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

Hyeonwook Kim; Xuesi Chen; Alex Cabral; Cindy Kaiying Lin; Udit Gupta; Josiah Hester

Source-linked AI summary

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

Hyeonwook Kim, Xuesi Chen, Alex Cabral, Cindy Kaiying Lin, Udit Gupta, Josiah Hester

arXiv:2608.23968v1Aug 25, 2026cs.HCcs.CY

TL;DR

LLM use carries substantial environmental footprints, but how eco-feedback shapes students’ sustainability awareness and adoption decisions remains underexplored. This study evaluates a latency–carbon trade-off interface with 89 university students and finds that eco-mode preference declines sharply as perceived latency increases, while recognizing carbon savings increases selection.

Problem
Evidence about how eco-feedback shapes LLM users’ sustainability awareness and decision-making remains limited despite LLMs’ substantial and growing environmental footprint.
Method
Researchers designed an eco-feedback LLM interface visualizing latency–carbon trade-offs and evaluated it with 89 university students.
Results
Higher perceived latency reduced eco-mode preference (OR = 0.192, p < .001), while recognizing carbon-saving impacts increased eco-mode selection (OR = 3.879, p < .001).
Takeaways & Limitations
Eco-feedback can support lower-carbon LLM use and increase awareness when visualizations remain compatible with usability and clearly communicate environmental impacts.
Takeaways & Limitations
Uncertain carbon modeling and simplified data-center assumptions mean the reported latency–GHG relationship represents general trends rather than precise real-world performance.

Abstract

from arXiv · show

Large Language Models (LLMs) are increasingly being embedded into all facets of society, from search to education, industrial, and financial applications. These systems' 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. Eco-feedback interfaces offer a promising approach to encourage more sustainable behaviors, yet their role in shaping LLM users' sustainability awareness and decision-making remains underexplored. We design and deploy the interface that visualizes latency-carbon trade-offs during live LLM interactions. We study its use with undergraduate computer science students (N=89, ages 18-24), enrolled in a computing ethics course, providing an empirical look at how a technically sophisticated and values-oriented user population responds to sustainability-aware AI interfaces. We found that the likelihood of choosing the eco-feedback system significantly decreased as perceived response latency increased (p < .001), while users' willingness increased when they recognized the carbon-saving impacts (p < .01). Also, students with stronger eco-mindedness demonstrated higher baseline willingness to adopt lower-carbon modes and reported increased awareness of the environmental impacts of LLM use, though this effect diminished as latency increased. These results position eco-feedback interfaces as a promising sustainability intervention and highlight their potential as an educational opportunity to promote more sustainable LLM use among university students and beyond.

I. INTRODUCTION … B. Environmental Psychology and Personal Norms

LLM use creates substantial and growing environmental costs, while users often lack awareness of their individual impacts. This paper motivates eco-feedback interfaces for university CS students and examines how latency, environmental representations, and personal norms may shape sustainable LLM use.

I. INTRODUCTION: LLM use contributes to greenhouse-gas emissions, water consumption, and rare-earth mineral extraction, with electricity demand projected to nearly double by 2026.Training GPT-3 alone is estimated to have produced 552 metric tons of CO2 equivalent emissions.
I. INTRODUCTION: Many users remain unaware of the direct sustainability impacts of their individual LLM use, motivating investigation into how environmental information shapes interaction choices.The introduction identifies this uncertainty as a central problem for sustainability-aware LLM design.
A. Eco-feedback Interfaces: Eco-feedback provides visual information about sustainability impacts to encourage behavior change, but its design and application in LLM systems remain largely unexplored.Educational settings may be especially relevant because students are actively forming technology-related values, habits, and professional norms.
I. INTRODUCTION: Young adults aged 18–24 are relatively willing to engage in pro-environmental practices, while approximately 65% have used generative AI tools.CS students are positioned as both frequent LLM users and future AI-system creators, with LLMs increasingly embedded in university learning and programming.
I. INTRODUCTION: The study evaluates an eco-feedback LLM interface with 89 university students in the US, using service latency as a design lever for reducing carbon emissions.The introduction frames the study around eco-mode preferences, perceived usability, eco-mindedness, and willingness to adopt eco modes under latency conditions.
I. INTRODUCTION: Eco-mode selection fell from approximately 45% of interactions in eco mode 1 to below 5% in eco mode 5 as latency increased.Mixed-effects models found that perceived latency reduced eco-mode preference (p < .001), whereas recognizing carbon-saving impacts increased eco-mode selection under low-latency conditions.
A. Eco-feedback Interfaces: Sustainable HCI eco-feedback systems commonly use gamification, motivational techniques, and informational displays to reduce energy consumption and greenhouse-gas emissions.Typical displays include green coloring, eco-related icons, cost savings, and relatable emissions comparisons such as miles driven or light-bulb usage.
B. Environmental Psychology and Personal Norms: Environmental psychology identifies personal norms—an individual’s internal moral obligation—as a direct and powerful driver of proenvironmental behavior.Eco-feedback systems increasingly incorporate personal norms using information from user surveys or system telemetry.

C. Sustainable and Carbon-Aware Computing: Architecture, Systems, Design … 3) Estimating Latencies:

The paper extends carbon-aware computing for LLMs by designing eco-feedback modes that trade controlled latency increases within SLA constraints for reduced carbon emissions. The interface communicates these latency–carbon trade-offs through user-selectable modes whose impacts depend on backend configurations and latency estimation components.

C. Sustainable and Carbon-Aware Computing: Architecture, Systems, Design: Prior work targets renewable-powered data centers and energy-efficient LLM clusters; this work focuses on response-time latency within SLA constraints.The extension addresses carbon-aware computing specifically through latency-aware LLM operation.
III. LLM ECO-FEEDBACK INTERFACE: The interface includes eco modes that increase latency to reduce carbon emissions and design elements that inform users about carbon outputs.The interface is presented in two parts: eco-mode design and carbon-output feedback.
A. Eco Modes Design: Users can reduce carbon emissions by accepting controlled delays in response generation under latency-based SLA trade-offs.The eco modes are designed around latency profiles constrained by Service Level Agreements.
1) LLMs and Service Level Agreements:: LLM SLAs define expectations such as latency and availability, while latency includes TTFT, TBT, and end-to-end request-to-response time.TTFT captures initial responsiveness, TBT captures latency between tokens, and end-to-end latency captures total request-to-response time.
2) Defining Eco Modes:: Users choose among eco modes with distinct latency profiles and carbon impacts determined by backend hardware configurations.The interface simulates latency–emissions trade-offs through controlled latency degradation.
2) Defining Eco Modes:: Tensor parallelism uses fewer GPUs to reduce emissions while increasing processing time, and batching reduces per-request emissions while increasing GPU processing time and user latency.Tensor parallelism affects both TTFT and TBT; batching combines incoming requests to reduce emissions for each request.
2) Defining Eco Modes:: Five eco-mode configurations were designed from measurements across 1, 2, 4, and 8 GPUs and batch sizes from 1 to 256, with increasing latency degradation and descending embodied carbon emission.The default baseline uses 4 GPUs, batch size 1, and no renewable energy scheduling for google/gemma-2-27b-it.
3) Estimating Latencies:: End-to-end latency tETE equals tTTFT plus tTBT multiplied by output tokens Tout, with an additional 150 ms overhead when renewable energy scheduling is used.The latency ratios are summarized from simulated input and output token sizes, GPU counts, and batch sizes.

4) Estimating Carbon Emissions:

This section estimates per-query carbon emissions using eco-mode configurations and decomposes total emissions into embodied and operational components. Across eco modes, higher latency corresponds to monotonically lower emissions, revealing a latency–emissions trade-off.

4) Estimating Carbon Emissions:: Carbon emissions per query are calculated from equations based on the eco-mode configurations.The passage introduces the equations as the basis for estimating query-level emissions.
4) Estimating Carbon Emissions:: As eco mode increases, latency rises while emissions decrease monotonically.The trade-off is observed across small, medium, and large output-token requests; larger requests have higher latency and variability.
4) Estimating Carbon Emissions:: CEtotal is the total carbon emission per query, decomposed into CEemb and CEop.CEemb represents amortized hardware manufacturing and deployment impacts, while CEop represents electricity-related inference emissions.

5) Latency and GHG Emission Evaluation across Eco Mode: … B. Experimental Procedure

The study evaluates latency and greenhouse-gas emissions across eco modes while examining an interface that makes sustainability trade-offs visible during live LLM use. It uses a functional system with undergraduate computer science students who encountered mode implications through direct interaction rather than advance explanation.

5) Latency and GHG Emission Evaluation across Eco Mode:: Latency and emissions were plotted across eco modes using TTFT, average TBT, end-to-end latency, and GHG emissions.End-to-end latency and emissions were evaluated after accounting for their dependence on input and output token lengths.
5) Latency and GHG Emission Evaluation across Eco Mode:: Small requests had the lowest, tightest latency distributions, whereas large requests were slowest and most variable, with differences widening at higher eco modes.End-to-end latency reached up to 252 seconds, especially for large requests.
B. Interface Design: The interface offered a default mode and five eco modes with real-time latency and estimated greenhouse-gas feedback.Four visualizations presented detailed metrics, relative savings, real-world impact equivalents, and latency–emissions trade-offs.
B. Interface Design: Annualized impact feedback assumed 1,217 queries per year and expressed savings through PET bottles, gasoline-car travel, and light-bulb usage.Annualized values were used because per-query savings were considered too small to interpret meaningfully.
A. Participants: The study recruited 102 undergraduate computer science students and retained 89 valid participants aged 18–24 for analysis.Participants were enrolled in a mandatory Computer Ethics course and were U.S.-based.
B. Experimental Procedure: Over three days and ten course sections, participants used personal laptops, individual credentials, and a functional LLM system supported by teaching assistants.The interface presented system configurations through a dropdown, while the specific latency and emissions implications of each mode were not explained beforehand.
B. Experimental Procedure: Participants learned the magnitude of latency–emission trade-offs through direct interaction and feedback during use rather than advance instruction.This procedure exposed mode implications within the functional eco-feedback interface.

1) Individual Survey: … A. User Preference and Satisfaction of Eco Modes

The study combined structured surveys, group discussion, and mixed qualitative–quantitative analysis to examine eco-mode preferences, satisfaction, and environmental feedback. Eco-mode selection and satisfaction declined as latency increased, with the lowest-latency mode receiving the strongest evaluations.

1) Individual Survey:: Participants freely explored the interface, completed randomized interaction tasks, and then answered an online survey combining Likert, open-ended, and ranking questions.The survey assessed interface understanding, user experience, and how effectively eco-feedback communicated environmental impact.
1) Individual Survey:: For each of five eco modes, participants compared default and eco modes and rated satisfaction, latency, carbon savings, performance differences, future use, and environmental norms.They also evaluated four eco-feedback visualizations for attention, satisfaction, and perceived impact using a validated 10-item Likert-scale instrument for personal environmental norms.
2) Group Discussion:: Group discussions elicited qualitative reflections on usability, eco-feedback interpretation, and design suggestions that complemented structured survey responses.Discussions were audio-recorded with consent to capture motivations and interpretations not evident in the survey.
C. Data Analysis: 85 initial codes were consolidated into 36 subthemes and four higher-level themes through reflexive thematic analysis of open-ended survey and discussion data.The first author led coding and codebook development, which three additional researchers iteratively reviewed; disagreements were resolved through discussion.
C. Data Analysis: Quantitative analyses in R examined eco-mode preference in relation to perceived latency, carbon-saving impact, default-mode differences, and environmental norms.Likert responses were treated as ordered indicators, while satisfaction and ranking items were reported descriptively.
C. Data Analysis: Eco-mode preference was modeled with a logit-link GLMM using participant-level random intercepts and fixed effects for mode, carbon-saving impact, latency, performance differences, and personal norms.Results were reported as odds ratios with 95% confidence intervals and p-values.
A. User Preference and Satisfaction of Eco Modes: Eco-mode selection decreased sharply from approximately 45% in Mode 1 to below 5% in Mode 5.Mode 1 had the lowest latency, whereas Mode 5 had the highest latency; the pattern was supported by mixed-effects logistic regression.
A. User Preference and Satisfaction of Eco Modes: Mean satisfaction declined from 3.64 in Mode 1 to 2.06 in Mode 5, while Mode 5 showed the greatest dispersion (SD = 1.03).Modes 2 and 3 had means of 3.03 and 2.76, whereas Modes 4 and 5 had means of 2.35 and 2.06; most users were dissatisfied with highly constrained modes.

1) Impact of Latency, Carbon-Saving, and performance difference: … B. Effectiveness of Eco-Feedback Visualizations

Latency reduced eco-mode preference, whereas recognizing carbon savings increased selection, and eco-minded students showed stronger baseline preference for eco modes. Despite high attention and satisfaction with eco-feedback visualizations, most participants remained reluctant to adopt the system regularly.

1) Impact of Latency, Carbon-Saving, and performance difference:: OR = 0.192, p < .001: higher perceived latency was strongly associated with reduced eco-mode preference.Carbon-saving recognition increased eco preference when modeled alone (OR = 3.879, p < .001).
1) Impact of Latency, Carbon-Saving, and performance difference:: Mode 4: OR = 0.136, p = .010; mode 5: OR = 0.131, p = .028: highest-latency configurations remained less likely than mode 1 in the combined model.The model accounted for carbon-savings recognition, latency, and performance differences.
2) Influence of Eco-mindedness:: Medium: OR = 8.95, p = .046; High: OR = 50.34, p = .006: eco-minded participants were more likely than low eco-minded participants to select eco modes.The comparison used personal-norm groups and a mixed-effects model; participant-level random intercepts accounted for repeated measures.
2) Influence of Eco-mindedness:: Only system modes with statistically significant effects were reported in the combined model after accounting for carbon savings, latency, and performance differences.Models included participant random intercepts.
3) Likelihood of Long-Term Adoption:: 48.3% somewhat unlikely and 16.9% extremely unlikely: most participants were reluctant to adopt the eco-feedback system for regular use.Among 89 participants, 5.6% were very likely, 16.9% somewhat likely, and 12.4% neither likely nor unlikely to use it.
B. Effectiveness of Eco-Feedback Visualizations: 80.9% reported paying at least some attention to the displayed eco-feedback information.Attention, satisfaction, and preferences were analyzed across four visualization formats using Likert-scale responses and qualitative feedback.
B. Effectiveness of Eco-Feedback Visualizations: Interpretive, outcome-oriented visualizations were better received than purely descriptive metrics, while familiar real-world carbon analogies were preferred to abstract units.Satisfaction was generally high across all four formats.

C. User Responses and Learning Effects … 1) Sustainability Within Acceptable Carbon Saving:

The eco-feedback interface helped students recognize and learn about the environmental impacts of everyday LLM use, but latency constrained sustained adoption. Eco mode 1 offered measurable carbon savings, while interface support and clearer benefits could improve tolerance for lower-carbon interaction.

C. User Responses and Learning Effects: Participants viewed the eco-feedback interface as both an interaction aid and a source of new understanding about environmental impact.Several participants immediately recognized the system’s sustainability focus.
C. User Responses and Learning Effects: Participants reported confusing technical terms and metrics, while the interface helped them learn that everyday chats produce carbon emissions.They also learned about the trade-off between response speed and environmental sustainability.
C. User Responses and Learning Effects: Higher eco modes were often described as unusable because response latency dominated participants’ judgments of speed and usability.Comparisons with existing language-model tools shaped expectations and sometimes created uncertainty about differences between eco modes.
C. User Responses and Learning Effects: The interface exposed gaps in digital-sustainability understanding and supported learning about environmental impacts within clear usability limits.This finding links interaction with the system to both sustainability awareness and constrained adoption.
VI. DISCUSSION: The discussion addresses implications for sustainable eco-feedback design in LLM systems, alongside study limitations and future directions.This frames the reported findings as design-relevant rather than solely descriptive.
1) Sustainability Within Acceptable Carbon Saving:: 12 g CO2 per query and 14.6 kg CO2 annually per user were reduced by eco mode 1, which had a 43% selection rate.The annual estimate assumes 1,217 queries per year, indicating meaningful carbon-saving potential when lower-impact modes remain acceptable.
1) Sustainability Within Acceptable Carbon Saving:: Perceived latency remained the primary barrier to sustained eco-mode use, making wait-time management and clearer carbon-saving communication important design priorities.Participants also shortened prompts or changed interaction patterns under delay, so visible system status and anticipatory feedback can help preserve user agency.

2) Eco-Feedback as a Learning-Oriented Interaction: … 1) Backend Systems Changes that Influence Design:

Eco-feedback can make the environmental implications of LLM use more understandable through situated, interpretable interaction, but sustained adoption depends on clear communication, personal norms, and manageable latency. Backend configurations currently create coarse latency–carbon trade-offs, motivating smoother and more realistic system designs.

2) Eco-Feedback as a Learning-Oriented Interaction:: Eco-feedback made otherwise invisible aspects of digital sustainability more legible during routine LLM interaction.Concrete feedback supported reflection on how prompt length and mode selection relate to environmental impact.
2) Eco-Feedback as a Learning-Oriented Interaction:: Eco-feedback interfaces function as lightweight, experiential learning touchpoints embedded in everyday AI use.Learning emerges through situated, feedback-driven interaction rather than relying solely on explicit instruction.
1) Eco-Feedback Interpretability:: Visualizations using concrete, everyday equivalents were perceived as more effective than abstract emissions metrics, although some participants misunderstood common symbols and carbon units.Leaf icons and savings units did not always convey the scale of environmental impacts.
1) Eco-Feedback Interpretability:: Eco-feedback should ground information and visuals in users’ everyday experiences and explain eco modes, carbon reduction, and available interface information through onboarding.Participants struggled with latency and trade-off metrics when presented without sufficient context.
2) Personal Norms and Personalization:: Greater eco-mindedness was associated with stronger eco-mode preference in eco mode 1, but this effect weakened substantially in higher modes as latency and interaction friction increased.Environmentally motivated users also reduced eco-mode adoption when delays became more pronounced.
2) Personal Norms and Personalization:: Personal norms can identify initial willingness to engage with lower-carbon modes but are insufficient to sustain adoption under high-delay conditions.Potential strategies include low-latency eco defaults, gentle prompts when time pressure is low, and framing eco mode 1 as a recommended baseline.
1) Backend Systems Changes that Influence Design:: The study’s fixed four-GPU, batch-size-one backend and lack of renewable energy scheduling produced coarse, discrete latency shifts unlike complex production systems.Large latency gaps contributed to a large drop-off in eco-mode preference between eco modes 1 and 2.
1) Backend Systems Changes that Influence Design:: Future systems could vary prefill chunk sizes, GPU frequencies, and query scheduling policies to create smoother latency transitions and more realistic latency–carbon trade-offs.Such evaluations could identify more precise latency thresholds acceptable to different users.

2) Uncertainty in Carbon Modeling and Limitations in Impact Captured: … VII. CONCLUSION

The study’s latency–GHG findings and carbon-reduction estimates require cautious interpretation because carbon modeling, sampling, environmental coverage, and interface design impose important limitations. The conclusion frames eco-feedback as supporting lower-carbon LLM use only within usability-compatible bounds.

2) Uncertainty in Carbon Modeling and Limitations in Impact Captured:: Carbon modeling is uncertain because supply-chain data, hardware lifecycles, cloud-provider transparency, and production data-center resource management are simplified.The assumed correspondence between batch size, GPU allocation, and renewable-energy scheduling does not capture complex LLM-serving realities.
D. Research Limitations and Future Work: Carbon-reduction estimates should be interpreted cautiously, with eco-feedback emphasizing relative trade-offs and transparent communication rather than precise numerical claims.Variable energy sources and system-level optimizations limit the precision of inference-time carbon accounting, while overstated savings risk greenwashing.
1) Sample and Demographic Scope:: The sample comprised undergraduate computer science students at a single U.S. university, which may limit generalizability and introduce sampling bias.Participants may have been more technically informed and environmentally attentive than broader LLM-user populations.
2) Environmental Metrics Beyond Carbon:: Future work should integrate water consumption and human health impacts to provide a more holistic view of environmental trade-offs.Broader environmental metrics could engage audiences concerned with sustainability outcomes beyond carbon accounting.
3) Interface Design Restrictions:: Limited interaction controls, including absent stop or scroll functionality, increased perceived waiting costs and reduced usability under higher latency.Multiple environmental equivalence formats, including PET bottles, gasoline car travel, and light bulb usage, may also have varied participants’ interpretations.
VII. CONCLUSION: The interface used service latency to reduce digital carbon emissions and examined how latency, eco-feedback design, and environmental norms shaped eco-mode preference, usability, and anticipated adoption.The study included 89 participants and found that lower-carbon LLM use was supported only within usability-compatible bounds.

APPENDIX A USER STUDY MATERIALS · A. Eco-Mode Preference and Trade-Off Questions · B. Eco-Feedback Visualization Evaluation

Appendix A describes the survey materials used in the study, combining Likert-scale, ranking, and open-ended questions. Participants evaluated eco-mode trade-offs and eco-feedback visualizations through preference, perception, satisfaction, attention, understanding, influence, and ranking measures.

APPENDIX A USER STUDY MATERIALS: The survey included Likert-scale, ranking, and open-ended questions.
A. Eco-Mode Preference and Trade-Off Questions: For each Eco Modes 1–5, participants compared the eco mode with the default mode.
A. Eco-Mode Preference and Trade-Off Questions: Participants reported which experience they preferred: the default mode or Eco mode X.
A. Eco-Mode Preference and Trade-Off Questions: They rated how different Eco mode X felt from the default mode.
A. Eco-Mode Preference and Trade-Off Questions: Trade-off questions assessed how carbon savings, delayed response, and perceived performance differences affected preference.
A. Eco-Mode Preference and Trade-Off Questions: Participants rated their satisfaction with the Eco mode X experience.
B. Eco-Feedback Visualization Evaluation: Participants evaluated right-side eco-feedback information, including descriptive metrics, You Saved, Your Impact, and the Eco Mode Tradeoffs graph.
B. Eco-Feedback Visualization Evaluation: For the visualizations, they reported attention, understanding, satisfaction, influence, and which display had the biggest impact on their experience.The ranked displays were descriptive metrics for each response, You Saved, Your Impact, and the Eco Mode Tradeoffs graph.

C. Future Use and Reflection · D. Personal Environmental Norms · E. LLM Familiarity

The survey assessed students’ future willingness to use the eco-feedback interface, reflections on usability, personal environmental norms, and prior familiarity and use of LLMs. These measures covered intended adoption, interpretation, sustainability attitudes, and everyday LLM practices.

C. Future Use and Reflection: Students reported how likely they would be to use the eco-feedback LLM service daily.This item assessed future willingness to adopt the interface.
C. Future Use and Reflection: Students reflected on whether anything about the eco-feedback interface was confusing or difficult to use.The open-ended question invited explanations of usability problems.
D. Personal Environmental Norms: Personal environmental norms were measured with a validated 10-item Likert-scale instrument covering sustainability-related attitudes and everyday actions.The instrument assessed personal environmental norms and routine sustainability behaviors.
E. LLM Familiarity: Background questions assessed participants’ familiarity with LLMs such as ChatGPT, Claude, and Gemini.This captured prior experience with commonly used LLM systems.
E. LLM Familiarity: Participants reported how often they used LLM tools.Usage frequency was included as a background measure of LLM experience.
E. LLM Familiarity: Participants identified their primary purposes for using LLMs by selecting all applicable options.The survey captured the contexts in which students mostly used LLMs.