🤖 AI

From Perception to Action: Can UI Interventions Foster Sustainable LLM Chatbot

This paper demonstrates that sustainability-oriented UI interventions, such as mode switching and energy feedback, can effectively increase user awareness and foster energy-responsible behaviors in LLM chatbots without compromising usability, complementing backend efficiency efforts.

Original authors: Nitish Patkar, Pooja Rani, Jack Glässer, Simon Lüscher, Martin Kropp

Published 2026-06-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nitish Patkar, Pooja Rani, Jack Glässer, Simon Lüscher, Martin Kropp

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are driving a car. You know that driving fast uses more gas, and driving slowly saves gas. But what if your car's dashboard didn't tell you how much gas you were using? Worse, what if you didn't even realize that your car was burning fuel just by sitting there?

This is exactly the situation with AI chatbots (like the ones you talk to on your phone or computer). They are incredibly useful, but they "eat" a lot of electricity to think and answer your questions. Most experts have been trying to fix this by building better engines (more efficient computer chips) or better gas stations (greener data centers). But this paper asks a different question: Can we change the steering wheel and the dashboard to help you drive more efficiently?

Here is the story of how the researchers tried to do just that.

The Problem: We Think We Know, But We Don't

The researchers first asked 77 people, "Do you know how much electricity your AI chatbot uses?"

The Confusion: Most people said, "Yes, I know it uses energy!" But when they tried to guess how much, they were wildly wrong. It's like thinking a sip of water is the same as a whole swimming pool.
The Gap: People cared about the environment, but they weren't willing to slow down their car (accept a "dumber" AI) just to save a little gas. They wanted the best performance, even if it cost more energy.

The Experiment: Building a "Green Dashboard"

The team built a special version of a chatbot with five new features designed to nudge users toward saving energy, without making the chatbot harder to use. Think of these as new features on your car's dashboard:

The Three-Mode Switch (The Gear Shift): Instead of just one way to drive, you can choose:
- Eco Mode: Uses the least energy (like driving at a steady, slow speed).
- Balanced Mode: A mix of speed and efficiency.
- Performance Mode: The fastest, most powerful, but uses the most energy (like flooring the gas pedal).
The Energy Note (The Fuel Gauge): Every time the AI answers, it shows a tiny note saying, "This answer used about as much energy as charging your phone for 10 seconds."
The Prediction (The Trip Computer): Before you even hit "send," it estimates how much energy your question will cost.
The Dashboard (The Trip Log): A screen showing your total energy use over the week.
The Analogies (The "Real World" Translator): Instead of saying "0.002 Watts," it says, "This is like charging your iPhone once."

The Test: A Week of Driving

They invited 11 people to use this special chatbot for five days in their normal work. They wanted to see if these new dashboard features would change how people drove.

What Happened?

People Switched Gears: The participants used "Eco Mode" for 56% of their questions. They realized that for simple questions (like "What's the weather?"), they didn't need the "Performance Mode" gas-guzzler.
Huge Savings: Even though they only used Eco mode for about half their questions, the total energy saved was massive. If they had used "Performance Mode" for everything, they would have used 47% more energy.
The "Big Spender" Effect: Interestingly, the "Performance Mode" was used less than 25% of the time, but it was responsible for 89% of the total energy used. It's like driving 10 miles at 100 mph uses more gas than driving 100 miles at 30 mph.
No Shorter Questions: People didn't start asking shorter questions. They just switched to the "Eco Mode" gear when they didn't need the AI to be a genius.

The Takeaway: Awareness is Good, But Choice is Better

The study found that simply telling people "AI uses energy" isn't enough. People need easy choices.

The "Nudge" Worked: When the researchers gave people a clear, easy button to switch to "Eco Mode," people actually clicked it. They didn't need to be lectured; they just needed a convenient option.
The "Trade-off" Reality: People were happy to use Eco mode for simple tasks. But when they needed a really complex answer, they still wanted the "Performance Mode," even if it used more energy.
The Dashboard Helped: Seeing the energy cost in familiar terms (like "charging an iPhone") helped people understand the impact better than just seeing numbers.

The Bottom Line

You can't just fix the engine (the AI technology) to save the planet; you also need to teach the driver (the user) how to drive efficiently.

This paper proves that if we design chatbots with clear choices (like an Eco button) and simple feedback (like "this costs as much as a cup of coffee"), people will naturally start making greener choices. However, the researchers warn that this isn't a magic cure-all. It works best when combined with better technology on the backend. It's about giving users the tools to be part of the solution, rather than just passive passengers.

Technical Summary: From Perception to Action: Can UI Interventions Foster Sustainable LLM Chatbot

Problem Statement

Large Language Model (LLM) chatbots are increasingly integrated into daily workflows, raising significant sustainability concerns due to their energy consumption and carbon footprint. While the majority of mitigation strategies focus on backend optimizations—such as efficient model architectures, hardware accelerators, and data center improvements—the user interface (UI) layer remains underexplored. Current research largely overlooks how user interactions and UI interventions can influence cumulative energy consumption during everyday use. The authors posit that while users may express environmental concern, a gap often exists between this attitude and actual behavior, driven by invisible energy costs and decision-making frictions. The core problem is determining whether sustainability-oriented UI interventions can effectively increase energy awareness and steer users toward more computationally efficient interaction patterns without compromising usability.

Methodology

The study employed a three-stage methodology to investigate this problem:

Baseline Survey (RQ1): An online survey was conducted with 77 participants to assess current user awareness of LLM energy consumption, receptiveness to intervention concepts, and willingness to accept performance trade-offs. The survey utilized estimation-style questions (e.g., comparing LLM queries to Google searches or iPhone charging) to gauge mental models and knowledge gaps rather than technical accuracy.
Prototype Implementation: Guided by survey findings and principles of persuasive technology and choice architecture, the authors implemented a web-based LLM chatbot prototype featuring five specific UI interventions:
- Three-Mode Switch: Allows users to select between "Energy-efficient," "Balanced," and "Performance" modes.
- Energy Note: Displays a short energy estimate with every LLM reply.
- Prompt Prediction: Provides an energy estimate before the user sends a prompt.
- Metrics Dashboard: Summarizes energy use over recent activity.
- Energy Analogies: Reframes energy amounts into relatable everyday units (e.g., phone charging).
- Technical Note: Energy estimation was calculated using a token-based model ( $E = \alpha \cdot T_{in} + \beta \cdot T_{out} + \zeta$ ), with coefficients scaled according to specific model sizes (GPT-4o, GPT-4o-mini, GPT-4.1-nano) to approximate relative consumption.
Field Study (RQ2): A five-day field study was conducted with 11 frequent LLM users who utilized the prototype in their daily workflows. Data collection included daily awareness check-ins (5-point Likert scale), an end-of-week questionnaire, and system-level telemetry (token counts, mode selection, energy estimates). The study analyzed behavioral patterns, awareness trends, and energy impact, comparing observed behavior against a counterfactual scenario where all interactions occurred in "Performance" mode.

Key Contributions

The paper makes three primary contributions to the field of Human-Computer Interaction (HCI) and Green Software Engineering:

Survey Data: A baseline survey on end-user awareness and perception regarding the energy footprint of LLM chatbots, highlighting the discrepancy between perceived and actual energy costs.
Empirical Evidence: Data on how specific UI interventions influence energy-aware interaction patterns in LLM chatbots, demonstrating that design choices can shift routine behavior.
Replication Package: A publicly available package containing detailed survey responses, analysis scripts, and the implemented working LLM chatbot prototype to support further research.

Results

Baseline Survey Findings:

Awareness vs. Knowledge: While 94.8% of respondents reported some awareness of AI energy use, 88.3% misestimated actual consumption (e.g., significantly overestimating or underestimating the energy cost relative to a Google search or phone charge).
Attitude-Behavior Gap: Although environmental concern was high, only 39.0% of participants indicated a willingness to accept a performance trade-off for lower energy use.
Intervention Receptiveness: 68.8% favored an "Energy Efficient Mode," but willingness dropped when explicit performance trade-offs were introduced.

Field Study Findings:

Behavioral Shift: In the field study, the "Energy-efficient" mode accounted for 55.8% of logged prompts (155/278). Notably, 90.9% of participants self-reported actively choosing the Eco-mode when high accuracy was not required.
Energy Savings: The observed total estimated energy use was 164.72 Wh. Under a counterfactual scenario where all prompts were executed in "Performance" mode, the energy use would have been 309.44 Wh. This implies a 46.77% reduction in the energy footprint due to observed behavioral changes.
Mode Impact: Despite being used in fewer than 25% of prompts, "Performance" mode was responsible for approximately 89% of the total energy footprint.
Mechanism of Change: Participants did not significantly reduce prompt length. Instead, the primary behavioral mechanism was mode switching rather than prompt shortening.
Usability: All interventions scored above 4/5 in usability. The "Three-Mode Switch" and "Energy Note" were rated as the most influential interventions.
Awareness Trend: Self-reported energy awareness increased over the five-day period (from a mean of 3.27 to 4.44), driven by repeated, low-effort interactions.

Significance and Claims

The authors claim that sustainability-oriented UI interventions can successfully improve user awareness and support more energy-responsible interaction patterns in LLM chatbots. However, the paper maintains a modest stance on the scope of these findings:

Complementary Role: The authors explicitly state that UI interventions should be interpreted as complementary to backend and model-level optimizations, not as a substitute. UI design can shift routine, low-stakes behavior, but high-cost interactions (often requiring high accuracy) still necessitate backend efficiency.
Behavioral vs. Causal: The energy savings reported are model-based behavioral estimates (counterfactuals) rather than hardware-metered causal effects.
Addressing Decision Friction: The study suggests that unsustainable use is often driven by everyday decision-making frictions rather than a lack of environmental concern. By making trade-offs visible and easy to revise (e.g., via a Three-Mode Switch), interfaces can support informed decisions without imposing constraints.
Limitations: The authors acknowledge limitations regarding the small sample size (n=11 for the field study), the potential for rebound effects (users compensating for slower responses with more prompts), and the generalizability of findings given the technical background of the participants. They conclude that these results are "promising indications" that merit replication in larger, more diverse studies.