects. The findings highlight the need for careful design interventions to balance sustainability goals with user experience. Future research should explore alternative labeling formats, long-term impacts, and cross-cultural applicability to optimize ECD’s effectiveness.