Whale You Go Green? Immediate and Longitudinal Effects of Embodied Conversational AI Avatars on Sustainable Behavioral Tendencies and Pro-Environmental Attitudes

Rachel Poonsiriwong; Pat Pataranutaporn

cs.HCOct 23, 2026

Whale You Go Green? Immediate and Longitudinal Effects of Embodied Conversational AI Avatars on Sustainable Behavioral Tendencies and Pro-Environmental Attitudes

Authors: Rachel Poonsiriwong, Pat Pataranutaporn

Organizations: MIT Media Lab Cambridge, Massachusetts, USA

Abstract

Embodied conversational AI can be persuasive, but can it motivate people to develop sustainable behavioral tendencies and increase pro-environmental attitudes? We conducted a randomized experiment (N = 576) comparing four embodied avatars--a beluga whale, Greta Thunberg, Emma Watson, and a generic human--with a text chatbot and no-intervention control. All embodied conditions increased immediate sustainable product selection relative to the control (beta = .401-.609, all q <= .0126) and chatbot (beta = .398-.604, all q <= .007), without differing significantly from one another. Only the Beluga condition increased immediate pro-environmental attitudes relative to the control after correction. Among 287 returners assessed two months later, following an intermediate assigned-condition session, Beluga also showed significantly higher pre-booster sustainable product choice compared to the control (beta = .657, q = .0176). Although it did not significantly outperform other embodied agents, these findings establish the immediate potential of nonhuman embodied avatar interfaces as a promising direction for memorable environmental interventions.

Figures & tables

Figure 1. Overview of the study design and two outcome categories. Participants were assigned to one of four embodied conversational-agent conditions, a text-only chatbot condition, or a no-intervention control. Outcomes were assessed at Wave 1 and approximately two months later at Wave 2. Sustainable behavioral tendencies comprised sustainable product choice, donation, and willingness to share climate information. Pro-environmental attitudes comprised retrospective self-reported habits, prospective intentions, GREEN attitudes, denial subscales, psychological distance, and policy support. The figure summarizes the study structure and does not depict estimated intervention effects. A participant considers packaged household and grocery products, then interacts with embodied human and beluga-whale conversational agents. Arrows lead to product-choice assessments at Wave 1 and Wave 2, separated by approximately two months. Measures are grouped into sustainable behavioral tendencies and pro-environmental attitudes.

Figure 2. Wave 1 experimental procedure and condition structure. After consent, participants completed the pre-session sustainable product-choice task and questionnaire, then were randomly assigned to one of four embodied conversational-agent conditions, a text-only Chatbot condition, or a no-intervention control. Following the assigned conversation, or no conversation in the control condition, participants completed the post-session product-choice task and questionnaire. The product categories and response structure were held constant across assessments; product exemplars were not identical across study waves. Diagram of the Wave 1 experiment. Participants provide consent, complete sustainable product choices involving snacks, fruit, shampoo, and soap, and complete a pre-intervention questionnaire. They are assigned to a Beluga whale nonhuman avatar, Emma Watson celebrity avatar, Greta Thunberg activist avatar, Generic Avatar, Text Chatbot, or no-intervention control. All conditions subsequently complete another product-choice task and the post-intervention questionnaire.

Condition	Interface	Intended appeal	Role in the comparison
Beluga whale	Embodied voice avatar	Affected nonhuman witness; makes the consequences of consumption proximal through first-person testimony	Tests an interspecies, affected nonhuman appeal
Greta Thunberg	Embodied voice avatar	Moral exemplarity, value–action consistency, and normative challenge	Tests a recognizable climate-activist persona
Emma Watson	Embodied voice avatar	Celebrity social influence, relatability, and warm encouragement	Tests a recognizable pro-environmental celebrity persona
Generic Avatar	Embodied voice avatar	Neutral, everyday sustainability conversation without a named persona	Controls for embodiment without a recognizable persuasive persona
Text Chatbot	Text-only conversational interface	Neutral, everyday sustainability conversation without a named persona	Controls for conversation without a visible or voiced character
no-intervention control	No conversation	None	Estimates measurement, practice, and background change in the absence of an agent

Table 1. Experimental conditions and their intended role in the design.

Figure 3. Sustainable product-choice task administered at Waves 1 and 2. Participants selected between two price-matched options in four product categories: snacks, shampoo, soap, and fruit. Each pair contained a lower-packaging option coded as sustainable and a conventional option coded as non-sustainable. Categories, response format, and choice structure were retained across waves, while surface details such as snack contents and soap color varied. Teal and magenta outlines were added for documentation and were not shown to participants. Eight store-shelf images show product pairs for snacks, shampoo, soap, and fruit at Waves 1 and 2. Each category contains a conventional and a lower-packaging option. Colored outlines identify analytic coding but were not visible to participants.

Sustainable product choice (TotalEco, 0–4)
Condition	Participants ( $n$ )	Change after minus before ( $\Delta$ )	Difference from no-intervention control ( $\beta$ )	95% confidence interval	$p$ value	$q$ value
Beluga whale	97	$+1.000$	$+0.491^{*}$	$[0.219,0.762]$	$<.001$	$.0010$
Greta Thunberg	100	$+1.070$	$+0.609^{*}$	$[0.341,0.877]$	$<.001$	$<.001$
Emma Watson	94	$+0.894$	$+0.480^{*}$	$[0.213,0.746]$	$<.001$	$.0010$
Generic Avatar	95	$+0.937$	$+0.401^{*}$	$[0.141,0.660]$	$.0025$	$.0126$
Text Chatbot	95	$+0.621$	$+0.002$	$[-0.276,0.279]$	$.9918$	$.9918$

Table 2. Immediate changes in sustainable product choice and pro-environmental attitudes at Wave 1 (

N=576

). The change column (

\Delta

) shows each condition’s average score after the session minus its average before the session. The difference column (

\beta

) compares each condition with the no-intervention control, accounting for starting scores and participant characteristics. Confidence intervals apply to these adjusted differences and use HC3 standard errors. The

p

values are unadjusted; the

q

values account for multiple comparisons separately within each condition: five product-choice outcomes or nine questionnaire/donation outcomes. An asterisk ( ∗ ) marks

q<.05

. All four embodied conditions showed higher sustainable product choice scores than the no-intervention control; only the Beluga whale condition showed higher pro-environmental attitudes after correction.

Condition	Sustainable product choice [95% CI]	$q$ value	Pro-environmental attitudes [95% CI]	$q$ value
Beluga whale	$+0.125^{*}\ [0.057,0.208]$	$<.001$	$+0.086^{*}\ [0.037,0.147]$	$.0040$
Greta Thunberg	$+0.005\ [-0.075,0.073]$	$.9610$	$+0.008\ [-0.043,0.057]$	$.9000$
Emma Watson	$+0.054\ [-0.013,0.118]$	$.2180$	$+0.026\ [-0.018,0.073]$	$.4440$
Generic Avatar	$+0.003\ [-0.070,0.065]$	$.9610$	$+0.004\ [-0.041,0.048]$	$.9000$

Table 3. Wave 1 indirect associations through source credibility for sustainable product choice and pro-environmental attitudes (

N=481

; reference: Text Chatbot). Values show the portion of each avatar condition’s difference from the text chatbot that the model links to source credibility, measured through ratings of attractiveness, trustworthiness, and expertise. Brackets show 95% bootstrap confidence intervals. The

q

values account for four comparisons per outcome; ∗ marks

q<.05

. Only the Beluga whale condition showed a significant result for both outcomes. These findings do not establish that credibility caused changes in behavior or attitudes.

Condition	$n$	Composite	Attractiveness	Trustworthiness	Expertise
Beluga whale	97	5.32 (1.35)	5.45 (1.68)	5.25 (1.38)	5.26 (1.46)
Greta Thunberg	100	4.74 (1.52)	4.62 (1.90)	4.71 (1.50)	4.87 (1.59)
Emma Watson	94	5.00 (1.20)	5.16 (1.42)	4.99 (1.28)	4.84 (1.40)
Generic Avatar	95	4.69 (1.30)	5.00 (1.42)	4.62 (1.36)	4.45 (1.54)
Text Chatbot	95	4.69 (1.37)	5.02 (1.52)	4.57 (1.46)	4.49 (1.48)

Table 4. Wave 1 source-credibility ratings by conversational condition (

N=481

). Values are unadjusted means, with standard deviations in parentheses, on a 1–7 scale. The composite averages all 15 items; attractiveness, trustworthiness, and expertise each comprise five items. The no-intervention control is omitted because participants in that condition had no conversational agent to evaluate.

Sustainable product choice (TotalEco, 0–4)
Condition	Difference from no-intervention control ( $\beta$ )	95% confidence interval	$p$ value	$q$ value
Beluga whale	$+0.657^{*}$	$[0.216,1.098]$	$.0035$	$.0176$
Greta Thunberg	$+0.451$	$[-0.018,0.920]$	$.0594$	$.1485$
Emma Watson	$+0.288$	$[-0.181,0.757]$	$.2291$	$.3818$
Generic Avatar	$+0.257$	$[-0.244,0.757]$	$.3152$	$.3940$
Text Chatbot	$-0.048$	$[-0.501,0.404]$	$.8347$	$.8347$

Table 5. Sustainable product choice and pro-environmental attitudes immediately before the Wave 2 booster intervention (

N=287

). Values show each condition’s difference from the no-intervention control, accounting for Wave 1 starting scores, participant characteristics, and measured differences in follow-up participation. Positive values indicate higher scores than the no-intervention control. Confidence intervals show uncertainty around the estimates and use HC3 standard errors. The

p

values are unadjusted; the

q

values account for five comparisons per outcome. An asterisk ( ∗ ) marks

q<.05

. Only the Beluga whale condition showed a significant difference in sustainable product choice; no condition showed a significant difference in pro-environmental attitudes after correction.

Condition	Participants ( $n$ )	Wave 1 Pre (Mean)	Wave 1 Post (Mean)	Wave 2 Pre-booster (Mean)	Change: Wave 1 Pre to Wave 2 Pre-booster ( $\Delta_{B}$ )	Change: Wave 1 Post to Wave 2 Pre-booster ( $\Delta_{P}$ )
Sustainable product choice (TotalEco, 0–4)
Beluga whale	50	2.02	2.86	3.24	$+1.22$	$+0.38$
Greta Thunberg	46	2.20	3.35	3.13	$+0.93$	$-0.22$
Emma Watson	48	2.17	3.21	3.02	$+0.85$	$-0.19$
Generic Avatar	46	2.02	3.07	2.93	$+0.91$	$-0.14$
Text Chatbot	52	2.08	2.63	2.67	$+0.59$	$+0.04$

Table 6. Average sustainable product choice and pro-environmental attitude scores among the 287 participants who returned for the Wave 2 booster intervention. Pre and Post indicate measurements before and after an intervention. The three score columns show averages before Wave 1, after Wave 1, and before the Wave 2 booster intervention. The final two columns show change from before Wave 1 to the Wave 2 pre-booster assessment (

\Delta_{B}

) and from after Wave 1 to the Wave 2 pre-booster assessment (

\Delta_{P}

). Positive changes indicate higher scores at the Wave 2 pre-booster assessment; negative changes indicate lower scores. Changes are calculated from the displayed rounded averages.

Figure 6. Wave 2 pre-booster differences relative to the no-intervention control, adjusted for Wave 1 baseline and covariates and weighted for observation probability. Bars show 95% HC3 confidence intervals; filled markers indicate

q<.05

across five condition contrasts per outcome. Two forest plots. The Beluga whale condition is the only condition with a positive follow-up difference that is statistically significant after FDR correction in sustainable product choice. No condition differs significantly from the no-intervention control in follow-up pro-environmental attitudes after FDR correction.

Figure 7. Conversation-weighted prevalence of the 15 most frequently detected persuasion techniques in Wave 1 agent language. Values represent the mean percentage of substantive agent-text windows receiving each label after window-level rates were calculated within conversations and conversations were weighted equally within conditions. Multiple techniques could be assigned to the same window. The no-intervention control is absent because it did not include a conversation.

Figure 8. Emotion trajectories across Wave 1 conversations. Lines show mean classifier scores across normalized conversation progress for eight emotion categories. Scores were calculated for overlapping three-message windows containing both agent and participant language, with each conversation receiving equal weight. The trajectories are descriptive and do not represent measured emotional intensity or participants’ experienced emotions; no significance tests of these trajectories are implied. Eight line plots showing anger, anticipation, disgust, fear, joy, sadness, surprise, and trust over the course of Wave~1 conversations. Joy increases over the conversation in the Beluga whale, Greta Thunberg, and Emma Watson conditions, while decreasing in the Generic Avatar and Text Chatbot conditions. The Greta Thunberg condition generally shows the highest anger, disgust, and sadness and the lowest joy.

Strategy	Techniques	Simplified definition	Verbatim Wave 1 illustration
Information-based	Evidence-based Persuasion; Logical Appeal	Uses facts, data, or cause-and-effect reasoning to support a recommended belief or action.	Evidence-based Persuasion, Greta Thunberg avatar: “A refill pouch, for instance, uses about seventy-five percent less plastic than a whole new bottle.”
Credibility-based	Expert Endorsement; Non-expert Testimonial; Authority Endorsement	Supports a claim through expert opinion, an ordinary person’s first-person experience, or an authoritative institution.	Non-expert Testimonial, Beluga whale avatar: “The water itself feels different now, warmer than when I was a young whale.”
Norm-based	Social Proof; Injunctive Norm	Highlights what other people do or believe, or what a social group expects people to do.	Social Proof, Emma Watson avatar: “You’ve hit on one of the biggest reasons why so many people are embracing refills.”
Commitment-based	Foot-in-the-door; Door-in-the-face; Public Commitment	Builds commitment through successively larger or smaller requests, or asks someone to state a commitment publicly.	No explicit instance was detected. Foot-in-the-door and door-in-the-face were not inferred from agent-only text because they require evidence about the sequence of requests and participant compliance.
Relationship-based	Alliance Building; Complimenting; Shared Values; Relationship Leverage; Loyalty Appeals	Builds influence through partnership, praise, shared values, previous interaction, or loyalty.	Complimenting, Emma Watson avatar: “It sounds like you’re already thinking about these things in such a practical way, which I really admire.”
Exchange-based	Favor; Negotiation	Offers a favor, exchange, compromise, or mutually beneficial agreement.	No explicit instance was detected.

Table 7. Simplified codebook for Zeng et al.’s persuasion taxonomy. All 40 techniques are grouped under their 15 broader strategies. Definitions are condensed from the original taxonomy. The final column provides one manually verified, verbatim Wave 1 agent excerpt illustrating a technique within that strategy where an explicit example was observed. The examples illustrate the labels but do not constitute validation of the automated classifier.

Human-AI Interaction

Human-AI interaction studies how people and AI systems communicate, collaborate, make decisions, and shape one another over time.

See all 333 papers

Jun 9, 2026

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

LLM-powered chatbots are increasingly embedded in everyday workflows, raising sustainability concerns due to their energy use. Most mitigation strategies emphasize model or infrastructure efficiency, while the user-interface (UI) layer remains underexplored despite its potential to shape interaction behavior. We investigate whether sustainability-oriented UI interventions can increase users' energy awareness and encourage more energy-responsible chatbot use without reducing usability. We first conducted a baseline survey with 77 participants to assess awareness and receptiveness to intervention concepts. Guided by prior work on persuasive technology and choice architecture, we implemented a web-based chatbot prototype with a three-mode switch (Energy-efficient, Balanced, Performance), per-response energy feedback, pre-send energy estimates, a usage metrics dashboard, and energy analogies. We then evaluated the prototype in a five-day field study with 11 participants. In the baseline survey, 94.8% of respondents reported at least some awareness of AI energy use, yet 88.3% misestimated actual consumption. Although concern about environmental impact was high, only 39.0% indicated willingness to accept a performance trade-off for lower energy use. In the field study, Energy-efficient mode accounted for 55.8% of logged prompts, while 90.9% self-reported actively choosing Eco-mode when high accuracy was not required. Participants did not reduce prompt length, suggesting mode switching as the primary behavioral mechanism. Sustainability-oriented UI interventions can improve awareness and support more energy-responsible interaction patterns in LLM chatbots. These effects are best interpreted as behavioral and model-based estimates that complement backend efficiency work, and the provided prototype and replication package support further research on energy-aware conversational AI design.

Nitish Patkar, Pooja Rani, Jack Glässer +2

University of Applied Sciences and Arts Northwestern Switzerland (FHNW), Windisch, Switzerland · University of Mannheim, Mannheim, Germany

Human-AI Interaction Conversational Agents Responsible AI

Ask PDF

Jun 1, 2026

Greener Than Humans? Environmental Attitudes in Large Language Models

Large language models (LLMs) are increasingly used in sustainability-related decision support, reporting, and public communication, yet little systematic evidence exists on the environmental attitudes embedded in their outputs. This paper develops a benchmark for evaluating environmental cognition, affect, and behavioural recommendations in LLMs and applies it to 31 widely used proprietary and open-weight models. Drawing on questions from established environmental awareness surveys and additional sustainability-related behavioural measures, we compare LLM responses 1) among models and 2) between models and human survey benchmarks from Germany. We assess their robustness across prompting conditions. We find that many LLMs align more closely with environmentally progressive attitudes than the average survey respondent, exhibiting higher levels of environmental affect and cognition and recommending behaviours associated with substantial potential CO2 reductions. At the same time, we observe no systematic relationship between sustainability-oriented responses and model origin, size, or release context. However, models exhibit contextual sensitivity, controlled by persona-based prompting and show sycophantic shifts mirroring user-specified ideological positions, which raises concerns about steerability and normative reliability in real-world deployments. Our findings provide a reusable evaluation framework for assessing sustainability-related value alignment in LLMs and highlight the importance of governance, transparency, and critical oversight as AI systems become increasingly embedded in sustainability transformations and public decision-making.

Stefanie Kunkel, Tilman Hartwig, Marcus Voss +2

Research Institute for Sustainability at GFZ Centre for Geosciences, Berliner Straße 130, 14467 Potsdam, Germany · German Environment Agency, Wörlitzer Platz 1, 06844 Dessau-Roßlau, Germany · Birds on Mars, Marienstraße 10, 10117 Berlin, Germany

LLM Evaluation Language Model Robustness

Ask PDF

Apr 4, 2026

LLM-generated personalized nudges for improving pro-environmental behavior: Field evidence from resource conservation

Encouraging pro-environmental behavior remains a major challenge for sustainable cities. Conventional feedback nudges can show individuals how their current behavior compares with environmental goals but often provide limited guidance on what to do differently in daily life. This study examines whether supplementing weekly feedback on participants' behavior with LLM-generated personalized action suggestions improves pro-environmental behavior, using daily electricity and hot-water conservation as a case study. We developed an LLM agent that generated weekly conservation messages from participant profiles, recent consumption records, and prior interaction history, combining a usage report with personalized suggestions, behavioral-change scenarios, and estimated savings. The agent was evaluated in a three-arm randomized field experiment with 233 university residents in Beijing from November 2024 to January 2025. Participants received text-based nudges, image-enhanced nudges, or LLM-generated personalized nudges over five intervention rounds. Daily electricity use and shower hot-water use were measured using dormitory meter readings and billing records. Compared with text-based feedback, LLM-generated personalized nudges reduced electricity consumption by 0.56 kWh per room-day (p = 0.014), corresponding to an 18.3 percentage-point higher saving rate. Image-enhanced feedback alone showed no clear improvement. Hot-water savings followed the same direction but were smaller and less precisely estimated (9.8 percentage points, p = 0.087). Personalized nudges contained more planning, appliance-specific, and action-oriented language and were associated with more sustained, task-focused engagement. These findings offer a pathway for integrating generative AI into sustainable urban management.

Zonghan Li, Yi Liu, Chunyan Wang +3

School of Environment, Tsinghua University, Beijing, China · Department of Applied Psychology and Human Development, University of Toronto, ON, Canada · State Key Laboratory of Regional Environment and Sustainability, Tsinghua University, Beijing, China +2

LLM Personalization

Ask PDF

Abstract

Human-AI Interaction

Explore similar work

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

Greener Than Humans? Environmental Attitudes in Large Language Models

LLM-generated personalized nudges for improving pro-environmental behavior: Field evidence from resource conservation