🤖 AI

Nudging Sustainable Choices through LLM-Generated Recommendation Explanations

This study demonstrates that while merely disclosing sustainability information in LLM-generated recommendation explanations fails to alter user choices, framing such information or invoking descriptive social norms effectively nudges consumers toward sustainable selections across both low- and high-involvement domains.

Original authors: Haya Halimeh, Dietmar Jannach, Oliver Müller

Published 2026-07-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Haya Halimeh, Dietmar Jannach, Oliver Müller

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 walking through a giant, digital supermarket where a helpful robot assistant is trying to guess what you want to buy. This robot is a "recommender system," a piece of software that suggests items based on what you've liked before. Now, imagine the world is trying to get everyone to make choices that are better for the planet, like picking a coffee that uses less plastic or a hotel that saves energy. Scientists are asking: Can we teach this robot to not just find what you like, but to gently nudge you toward what is good for the Earth without forcing you?

To do this, the researchers use two main ideas. First, they use "explanations." Instead of just saying, "Here is a coffee," the robot says, "Here is a coffee because you like dark roast." This is like a waiter telling you why a dish is good. Second, they use "nudges." A nudge isn't a shove; it's a subtle change in how information is presented that makes the right choice feel easier or more appealing, like putting fruit at eye level in a cafeteria. The big question is: If we tell the robot to explain why an item is eco-friendly, will people actually buy it? And does it matter how the robot says it?

This paper dives into that question by turning the robot's explanation into a super-smart, AI-powered storyteller. The researchers, Haya Halimeh, Dietmar Jannach, and Oliver Müller, wanted to see if they could use Large Language Models (LLMs)—the same kind of AI that writes essays and chats with you—to generate these eco-friendly explanations on the fly. They tested this in two very different worlds: buying a cheap cup of instant coffee (a quick, low-stakes decision) and booking a hotel for a vacation (a big, expensive, high-stakes decision).

The team set up a game where 529 real people had to choose between four items that all matched their personal tastes perfectly. Some items were "greener" (like having recyclable packaging or a green certification), and some were not. The twist was in the explanation the AI gave for each item. They tested four different ways of talking about the green features:

Just the facts: "This item is recommended because it has recyclable packaging."
The "Good News" frame: "This item is recommended because its recyclable packaging helps reduce landfill waste and supports a cleaner future." (Focusing on the positive benefit).
The "Everyone Else" nudge: "This item is recommended because many other customers who looked at similar products chose it for its recyclable packaging." (Focusing on what peers do).
No green info: Just explaining why it fits their taste, with no mention of the environment.

Here is what they found, and it's a bit of a plot twist. Simply telling people the green facts (the "Just the facts" group) didn't change their minds at all. People still picked the non-green items just as often as before. The AI didn't magically make them care just by listing the data.

However, when the AI used the "Good News" frame or the "Everyone Else" nudge, the results were dramatic. In these groups, the number of people choosing the sustainable option jumped from about 51% (basically a coin flip) to around 69% and 67%, respectively. The researchers found that it wasn't just having the information that mattered; it was how the information was wrapped up. The "Good News" frame, which highlighted the positive consequence of the choice, was the most powerful, making the decision feel easier and more satisfying.

Interestingly, the researchers noticed a funny gap between what people said and what they did. When people saw the plain facts, they thought the explanations were helpful and clear. But that didn't translate into buying the green item. It was only when the explanation was framed as a benefit or a social trend that people actually changed their behavior. This suggests that knowing something is "good" isn't enough; you need to feel the positive impact or see that others are doing it to actually make the switch.

The study also checked if this worked for both quick decisions (coffee) and big decisions (hotels). The answer was yes: the same rules applied to both. Whether you are spending $5 or $500, a well-framed explanation can nudge you toward a greener choice without taking away your freedom to choose. The researchers conclude that using AI to craft these specific, theory-backed stories is a powerful tool for encouraging sustainable habits, but it only works if you tell the story right. Just listing the facts isn't enough to move the needle.

Technical Summary: Nudging Sustainable Choices through LLM-Generated Recommendation Explanations

Problem Statement
Recommender systems (RS) significantly influence global consumption patterns, presenting a critical opportunity to steer users toward sustainable choices aligned with the UN Sustainable Development Goals. While prior research suggests that explanations can improve decision-making, existing sustainability-aware explanations are largely static, template-based, and limited in their ability to personalize content. Furthermore, empirical evidence regarding the effectiveness of such explanations in personalized settings is thin. A key gap exists in understanding whether merely disclosing sustainability information is sufficient to change behavior, or if the presentation of that information (via behavioral nudging mechanisms) is required. Additionally, it remains unclear whether these effects generalize across domains with varying levels of user involvement (e.g., low-stakes daily purchases vs. high-stakes travel bookings).

Methodology
The authors conducted two randomized user studies ( $N=529$ ) across two distinct domains: a low-involvement domain (instant coffee, $n=262$ ) and a high-involvement domain (hotel bookings, $n=267$ ).

Experimental Design: The studies employed a between-subjects design with four conditions. Participants first elicited explicit preferences (e.g., roast level, hotel facilities) which were used to generate preference-matched recommendations. Each participant was shown four items (two high-sustainability, two low-sustainability) accompanied by LLM-generated explanations.
- Preference-only Baseline: Explanations described the match between user preferences and item attributes, omitting sustainability content.
- Mechanism-free: Explanations disclosed the sustainability attribute (e.g., recyclable packaging, Green Key certification) factually without behavioral framing.
- Framing: Explanations presented the sustainability attribute using a positive, gain-oriented frame (foregrounding beneficial consequences).
- Descriptive Social Norms: Explanations presented the sustainability attribute by citing the behavior of similar peers (e.g., "many guests have chosen this...").
Technical Implementation:
- Recommender: A preference-based, knowledge-based recommender scored items against user-stated features, ensuring that high- and low-sustainability items were matched on preference fit to isolate the effect of the explanation.
- LLM Generation: Explanations were generated on-the-fly using Claude Sonnet 4.6. The system utilized a prompt-engineering approach where the core structure (two sentences, <50 words) was fixed, but the second sentence was dynamically instantiated based on the assigned condition (Table 1).
- Validation: A rigorous audit pipeline ensured the quality of the stimuli. This included statistical analysis of text properties (length, complexity), an "LLM-as-a-judge" evaluation to verify adherence to the intended mechanism and factual fidelity (99.2% support rate), and human validation of a stratified subset.
Analysis: The authors used binary logistic regression to analyze choice behavior (selecting a high-sustainability item) and linear regression for subjective evaluations (satisfaction, difficulty, clarity, usability, sustainability impact). Models controlled for demographics, domain expertise, and biospheric value orientation.

Key Contributions

Theory-Grounded LLM Prompting: The paper operationalizes behavioral nudging theories (framing and descriptive social norms) into prompt-level templates, enabling the scalable generation of personalized, theory-grounded sustainability explanations.
Empirical Validation of Presentation Strategies: The study isolates the effect of how sustainability information is presented versus that it is presented. It demonstrates that the mere presence of sustainability content does not drive behavioral change, whereas specific behavioral framings do.
Cross-Domain Generalizability: The research validates that these effects hold across both low-involvement (routine) and high-involvement (complex) decision contexts, suggesting the intervention is robust to varying levels of user deliberation.
Divergence of Perception and Behavior: The study highlights a critical divergence where plain disclosure improves subjective perception of the system without altering choice behavior, whereas framed explanations alter behavior without compromising user satisfaction.

Results

Choice Behavior (RQ1 & RQ2): Sustainability-aware explanations increased the likelihood of sustainable choices compared to the preference-only baseline. However, this effect was driven entirely by the presentation mechanism.
- Mechanism-free: Plain disclosure of sustainability attributes resulted in a sustainable choice rate of ~~54%, statistically indistinguishable from the baseline (~~51%).
- Framing: Increased sustainable choices to 69% (Odds Ratio = 2.39 vs. baseline).
- Descriptive Social Norms: Increased sustainable choices to 67% (Odds Ratio = 1.89 vs. baseline).
- Comparison: Both framing and social norms significantly outperformed the mechanism-free condition. Framing showed the strongest effect and the greatest reduction in reported choice difficulty.
Subjective Evaluation: All sustainability-aware conditions improved the perceived "sustainability impact" and "usability" of the explanations compared to the baseline. However, only the mechanism-based conditions (framing and norms) significantly reduced "choice difficulty." Crucially, adding sustainability content did not negatively impact "choice satisfaction" or "explanation clarity."
Generalizability (RQ3): Statistical tests for interaction between condition and domain were non-significant, confirming that the behavioral and experiential effects were consistent across both instant coffee and hotel booking scenarios.

Significance and Claims
The authors conclude that LLM-generated explanations serve as a scalable tool for nudging sustainable behavior, but only when the content is strategically framed using established behavioral mechanisms. The paper argues against an "information-deficit" model, positing that providing environmental facts alone is insufficient to shift choices in personalized RS. Instead, the presentation of information—specifically through positive framing or social norms—is the active ingredient that eases decision-making and drives sustainable selection.

The work demonstrates that generative AI can be used to create adaptive, theory-grounded interventions that preserve user autonomy (by not restricting options) while effectively guiding behavior toward social good. The authors emphasize that their findings are specific to the tested domains and mechanisms, noting that the divergence between improved perception (from plain disclosure) and actual behavior change (from framing) offers a nuanced insight for future adaptive explanation design.

Technical Summary: Nudging Sustainable Choices through LLM-Generated Recommendation Explanations

More like this