Source-linked AI summary
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
TL;DR
This study asks whether embodied conversational AI can support sustainable choices and pro-environmental attitudes, and whether those effects remain detectable over time. A randomized comparison found immediate choice advantages across embodied conditions, while Beluga showed the clearest corrected control differences at follow-up.
Problem
Climate change can feel distant from everyday decisions, and it remains unclear whether affected nonhuman avatars have effects beyond an interaction or advantages over human agents, text chat, or no intervention.
Source: §1 · p. 2Method
A randomized experiment assigned 576 participants to four embodied avatars, a text chatbot, or no intervention, measuring choices and attitudes immediately and assessing returners about two months later.
Source: §1 · p. 2Results
All four embodied conditions increased immediate sustainable selection versus text chat and no intervention; Beluga alone showed corrected control differences in immediate attitudes and pre-booster choice at Wave 2.
Source: §1 · p. 2, §7 · p. 17Takeaways & Limitations
Affected nonhuman agents merit further longitudinal study as a promising direction for memorable environmental interventions.
Source: §7 · p. 17The primary outcome was an unincentivized repeated screen-based choice task, so the findings do not establish verified purchasing changes outside the study.
Source: §5.5 · pp. 16–17
Abstract
from arXiv · showhide
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.
1 Introduction
The study asks whether embodied conversational agents can support sustainable choices and attitudes, and whether effects persist beyond an immediate interaction. It compares four avatars with text-only and no-intervention conditions, finding immediate choice advantages across embodied conditions and a narrower longitudinal pattern for Beluga.
1 Introduction
- Climate change is difficult to connect to everyday decisions, although individual action often occurs through ordinary choices that conversational AI could make personally relevant.Scalable persuasive messages also create risks of manipulation or unwanted influence.Source: §1 · p. 2
- A beluga avatar may frame environmental harm through an affected animal, but whether such a nonhuman character has lasting effects or advantages over human, text, or no-intervention conditions remains unclear.Prior work links anthropomorphized nature with connectedness and protective concern, and marine characters with sustainability intentions and preferences.Source: §1 · p. 2
- The experiment randomized 576 participants across four avatar conditions, a text chatbot, and no intervention, with 287 returning about two months later for follow-up and a booster.The four embodied conditions used spoken dialogue, while the chatbot used typed dialogue.Source: §1 · p. 2
- All four embodied conditions increased immediate sustainable selection relative to both no intervention and text chat, without significant differences among the embodied conditions.Beluga alone showed corrected differences from the no-intervention control in immediate GREEN attitudes and pre-booster sustainable choice at Wave 2; direct tests did not establish that it outperformed every other embodied condition.Source: §1 · p. 2
2 Related Work
Prior work presents conversational persuasion, embodiment, and environmental interventions as context-dependent rather than uniformly effective. The study addresses gaps in comparative and longitudinal evidence by testing distinct agent representations alongside text-only and no-intervention conditions.
2.1 Persuasive Conversational AI and Consumer Influence
- Conversational AI can personalize persuasion, but effects vary across beliefs, intentions, behavior, sources, recipients, and contexts, while manipulation risks make transparency and meaningful choice relevant.Prior work reports low detection of manipulative turns and reduced compliance after just-in-time warnings [60].Source: §2.1 · pp. 2–3
2.2 Embodiment, Represented Identity, and Celebrity Likeness
- Embodiment does not confer a uniform advantage, and represented identity, modality, and ongoing interaction may shape responses, motivating direct comparisons across human and nonhuman agents.An affected-animal framing connects to Animal–Computer Interaction’s treatment of animals as stakeholders, while prior marine-character work found especially strong empathy toward a beluga.Source: §2.2 · p. 3
2.3 Digital Interventions for Pro-Environmental Behavior
- Environmental interventions can affect beliefs, policy support, sharing, and behavior differently, while individual product choices do not capture the institutional and material systems shaping sustainability.Prior work rarely combines text-only and no-intervention controls with comparisons among recognizable human, generic human, and affected nonhuman agents over time.Source: §2.3 · p. 3
3 Methodology
The study uses a randomized six-condition design to distinguish embodied-interface effects from text conversation and no intervention, while examining attitudes, sustainable choices, credibility, and persistence. Its intervention combines condition-specific agents with shared factual content and modality differences.
3 Methodology
- The randomized design compares four embodied agents, a text chatbot, and no intervention, asking about attitudes, sustainable tendencies, credibility associations, and differences after repeated exposure.The study distinguishes immediate task performance, broader attitudes, and longitudinal condition differences approximately two months later.Source: §3 · pp. 3–4
3.1 Intervention Design
- The embodied interface comprised a visible avatar, synthetic voice, and spoken interaction; its generic-avatar versus text-chatbot contrast tests the combined configuration against text-only conversation.The generic avatar and chatbot were aligned in persona and topic, but the contrast does not separate visual presence, voice, and spoken interaction.Source: §3.1 · p. 4
3.1.1 Technical Implementation.
- The implementation used condition-specific prompts and spoken or typed dialogue, with a shared fixed factual block across all five conversational conditions.Images were stylized rather than photorealistic, and visual properties such as cuteness and uncanniness were design choices, not independently manipulated or measured; factual values were embedded rather than retrieved live.Source: §3.1.1 · p. 4
3.2 Procedure and Longitudinal Structure
The procedure measured choices and attitudes before and after the initial session, then followed returning participants through an intermediate assigned-condition session and a Wave 2 booster. The intermediate sample was too small and unequal for confirmatory condition comparisons.
3.2 Procedure and Longitudinal Structure
- At Wave 1, participants completed baseline and pre-session choice measures, received their assigned conversation or no conversation, and then completed post-session choice and survey measures.The shopping task required choosing between paired options; no purchase or delivery occurred.Source: §3.2 · p. 4
- Returning participants retained their randomized condition through an intermediate assigned-condition session and a Wave 2 follow-up booster.Wave 2 was the next assessment with sufficiently complete data for primary longitudinal comparisons.Source: §3.2 · p. 4
- Only 142 participants had complete, matching intermediate-session survey records, with 13–21 per conversational condition versus 56 controls, so those comparisons were exploratory.A client-side caching issue reduced complete, linkable application records; 141 participants also had complete application-based choice records.Source: §3.2 · p. 4
3.3 Ethical Considerations
The study used a six-condition randomized procedure and disclosed that public-figure agents were AI-generated simulations, not affiliated with or endorsed by the figures. The stimuli were bounded to a noncommercial study, with broader synthetic-identity risks noted.
3.3 Ethical Considerations
- The study was preregistered as AsPredicted #300,140 and exempted under MIT protocol #E-7428.Source: §3.3 · p. 4
- Before interaction, participants were told that the Greta Thunberg and Emma Watson agents were AI-generated simulations without their affiliation or endorsement.Source: §3.3 · p. 5
- The public-figure visual and voice assets were bounded stimuli for a noncommercial study; responsible deployment calls for synthetic-identity and persuasive-purpose disclosure, data minimization, and refusal options.The passage also cautions that publicly accessible assets are not necessarily in the public domain and flags false-endorsement and reputational concerns.Source: §3.3 · p. 5
3.4 Sample Size and Design Sensitivity
The sample-size plan targeted roughly 600 participants across six conditions, while the achieved sample supported detection of moderate effects more readily than small follow-up differences. The choice task retained its product categories and response structure across waves while varying exemplars.
3.4 Sample Size and Design Sensitivity
- With N = 576, the six-group design had 80% power for Cohen’s f ≈ .150; two groups of about 95 had 80% power for d = .41 at α = .05.At α = .01, the two-group threshold was d = .50.Source: §3.4 · p. 5
- At Wave 2 (N = 287), the corresponding 80%-power thresholds were f = .214 and d = .58 at α = .05, indicating limited sensitivity to small follow-up differences among embodied agents.At α = .01, the two-group threshold was d = .71.Source: §3.4 · pp. 5–6
- The choice task retained four categories—snacks, shampoo, soap, and fruit—and its response structure across waves, but changed product exemplars.Each pair contrasted a price-matched lower-packaging option with a conventional option.Source: Figure 3 · p. 6
3.5 Participant Recruitment and Analytic Samples
Wave 1 included 576 matched, attention-check-passing U.S. resident records across six conditions. The intermediate sample and Wave 2 follow-up were smaller, with Wave 2 retaining 287 participants.
3.5 Participant Recruitment and Analytic Samples
- Wave 1 included 576 complete, matched, attention-check-passing records, with 94–100 participants in each of the six conditions.Condition counts were Beluga 97, Greta Thunberg 100, Emma Watson 94, Generic Avatar 95, Text Chatbot 95, and control 95.Source: §3.5 · p. 6
- The intermediate sample comprised 142 matched, attention-check-passing pairs; product-choice records were available for 141.The condition counts ranged from 13 to 56 participants.Source: §3.5 · p. 6
- Wave 2 retained 287 participants, or 49.8% of Wave 1, after 27 of 314 post-survey records failed quality, matching, or eligibility checks.The retained count was 45–52 participants per condition.Source: §3.5 · p. 6
3.6 Statistical Analysis and Model Specification
The analysis used baseline-adjusted models for separate behavioral and attitudinal outcomes, with distinct procedures for follow-up, credibility associations, and multiple-testing correction. The product-choice task preserved its decision structure across waves while varying surface details.
3.6.1 Dependent Variables.
- Outcomes were analyzed separately across sustainable behavioral tendencies and pro-environmental attitudes; sustainable product choice summed four lower-packaging selections as TotalEco (0–4).The product-choice task kept categories, price matching, and response structure constant across waves while changing surface details to reduce exact-image retrieval cues.Source: §3.6.1 · pp. 6–7
3.6.2 Agent Perception Measures and Covariates.
- Source credibility used 15 items covering attractiveness, trustworthiness, and expertise, but cross-condition comparisons warrant caution because participants may interpret these dimensions differently across agent types.Source: §3.6.2 · p. 7
3.6.3 Scoring and Data Preparation.
- Multi-item scales used available-item means after reverse coding, while choice and plastic inventories were summed; required-case missingness was excluded without follow-up imputation.Cronbach’s α was used to assess consistency.Source: §3.6.3 · p. 7
3.6 Statistical Analysis and Model Specification
- Post-session outcomes were estimated with linear ANCOVA, adjusting intervention coefficients for the pre-session outcome and baseline covariates, with no intervention as the reference.Original-unit estimates, 95% confidence intervals, and HC3 standard errors were reported.Source: §3.6 · p. 6, §3.6.3 · p. 8
3.6.5 Follow-up and Across-wave Analyses.
- Primary Wave 2 models tested pre-booster outcomes after the intermediate assigned-condition session, adjusting for Wave 1 baseline scores and covariates and weighting for measured selection.Because participants could repeat their assigned intervention, estimates represent the intervention sequence rather than isolated Wave 1 persistence.Source: §3.6.5 · p. 8
3.6.6 Source Credibility Analysis.
- Credibility analyses estimated condition–credibility and credibility–outcome paths and indirect associations, but post-session timing leaves their temporal ordering unknown, so they are not causal mediation.Source: §3.6.6 · p. 8
3.6.7 Multiple-Testing Correction and Sensitivity Analysis.
- The analyses reported unadjusted p and BH–FDR-adjusted q values, treating q < .05 as significant; a post hoc sensitivity analysis pooled conditions into larger testing families.Source: §3.6.7 · p. 8
4 Results
Wave 1 showed a shared immediate choice advantage for embodied agents, while attitude and follow-up choice differences were concentrated in Beluga; later outcomes and exploratory analyses qualified that pattern.
4.1 Embodied Avatars Increased Immediate Sustainable Product Choice
- All four embodied conditions exceeded control on choice (β=.401–.609, all q≤.0126) and each exceeded chatbot directly (β=.398–.604, all q≤.0070).The embodied conditions did not differ significantly from one another.Source: §4.1 · p. 8, Table 2 · p. 9, Figure 4 · p. 9
- Beluga and Greta showed selected next-week intention shifts; only Beluga refillable intention survived pooled correction, while donation and sharing willingness did not differ by condition.These intention measures were not observed behavior.Source: §4.1.1 · p. 8, §4.1.2 · p. 8
4.2 The Non-Human Avatar Significantly Increased Pro-Environmental Attitudes
- Only Beluga showed higher GREEN attitudes than control after correction (β=.361, q=.0008); climate-related beliefs and policy support did not differ significantly across conditions.Beluga also exceeded Generic Avatar and Emma on corrected exploratory contrasts, but not Greta.Source: Figure 5 · p. 9, §4.2 · p. 10, §4.2.1 · p. 10
4.3 The Beluga Condition Retained a Sustainable Product-Choice Difference at Follow-Up
- Before the Wave 2 booster, only Beluga showed significantly higher sustainable product choice than control (β=.657, 95% CI [.216,1.098], p=.0035, q=.0176); direct tests did not establish superiority over other embodied conditions.Source: §4.3.1 · p. 10, §4.3.2 · p. 10
4.4 Pro-Environmental Intentions Predicted Later Retrospective Self-Reported Habits Across Avatar Types
- Across conditions, Wave 1 intentions predicted matching Wave 2 self-reports for water, energy, meat, transport, and refillables (β=.278–.368; refillable β=.307), not specifically Beluga.Refillable intention also predicted fewer single-use-plastic purchases (β=−.212).Source: §4.4 · pp. 10–11
4.5 Exploratory dialogue and feedback analyses suggest candidate explanations for effects observed with the Beluga whale avatar
- Beluga language featured Non-expert Testimonial (55.1%, 32.5 points above next); feedback included affected-animal empathy and a later refillability-memory account as exploratory candidate explanations.The analyses were not tests of causal mechanisms.Source: §4.5.1 · pp. 12–13, §4.5 · p. 12
4.6 Participant Characteristics Were Balanced Across Conditions
- Wave 1 showed no significant condition differences in measured demographics or baseline attitudes, dispositions, and plastic purchasing.The paper treats these as descriptive balance checks; random assignment underpins the comparison.Source: §4.6 · p. 14
4.7 Participant Characteristics Remained Balanced Across Conditions in Wave 2
- Returners differed from nonreturners on six baseline characteristics, but adding water use to sensitivity models left the Beluga pre-booster choice result significant (β=.633, q=.0217).Retention proportions ranged from 46.0% to 54.7% across conditions.Source: §4.7 · p. 14
5 Discussion
The discussion finds that embodied interfaces improved immediate sustainable product choice, while the Beluga condition showed the clearest later control contrast. It interprets the Beluga pattern as a possible decision cue, emphasizes outcome-specific effects and ethical and ecological limits, and reports computational resource use.
5.1 Embodied interfaces can support consumers in making sustainable product choices in the moment
- 0.401–0.609 choices higher than control and 0.398–0.604 higher than the text chatbot across embodied conditions; the conditions did not differ significantly from one another.The Generic Avatar also exceeded the chatbot, but this comparison jointly varied visual presence, synthetic voice, and spoken versus typed interaction, so no single feature was isolated.Source: §5.1 · p. 15
- At Wave 2 pre-booster, Beluga alone showed a corrected product-choice advantage over control (β = 0.657, q = .0176), after an intermediate assigned-condition session.The direct comparison with the other embodied conditions was not significant; the result describes the intervention sequence, not isolated Wave 1 persistence.Source: §5.1 · p. 15
5.2 An affected non-human may provide a memorable cue to encourage sustainable behavior
- Beluga dialogue paired reflective thinking with non-expert testimonial language, and participant feedback often referenced affected animals or ecosystems, suggesting a possible link to actionable product attributes.Cue accessibility and unaided recall were not measured, so a memorable-cue mechanism remains an interpretation rather than an established explanation.Source: §5.2 · p. 15
5.3 Conversational positivity is an incomplete indicator of sustainable product-choice effects
- The most positive agent and participant language did not consistently align with the clearest behavioral evidence, so sentiment and sustainable product choice warrant separate evaluation.The descriptive comparisons do not establish that positivity is irrelevant or that sadness improves persuasion.Source: §5.3 · p. 16
5.4 Changes in sustainable product choice do not establish broader changes in climate beliefs
- Product-choice effects did not establish broad climate-belief change: no corrected differences appeared in denial, psychological distance, policy support, donation, or sharing, and no Wave 2 pre-booster GREEN contrast survived correction.The authors recommend targeting specific beliefs and testing whether effects generalize beyond practiced product choices; intervention duration was not experimentally varied.Source: §5.4 · p. 16
5.5 Ethical implications and limitations
- The study establishes greater sustainable selection within a repeated screen-based task, not verified purchasing changes; the authors also call for disclosure, consent, and refusal options in commercial deployments not tested here.The task used price-matched options from four familiar categories, leaving generalization to products with more heterogeneous or identity-linked preferences open.Source: §5.5 · pp. 16–17
5.6 Computational Resource Use and Environmental Considerations
- The archived logs covered 722 sessions and 9,386 agent turns, but the illustrative footprint estimate excluded provider infrastructure and several other system components.Provider-specific environmental measurements were unavailable, and the estimate was explicitly not a complete study-footprint calculation.Source: §5.6 · p. 17
6 Future Work
Future work should test whether affected nonhuman agents make environmental consequences memorable, whether effects generalize to consequential purchasing, and how long decision cues remain accessible.
Future studies of affected nonhuman agents
- Future experiments could vary nonhuman identity, affected-stakeholder perspective, novelty, and cuteness while measuring empathy, guilt, moral standing, and recall.These tests could clarify whether affected nonhuman characters make consequences more salient and memorable at later decisions.Source: §6 · p. 17
Generalization to purchasing decisions
- Testing novel products, incentivized choices, simulated baskets, and actual purchases could show whether effects generalize beyond the standardized, price-matched task.The authors identify aesthetics, brand, price, fit, and gendered marketing as relevant product dimensions for future tests.Source: §6 · p. 17
Longitudinal and deployment studies
- Additional follow-ups, independently varied booster exposure, and unaided-recall measures could clarify cue accessibility over time and when booster conversations are useful.Deployment studies could also examine how disclosure, user control, and refusal affect persuasive outcomes and perceived autonomy.Source: §6 · p. 17
7 Conclusion
The study finds that embodied-avatar conditions increased sustainable selection on a repeated screen-based task relative to text chat and no intervention. Beluga showed the clearest corrected attitude and follow-up choice contrasts with control, motivating further study of affected nonhuman agents.
Conclusion
- Embodied-avatar conditions increased sustainable selection on a repeated screen-based task relative to both the text chatbot and no intervention.This conclusion is bounded to the tested implementation and task.Source: §7 · p. 17
- Beluga alone showed corrected post-intervention pro-environmental attitude and pre-booster Wave 2 product-choice advantages over the no-intervention control.The conclusion presents affected nonhuman agents as a promising direction for memorable environmental interventions and further longitudinal study.Source: §7 · p. 17
- The findings raise further questions about how embodied characters can shape decisions over time.Source: §7 · p. 17
8 Generative AI Disclosure
The disclosure describes generative AI use in manuscript preparation and in study-system components, including response generation, spoken-agent infrastructure, and visual-material creation.
Generative AI Disclosure
- Generative AI was used to refine phrasing in author-drafted manuscript text.Source: §8 · p. 17
- Claude Sonnet 4.6 generated typed chatbot responses and assembled personalized, condition-specific conversational context.Source: §8 · p. 17
- ElevenLabs provided spoken-agent infrastructure, while Google’s Nano Banana model created or modified static study-interface visuals.Exploratory persuasion-technique analysis used deterministic rules, and emotion analysis used a pretrained CardiffNLP RoBERTa classifier.Source: §8 · p. 17