When LLMs Slow Down: How Environmental Impacts Mediate University Students' LLM Usage

Synthesis: Large language models' carbon and water footprints raise important Sustainability concerns, particularly with adoption rates exceeding 80% among university students despite limited insight into the environmental impacts of individual usage. Kim and colleagues design and deploy an eco-feedback interface that visualizes latency-carbon trade-offs during live Large Language Models (LLMs) interactions, studying its use with 89 undergraduate computer science students in a computing Ethics course. They find that the likelihood of choosing the eco-feedback system is significantly shaped by sustainability awareness, providing an empirical look at how a technically sophisticated and values-oriented user population responds to sustainability-aware AI interfaces.

Key Findings

1. LLM adoption exceeds 80% among university students, yet there is limited insight into the environmental impacts of individual usage.
2. An eco-feedback interface visualizing latency-carbon trade-offs during live LLM interactions was designed and deployed.
3. The interface was studied with 89 undergraduate computer science students (ages 18-24) in a computing ethics course.
4. The likelihood of choosing the eco-feedback system is significantly shaped by sustainability awareness.
5. The study provides empirical evidence on how a technically sophisticated, values-oriented user population responds to sustainability-aware AI interfaces.

What this means for practice

• Instructors. Make the latency–carbon trade-off explicit in assignment design: students selected eco mode in roughly 45% of interactions at low latency but under 5% at high latency, so leave room for the slower, lower-carbon option within deadlines.
• Instructors. Teach sustainability awareness directly, since recognizing the carbon-saving impact significantly increased eco-mode selection under low-latency conditions (p < .001).
• Designers. Show impact in concrete, outcome-oriented terms (PET bottles, gasoline car travel, light bulbs) rather than abstract carbon metrics, which participants rated more favorably.
• Designers. Keep perceived latency low and add the interaction controls users expect; higher perceived latency significantly reduced eco-mode preference (p < .001), and the absence of stop or scroll functionality compounded the perceived cost of waiting.
• Researchers. Extend eco-feedback beyond carbon to water consumption and human health impacts, and test it with populations beyond technically informed computer science students.

Limitations

• Single-site sample: 89 undergraduate computer science students at one U.S. university, recruited from a mandatory Computer Ethics course, so participants were likely more technically informed and environmentally attentive than broader LLM user populations, and the course context may have encouraged socially desirable responses.
• No demographic attributes (gender, race, nationality) were collected, limiting analysis of how these factors shape eco-feedback interpretation and engagement.
• Interface design restrictions (no stop or scroll functionality, and several equivalence formats for environmental impact) likely influenced satisfaction under higher-latency conditions.
• Carbon modeling for computing systems carries inherent uncertainty, so the authors frame the results as communication rather than precise numerical claims.

Citation

Kim, Chen, Cabral, Lin, Gupta, & Hester (2026). When LLMs Slow Down: How Environmental Impacts Mediate University Students' LLM Usage.