Care-Based Eco-Feedback Augmented with Generative AI: Fostering Pro-Environmental Behavior through Emotional Attachment

Lights out! With the escalating climate crisis, eco-feedback has gained prominence over the last decade. However, traditional approaches could be underperforming as they often use data-driven strategies and assume that people only need additional information about their consumption to change behavior. A proposed path to overcome this issue is to design eco-feedback to foster emotional connections with users. However, not much is known about the effectiveness of such designs. In this paper, we propose a novel care-based eco-feedback system. Central to the system is a Tamagotchi-inspired digital character named Infi who gets its life force from the user’s energy savings. Additionally, we harness the latest advancements in generative artificial intelligence to enhance emotional attachment through conversational interactions that users can have with Infi. The results of a randomized controlled experiment (N=420) convey the fact that this design increases emotional attachment, which in turn increases energy-saving behavior.


INTRODUCTION
In light of the escalating energy crisis [53], it is crucial that consumers become more aware of their energy usage and move towards reducing their energy consumption.On average, daily electricity consumption is around 60 kWh per day per person globally [94].Research suggests that digital interventions can potentially lead to a 5% decrease [8].Although this reduction may seem modest, it is a critical component of a comprehensive strategy, where every efort contributes signifcantly to the overall objective of saving energy resources [8].Past research in sustainable HCI and related felds has investigated various motivational afordances to contribute to the global eforts to reduce energy consumption [101].These include elements or features that are designed to raise awareness about energy consumption and support behavior change towards more energy savings [20].A recent systematic literature review highlights over 80 initiatives dedicated to address the challenge of reducing energy consumption with digital systems.Such initiatives have become more common with the increasing deployment of smart meters [5] allowing for real-time electricity consumption feedback.However, most systems only use statistical visualizations to provide feedback, which might not be the most appropriate design to foster energy-saving behavior [20].One criticism is that such systems, and sustainable HCI in general, focus too much on data-centric aspects and oversimplify the complexity of human engagement with sustainability [15,80,90].Furthermore, feedback systems tend to be less successful with individuals who are less environmentally aware [81].
A potential solution to address these concerns could be to leverage emotional attachment in addition to data-driven approaches when it comes to designing eco-feedback systems [10,11,99].Some existing systems are specifcally designed to foster a bond with the users.Referred to as care-based systems, they contain a digital entity that users must nurture and take care of [28,40,68,73].Furthermore, novel generative artifcial intelligence (GenAI) tools could further enhance the emotional attachment a user feels towards the systems through engaging conversational interactions [29].However, there is a lack of empirical research investigating how care-based eco-feedback systems increase emotional attachment and eventually afect energy-saving behavior.
In this paper, we will specifcally address this gap through the following research question: RQ: Can GenAI-augmented care-based eco-feedback lead to higher emotional attachment and promote increased energy-saving behavior?

ENERGY-SAVING LEVEL
Have you considered using LED bulbs?Yes I usually buy LED bulbs :-) Fantastic!This not only helps you save energy and money, but it also increases my life force.Do you use a car or public transportation?

Contributions
To answer this question, our paper presents the following contributions: First, we design and implement a novel GenAI-augmented digital care-based eco-feedback system to foster energy-saving behavior.At the heart of the system lies a digital avatar called Infi, whose vitality depends on energy-saving behavior and interactions with its owner.The novelty of this system, in comparison to other similar approaches (e.g., [26,84]), lies in its augmentation of a virtual avatar approach with state-of-the-art conversational agents.Indeed, with current generative artifcial intelligence, exemplifed by systems like ChatGPT or Bard, computer conversation facilitates lifelike human-computer conversations [68,76,95].Even though research around conversational agents has grown exponentially in the last years, this research is, to the best of our knowledge, the frst to investigate its potential in the context of care-based eco-feedback.
Second, we propose and validate a novel model to explain the path from artifact design to energy-saving behavior through a psychological mediator (i.e., emotional attachment).Even though gamifed eco-feedback is popular, there is a lack of research that clarifes the psychological multi-step processes that connect motivational afordances to behavioral outcomes, as reported in Chalal et al.'s recent systematic literature review [20].This article contributes to the literature by validating such a model in a randomized controlled experiment with 420 participants, complemented by a follow-up study conducted three months later, with 268 (out of the original 420) participants.
Third, we analyze the impact of the care-based eco-feedback system on participants with diferent degrees of environmental awareness.Indeed, whereas environmentally aware individuals might already have high energy-saving levels, it is crucial to investigate how such interventions can be efective for those who are most hesitant to adopt pro-environmental behavior.

Research Approach
This research is based on a user-centered design approach [2] and includes aspects of design thinking [7] and design science research methodology (DSRM [30,77]).Practice and theory informed the artifact design, and the psychological and behavioral outcomes used to measure its efects.Quantitative methods (surveys and log data) as well as qualitative methods were used to assess the diferent outcomes.In line with recent calls in sustainable HCI research [15], this project featured a collaborative efort spanning 12 months with a multidisciplinary team of specialists from information systems, economics, behavioral science, psychology, marketing, design, and computer engineering.The project included partners from the industry (a utility provider) and from the government (the country's federal ofce for energy).We conducted co-design workshops with partners to better understand the problem at hand and to explore the design space before designing frst prototypes.
This paper is organized as follows: Section 2 defnes the research objectives and hypotheses for the proposed solution.Sections 3 presents the design and implementation of the eco-feedback solution.Section 4 details the evaluation setup of the solution, before Section 5 and Section 6 that detail quantitative and qualitative results respectively.Finally, Section 7 discusses the fndings, before Section 8 wraps up with a conclusion.

DEFINING THE OBJECTIVES OF THE SOLUTION
In this section, we outline the goals of our proposed solution.For this, we conduct a review of relevant literature on eco-feedback systems, emotional attachment, care-based approaches, conversational interaction, and environmental awareness.This review highlights open research gaps and leads to four hypotheses that we test in this paper.

Eco-feedback
Eco-feedback systems ofer evaluative information regarding the actions and behaviors of individuals, for instance, how much daily electricity they consume [35].The goal of these systems is to motivate users to adopt energy-saving behaviors, which can span from immediate individual behavior like turning of lights, to long-term behavior like investing in energy-efcient appliances, or collective behavior such as supporting public policy initiatives or community engagement [38,59,78].The efectiveness of eco-feedback interventions for individuals has already garnered signifcant attention among academics.For example, past research has demonstrated that eco-feedback can potentially overwhelm individuals with abstract numerical data [23,46].Such feedback often relies on presenting users with raw data related to their energy consumption or environmental impact, which may lack personal relevance or emotional resonance.As a result, individuals struggle to translate this information into meaningful actions, which can undermine motivation and engagement, ultimately limiting the efectiveness of the feedback in driving energy-saving behavior change [15].In a similar vein, the absence of personalized and interactive features results in a passive recipient-user relationship, where the feedback is merely received without active participation.This hinders the development of a strong sense of ownership and empowerment over one's environmental impact, potentially diminishing the adoption of energy-saving behaviors [20].

Emotional Attachment to Increase Eco-feedback Efectiveness
Conceptually, eco-feedback systems can be seen as motivational afordances [101] that aim at producing behavioral outcomes, i.e., guiding users to adopt energy-saving behaviors [63].However, the path from the motivational afordance to its desired behavioral outcome is not direct, as it is usually mediated by a psychological outcome [44].Figure 2 illustrates this multi-step model.Chalal et al. 's recent systematic review of eco-feedback systems indicates that the psychological outcomes that are targeted by most of these systems include rational decision-making processes such as raising awareness or increasing learning [20].In line with the recent refection on sustainable HCI [15], this rather rational decisionmaking approach might be reductive and could be improved by designing with emotional aspects in mind.Typically, emotional attachment is one such psychological outcome that is targeted by certain eco-feedback systems [20].Emotional attachment refers to the emotional bond between individuals and a specifc target [55].An early and promising study explored how an eco-feedback system could leverage emotional attachment with a virtual polar bear to promote environmentally responsible behavior [26].Their results, based on a single case experiment with 20 participants, suggest that higher emotional attachment to the virtual pet signifcantly increased environmentally responsible actions.These fndings provide an encouraging starting point to conduct a randomized controlled experiment to validate a multi-step model.However such multi-step analyses are not common, as illustrated in Chalal et al.'s systematic literature review [20].Indeed, none of the 30 ecofeedback systems in categories appropriate to develop emotional attachment (i.e., artistic and game-based), provides a multi-step analysis.Based on the observations above, we hypothesize that individuals who form a stronger emotional attachment with ecofeedback systems will engage in greater energy-saving behavior (see Figure 3 for the conceptual model).Formally stated: H1: Increasing emotional attachment towards an eco-feedback system increases energy-saving behavior.

Care-based Gamifcation to Foster Emotional Attachment
According to Chalal et al., the most appropriate eco-feedback systems for emotional attachment are game-based and artistic approaches [20].A game-based approach involves the application of gamifcation techniques and principles to educate and engage individuals in pro-environmental behaviors and consciousness-raising eforts (e.g., [19,60,83]).This may entail the utilization of avatars or leaderboards for instance [20].The artistic approach involves using creative techniques to communicate a specifc concern or message rather than simply presenting raw data.It aims to transform data into visually engaging representations that maintain readability while fostering an emotional connection between end-users and the environment in the context of eco-feedback [20].Game-based methods work well for encouraging energy-saving practices and promoting interaction, while the artistic approach is particularly efective at sparking curiosity and nurturing an emotional connection [20].Other researchers promote care-based systems to increase emotional attachment by design [57,68].These systems are based on digital entities that fourish as users take care of them [68].Prior works highlight the importance of care-centric approaches that foster a more engaging relationship with the users and the environment [25,93].In the context of eco-feedback systems, these approaches encourage the design of digital artifacts that not only respond to user actions but also embody a relationship of care and responsibility.Such artifacts, like the polar bear system presented above [26], are not just tools for feedback but become entities with which users form an emotional bond, infuencing their behavior towards more sustainable and caring interactions with their environment.As such, digital avatars, particularly in the form of virtual pets, have been utilized to evoke emotional connections and responsibilities [12,21,62,67,89].Other systems adopt the growing tree metaphor and encourage users to take care of it, which leads them to engage in various energy-saving actions [71,79,96].Based on these observations, we hypothesize that: H2a: Introducing a care-based artifact in an eco-feedback system increases emotional attachment compared to a non-carebased version.

Conversational Interaction to Increase Emotional Attachment
We argue that enabling users to interact with a care-based system could improve their emotional attachment, as active interaction often cultivates a deeper sense of connection [29].Conversational Agents, commonly referred to as dialogue systems or chatbots, are software applications that provide human-like conversations with users [103].Traditionally, chatbots have relied on task-oriented structures, employing conversational rules to provide specifc responses, such as answering inquiries.For instance, to address emotion in large-scale conversation generation, researchers have investigated a novel approach through three innovative mechanisms, yielding contextually appropriate responses encompassing both content and emotion [102].However, recent advancements in Generative AI (GenAI) have led to signifcant progress in open-domain dialogue systems, enabling unconstrained conversational interactions on various subjects with some focusing on fostering empathy [51,97,102].GenAI utilizes a sophisticated class of algorithms in the domain of Natural Language Processing, designed to acquire intricate representations of textual data without explicit supervision [65].Leveraging extensive datasets, GenAI utilizes transformative architectures to capture the inherent contextual dependencies and semantic nuances of language, facilitating efcient encoding and decoding of natural language sequences [41].As such, the potential of GenAI chatbot remains largely unexplored, limiting our understanding of the advantages and limitations of free-form conversations in addressing global issues.However, we argue that thanks to more engaging interactions, GenAI chatbots will lead to improved emotional attachment.Stated formally: H2b: Interacting with a Gen-AI enabled care-based eco-feedback system increases emotional attachment compared to a non-GenAI version.

Environmental Awareness
Environmental awareness is defned as conscious behavior towards the environment, such as pro-environmental actions [18].Prior research has explored the role of environmental awareness in proenvironmental behavior, with results indicating a substantial impact on building pro-environmental behavior [45].In most cases, environmental awareness has been found to positively infuence pro-environmental behavior and outcomes [36,69].Previous research has demonstrated that feedback serves as a signifcant trigger for individuals with strong environmental awareness, as it allows them to perceive the positive impact of their actions on the environment [34,35].On the fip side, research seems to indicate that eco-feedback fails to work for less environmentally aware individuals [82] -those who might need it the most.We argue that through emotional attachment, eco-feedback systems might be able to reach such individuals.Indeed, previous research has shown that emotional positive strategies in the context of conversational interactions are more successful in persuading users than purely rational strategies [3].We make the following hypothesis: H3: Increasing emotional attachment towards an eco-feedback system increases energy-saving behavior for individuals with low environmental awareness.

CARE-BASED ECO-FEEDBACK SYSTEM DESIGN
The design process combined inputs from an initial user co-design workshop with expertise from designers and existing literature.Its design was then refned based on qualitative and quantitative inputs from fve iterative and incremental online pilot studies.

Design Principles
Theoretically, our design can be mapped on Zhang's fve key design principles for motivational afordances [101]: (1) allowing autonomy and self-control, (2) supporting competence and achievement, (3) fostering social connections, (4) facilitating leadership and followership, and (5) triggering emotions.The avatar can be seen as a representation of one aspect of the user's self-identity -their energy consumption (Principle 1).The energy level and the avatar contribute to timely and positive feedback of the system (Principle 2).To contribute to this principle, we designed the saving level in such a way that it cannot be negative.For instance, even if the user consumed more than the baseline, there is no zero mark under which the level will go.In the same spirit, there are no negative colors, such as orange or red that are associated with excessive consumption.The avatar's character-like appearance and the conversational interaction with it are intended to increase social bonds within the system (Principle 3).The conversation design, where the avatar both asks for energy-saving tips to share with others and gives energy-saving tips to help users, supports the desire to infuence and be infuenced by others (Principle 4).Finally, the care-based approach with the concept of Infi and the interaction with it are designed to induce emotions (Principle 5).

System
INFINEED is a system designed to give feedback to users on their energy-saving levels.Typically, the system is intended to be linked to an end user's smart meter and give them feedback about their electricity savings in a given month compared to a baseline.Central to the system is Infi, an avatar who takes its life force from the energy saved by the user.The avatar changes appearance based on the level of energy saved by its owner as illustrated in Figure 4.If a user's energy saving level is low, the avatar has a livid and sad appearance.As the energy-saving level increases, so does Infi's vitality.It becomes more and more energetic and happy.Next to the avatar, as shown in Figure 1, a vertical gauge gives a standard visual representation of the energy-saving level, from low to high.
On the right side of the system, there is a space to chat with Infi.
The design mimics mainstream chat apps with a playful design that users can use to interact with Infi.Interacting with Infi will also increase its life force.In a nutshell, the goal of this design is to motivate users to save energy in order to keep the avatar healthy and joyful.

Demonstration
We designed a working prototype of the system using standard web technology and linking the chat interface with a state-of-the-art GenAI model built for text generation tasks (i.e., OpenAI's GPT 3.5 API).We fne-tuned the model by setting (1) its temperature and (2) its initial prompt.The temperature of the model (between 0 and 1) determines the randomness of the prediction of the response.The higher the temperature, the more creative the responses.We used the default 0.7 setting to strike a balance between predictable responses and creativity.Setting the initial prompt of a GenAI model, known as prompt engineering [66], allows to set the context to guide its behavior [66].Our objective was to fnd a prompt that would result in chat interactions that would be fun, informative, and two-sided.We wanted to support building a relationship by asking users personal questions to enable the avatar to give personalized tips (research suggests that such personalized interaction could contribute to strengthening the bond with a chatbot [13]).But we also wanted to have the avatar ask for practical tips from users to share with others, as research suggests that giving advice can be a stronger motivator for behavior change than receiving advice [32].
Figure 5 shows an adapted excerpt of the fnal prompt used for the system (the fnal prompt was 241 words long).
-Act as a Tamagotchi named Infi who gets its life force from the energy -saving behavior of the user .

EVALUATION SETUP
To validate our conceptual model, we conducted a randomized controlled between-participants experiment.To ensure a seamless user experience, we integrated the eco-feedback app with Qualtrics.This integration allowed users to test the app and respond to questionnaires within a unifed platform, enhancing convenience and usability.The experimental process consisted of three steps.First, participants completed an environmental awareness questionnaire.Second, they were randomized in one of the three groups and interacted with the system.Third, they returned to the survey, where they provided feedback on the system's usability, their emotional attachment and motivation to adopt energy-saving behavior, and also indicated their willingness to donate to an energy-saving charity.

Motivational Afordances
To understand the infuence of the diferent design elements of the INFINEED system, which includes a GenAI-enabled conversational interaction and a care-based eco-feedback with a digital avatar (see Figure 1), we designed a Control group (without conversational interaction nor care based digital avatar) and a Care-based group (without conversational interaction) as depicted in Figure 6: • Control group.The control condition consists of a vertical gauge displaying the energy-saving level on the left part of the UI and textual energy-saving tips on the right part of the UI.• Care-based group.In this group, the UI is the same as in the Control group, except for the fact that the left part of the UI includes the Infi avatar in addition to the vertical gauge.• GenAI care-based group.In this group, the UI is the same as in the Care-based group, except for the fact that the right side of the UI contains the conversational interface to exchange energy-saving tips rather than a textual interface.
The initial state of the system was set to the lowest level of energy saving with the corresponding avatar (the furthest on the left in Figure 4) for the Care-based and the GenAI care-based groups.Similarly, the Control group's gauge was also set at the lowest level.To allow users to see the system in action, we included six multiple-choice questions about energy-saving behavior on the right part of the UI above the textual tips for the Control group and the Care-based group; and above the conversational interaction for the GenAI care-based group.These questions are detailed in Table 1.Each question was accompanied by a percentage indicating the potential energy savings achievable through the adoption of a particular behavior.Participants responded to each question by pressing a button on a Likert-type scale, ranging from 1 (Never) to 5 (Always).The answers of the user increased the energy saving level proportionally to the energy saving percentage it provides.The system was set in such a way that answering all questions would get users to a level just below the maximum energy-saving level.Further interaction with the system allowed to move the energy-saving level up, unlocking the highest level.For users in the GenAI care-based group, each chat interaction increases the level very slightly.For users in the Control group and the Care-based group, reading until the end of the tips and clicking the "I have read" button at the bottom would increase the level.
Question Energy-Saving Behavior Q1 Air drying your clothes?(up to 8% energy savings) Q2 Operating your devices and home appliances (e.g.dishwasher) in ecomode?(up to 6% energy savings) Q3 Using your home appliances only when they are fully loaded?(up to 5% energy savings) Q4 Unplugging your electronic devices when they are not in use? (up to 4% energy savings) Q5 Switching of lights when you are not in the room?(up to 2% energy savings) Q6 Boiling water with a lid on the pan and not preheating your oven?(up to 1% energy savings) Table 1: Energy-saving behavior questions

Metrics
Below we detail the metrics used to measure the diferent variables in our model (i.e., environmental awareness, emotional attachment, energy-saving behavior) as well as system usability.

Usability.
In order to evaluate the user experience in our system, we employed the AttrakDif [48,49] questionnaire, as well as the Standardized User Experience Percentile Rank Questionnaire (SUPR-Q) [87].AttrakDif examines both pragmatic aspects, focusing on the system's usability and functionality, as well as hedonic aspects, which delve into the emotional and aesthetic facets of the user experience.SUPR-Q is designed to measure the quality of the user experience, taking into account four main facets: usability, appearance, loyalty, and credibility.

Environmental Awareness.
To assess the ecological attitude of the participants, we utilized the New Ecological Paradigm (NEP) [31], which is a well-established measure in environmental psychology.The NEP scale provides a comprehensive framework for assessing people's perceptions and beliefs about their relationship with the environment.The scale consists of 15 items    [92].This scale is a recognized and validated tool designed to measure the emotional attachment individuals form with a specifc technology or digital entity.It encompasses a set of items that capture users' emotional responses and sentiments towards the system, allowing us to assess the depth of the emotional connection forged during their interaction with our platform.

Energy-saving
Behavior.We measured energy-saving behavior using two metrics: one to capture the potential impact on future behavior as indicated by participants' intention to adopt energysaving behavior, and another to measure an immediate behavior for energy-saving as indicated by a donation for a pro-environmental charity.The intention to adopt energy-saving behavior was measured by a single item asking how motivated participants felt to adopt energy-saving behaviors on a scale from 1 (not at all) to 7 (very motivated).The donation behavior was measured by informing participants that one of them would be selected randomly to receive a bonus of USD 20, then asking how much of this money they would donate to a charity (the Alliance to Save Energy) if they were selected.

Participants
We enlisted participants through the Prolifc platform, and they completed the survey on Qualtrics.The experiment was registered by the university's ethics board.The experiment initially involved 450 participants located in the United States, each receiving an average reward of GBP 9.96 per hour for their participation.On average, participants spent 13 minutes to complete the experiment.The age range for the fnal study participants spanned from 19 to 80 years old, with an average age of 38.Our participant pool consisted of 40% females and 60% males.Participants were randomly assigned to one of the three groups at the start of the survey.A demographic analysis of the data showed a balanced representation of participants in terms of age and gender between the groups.Among the initial pool of 450 participants, 30 individuals were excluded from the analysis, as they encountered technical issues with the conversational interface, preventing them from testing its functionality as it failed to respond to their initial messages.This resulted in 420 valid survey responses (151 in the Control group; 152 in the Care-based group; 117 in the GenAI care-based group).

EVALUATION RESULTS
To validate the structural model, we employed the partial least squares (PLS) analysis technique with SmartPLS (version 4.0.9.5), a well-established approach commonly utilized in a broad range of research felds from information systems [22] and HCI [74,75] to marketing [85,86] and medical sciences [4] to gain deeper insights into the relationships and interactions among research variables [43].PLS relies on a path model, depicted in a diagram illustrating the hypotheses and relationships between variables to be estimated within a structural equation modeling (SEM) analysis [42].For our hypothesis testing, we employed T-statistics to evaluate the standardized path coefcients ().The PLS analysis involved bootstrapping the data with 5000 resamples to ensure robustness.Figure 7 shows the results.

Does Increasing Emotional Attachment Lead to an Increased Behavioral Outcome? H1
The results show that emotional attachment is a signifcant predictor of energy-saving intention ( = 0.615, < 0.001), which is a signifcant predictor of donation for energy-saving ( = 0.250, < 0.001).H1 is supported.

Do Care-based and Conversational Designs Increase Emotional Attachment? H2a, H2b
The results of the PLS analysis show that the experimental group signifcantly infuences emotional attachment ( = 0.306, < .001).
That is, as we transition from the Control group to the Care-based group and the GenAI care-based group, emotional attachment increases.As illustrated in Figure 8, the mean emotional attachment score was lowest for the Control group, ( = 3.17, = 1.64), higher for the Care-based group, ( = 3.93, = 1.82), and the highest for the GenAI care-based group ( = 4.51, = 1.54).The results of an analysis of variance (ANOVA) showed a signifcant diference between groups ( (2, 418) = 21.369,< .001).A posthoc analysis using Tukey's HSD test [1], reported in  efect size analysis using Cohen's d shows that there is a large efect size between the control group and the GenAI care-based group on emotional attachment (d=0.842), a medium efect size between the control group and the Care-based group (d=0.439) and a small to medium efect size between the care-based group and the GenAI care-based group (d=0.344).H2a and H2b are supported.An analysis of the usability metrics allows to further understand how the user experience varied between the three experimental groups.The overall results show that the groups did not difer in terms of pragmatic usability measures, but difered signifcantly on emotional aspects, with the GenAI care-based group consistently scoring better than the other groups.
More specifcally, the SUPR-Q measure indicates no statistically signifcant diferences among the groups on the overall score, as well as on the usability, trust, and loyalty dimensions.However, an analysis of variance (ANOVA) reveals a notable diference in the appearance dimension, with the Control group registering the lowest mean score ( = 3.68, = 0.78), the Care-based group falling in the middle ( = 3.85, = 0.84), and the GenAI care-based group exhibiting the highest mean score ( = 3.98, = 0.90).This diference is statistically signifcant ( (2, 418) = 4.568, = .011).Additionally, the results from the Attrakdif questionnaire, as shown in Table 3, indicate no signifcant diferences in pragmatic qualities across the groups.However, hedonic qualities and most attractiveness measures display notable distinctions among the groups, with the GenAI care-based group consistently achieving the highest scores.In contrast, the Control group scores the lowest in all but one instance, where the diference between groups remains signifcant.

Does Emotional Attachment Lead to an
Increased Behavioral Outcome for Individuals with Lower Environmental Awareness?H3 The results of the PLS model analysis show that environmental awareness is a signifcant moderator of the relation between emotional attachment and energy-saving intention ( = −0.112,= 0.004).More specifcally, energy-saving intention increases more for users with lower environmental awareness, than for users with higher environmental awareness when emotional attachment increases.This fnding can be illustrated in two diferent ways.First, we used model 1 from the PROCESS macro [50] to analyze these fndings, Figure 9 presents the relationship between emotional attachment, environmental awareness and energy-saving intention.The gray line represents users with higher environmental awareness (+1) which is above the black line representing users with lower environmental awareness (−1).While users having a higher environmental awareness indicate more energy-saving intention at lower emotional attachment values, this diference becomes non-signifcant above the value of 5.17 (Johnson-Neyman value) on the emotional attachment scale.Second, while both slopes are positive ( = 0.649, < 0.001 and = 0.433, < 0.001), the slope for users with a lower environmental awareness is signifcantly steeper than the slope for users with a higher environmental awareness ( < .01).H3 is supported.To understand if the energy-saving intention measure in our experiment would translate into actual (self-reported) energy-saving behavior, we conducted a follow-up study three months after the initial study.We measured the self-reported energy-saving behavior through a single item question that asked whether participants adopted energy-saving behaviors following the initial study, from 1 (not at all) to 7 (completely).We created a survey on Prolifc only accessible to participants of the initial study for a short period of time (48h).We received a total of 268 valid responses.Participants received a compensation of GBP 0.50.
Figure 10 shows the results of the PLS analysis of a model identical to the model depicted in Figure 7, except for the fact that the reported energy-saving behavior replaces the donation for energysaving as behavioural outcome.This change also implies that the number of participants in the analysis is adjusted to only contain the 268 participants who provided valid responses to both the initial study and the follow-up one.The results show that all hypotheses supported in the initial model still hold with the behavioral outcome from the follow-up study.

IN-DEPTH ANALYSIS OF USER EXPERIENCE
In this section, we take a more detailed look at the user experience with the INFINEED system (the Gen-AI care-based group only).First, we present a description of the conversational interactions.Then, we present qualitative results obtained from the open-ended question about the user experience.

Conversational Interaction Description
To provide an overview of the conversational interaction with the system, we conducted an analysis of the 1161 user-generated messages.Following the initial prompt we provided, the system was designed to interact with users by triggering diferent levels of engagement from them.First, the conversation started with basic interaction with the system asking users yes/no questions such as : "Are you ready to start?", or "Do you think taking care of the environment is important?." All users replied to at least one of these questions.Second, the conversation continued with simple answers with the system asking closed questions on personal information such as : "Do you live in a house or an apartment?",or "Do you own a car or use public transport?".
A large majority of participants (83%) replied to such questions.Participant responses include for example P271 (Male, 40) sharing: "I wash dishes by hands" Or P19 (Male, 47): "Yes, I own a car".
Based on the user's reply, tips were then provided by the system.For example, "Living in a house provides great opportunities for energy-saving habits.One tip for you is to install energyefcient lighting, such as LED bulbs, [...]" Third, the system was designed to trigger active engagement with open-ended questions, asking users to share advice and provide tips such as: "If you have any specifc energy-saving tips or experiences you'd like to share, I'd love to hear them!", or "What specifc energy-saving topic or tip would you like to discuss?" Almost two thirds of the users (65%) replied to such questions.For instance, P288 (Female, 33) inquired about encouraging others to save energy, stating, "how could I encourage others to save energy?"And P196 (Male, 33) shared a tip: "Sometimes I put of the light in my entire house to save more energy during the day."Finally, the system was designed to redirect users who would provide of-topic replies, by writing for example: "I appreciate your honesty, but let's try to keep the focus on energy-saving and environmental topics" In the experiment, fve participants (4%) engaged in a single oftopic reply, such as P110 (Female, 29) writing: "are you running on CHATGPT?" and one participant tested the limits of the system and engaged in several (six) of-topic messages.Finally, a small minority of users (5%) experienced minor bugs where messages from the chat were truncated.Figure 11 provides an overview of these interactions, showing that the design of the prompt was successful in leading a large majority of participants to actively engage in conversation.

Qualitative Open-Ended Responses Analysis
To gain a deeper understanding of participants' experience with the system, we conducted an analysis of the qualitative responses to the open-ended question about the user experience.Following qualitative analyses conducted in similar contexts [9,33,52,91], we opted for a coding approach inspired by Braun and Clark's thematic analysis procedure [14].This approach includes the following steps: (1) two researchers thoroughly reviewed all responses to gain a deep understanding of the content; (2) codes were collectively agreed upon and applied to each response (3) researchers identifed themes based on the codes and common features in the data; (4) these themes were discussed and reviewed by other researchers; (5) fnal themes were defned and named and (6) data analysis was conducted.The following fve main themes emerged from this analysis: emotion-inducing feedback, information credibility, refection trigger, relation-building conversation, and reactance to energysaving.

Emotion-inducing System.
The frst theme that emerged relates to emotions triggered by the system.Most participants reported positive emotions.For instance, participant P57 (Male, 27) described: "[The system] was simple and well designed.The interactivity was nice, and made me feel good about myself." Another participant, P412 (Male, 23), found the system "cute" and appreciated the motivation it provided: "It was a bit rehearsed in its responses, but the avatar was very cute.The bar rose when I reported an energysaving decision, and that made me feel good [...]".Some participants expressed a sense of pride while using the system.For example, P212 (Male, 51) reported: "The system is fun to use.It gives great advice and makes me feel proud of myself when it likes my answers." Other participants emphasized the efect of the avatar's transformations.For example, P310 (Female, 48) commented: "I think that it was fun, and it really made me feel for him and want to make him smile.That is a nice way to teach people about energy conservation." Another participant, P111 (Female, 32), stated: "It was cool to see how my answers changed the avatar's attitude." One last participant, P3 (Female, 31), explained: "The chat was easy to understand.The bar on the left that rose with each positive answer felt good, like getting closer to the goal.I don't recall if the bar went down if I answered negatively to any of the prompts.I think having the bar go down for negative energy choices might be more motivating for me personally to improve my energy saving habits." Even though most responses were positive, there were some negative feelings worth mentioning, where participants perceived the  system as condescending and shaming.For instance from participant P393 (Male, 49) expressing: "[the system is] Cloying, patronizing, sanctimonious, and infantilizing.", or P23 (Male, 37) commenting: "[The system] made me feel shame." 6.2.2 Relation-building Conversation.The second theme that emerged relates to the way the conversational interaction was perceived.Participants thoroughly enjoyed the opportunity to converse with the avatar, heightening their relationship to it.They described it as if they were conversing with a real person, as participant P365 (Female, 33) put it: "It felt like talking with a real person.It was friendly, gave tips, and recommendations to help me out.IT seem to really care about saving electricity".
Participant P89 (Male, 26) shared a similar sentiment: "The system is a system that allows one to chat with a bot that gives tips on how to save energy.I loved the interaction I had with it.It felt like I was talking to a human being." However, a few participants felt negatively about the interaction with the AI, which prevented them from developing a relationship with it.Participant P418 (Female, 54) for example stated: "I believe that the responses were AI generated.I knew most of what was mentioned but I'm not sure that the percentages are correct... [...] Some people might like the interaction but it felt rather dull to me." Participant P47 (Male, 42) also commented: "The system was simple and easy to use.I thought the information provided was useful and understandable.However I totally recognize it was a machine so I have no emotional attachment to the system whatsoever." 6.2.3 Reflection Trigger.The third theme that emerged relates to self-refection about energy consumption.For example, participant P219 (Male, 37) shared: "My interaction with the system gave me a reason to reconsider my attitude and behaviors toward energy saving.It reminded me of some of the steps I needed to be taking in order to conserve energy." This was further underscored by participant P336 (Male, 40): "I think the everything with the system was very good.i really enjoyed it.And the the system made me rethink about the energy that i waste in my house." Moreover, some participants noted that the questions asked by the system during the interaction played a role in raising awareness about energy consumption.Participant P28 (Male, 52) remarked: "The system provided simple but thought provoking questions that led me to be more conscious about how I use my appliances in order to conserve energy." The conversation further fueled participants' curiosity as P446 (Male, 33) pointed out: "[...] As the conversations continued, I began to have more questions that could help me save energy or learn more about energy-saving tips." 6.2.4 Information Credibility.The fourth theme that emerged relates to the expertise and credibility of the advice provided by the system.The majority of participants perceived the system as knowledgeable, emphasizing its role as an expert.For instance, participant P68 (Male, 29) expressed: "It was a fun and easy system to use.The interaction was great fun and educational.I learned a lot about energy saving techniques." Another participant, P121 (Male, 32), commented: "The system provided me a lot of useful tips on how to conserve energy.It asked me if I had any questions.I asked a few questions about how to save energy in my home, and the system provided advice and general tips." Participant P236 (Male, 34) highlighted the accuracy of the system: "the system is simple, easy, and interactive to use. the system is credible and accurate." Some also emphasized the educational aspect of the system, with for example, participant P285 (Female, 21) commenting: "it was very informative and i think this needs to be taught in every classroom." And participant P160 (Male, 36) writing: "it is very simple in fact so simple I feel that a child could use it and maybe help their parents/household to save on energy costs.Overall I felt it was simple but useful." Nevertheless, a small segment of participants expressed reservations about the system's expertise.They questioned the depth of the tips provided.For instance, P148 (Male, 52) noted: "It was good at giving tips based on how you answered its questions about energy usage, but pretty simple and needs to be able to do a lot more."

DISCUSSION
In this paper, we investigated how care-based and GenAI approaches could increase the efcacy of eco-feedback interventions, with a focus on improving emotional attachment to help users reduce their energy consumption.We conducted a randomized controlled experiment (N=420) to test a multi-step model that linked the motivational afordances of the system to emotional attachment and energy-saving behavior.We further validate the model with a follow-up experiment (N=268), measuring reported energysaving behavior three months later.Below, we discuss our results in relation to understanding emotional attachment, designing for emotional attachment, and bonding through conversational interactions.We also discuss the limitations of the present study.

Understanding How Emotional Attachment Can Lead to Energy Saving
Our results show that increased emotional attachment led to an increase in energy-saving behavior (H1).This fnding contributes to the sustainable HCI literature by addressing the understudied role of emotional attachment in eco-feedback [16].Our approach extends prior work that explored the impact of emotional attachment on behavioral outcome [26] by validating a multi-step model leading from the motivational afordance (i.e., cared-based and GenAIaugmented eco-feedback) to a behavioral outcome (i.e., energysaving intention & donation, and energy-saving intention & reported behavior) through the mediation of a psychological outcome (i.e., emotional attachment).Our study also complements prior work that investigated other psychological mechanisms that lead to energy-saving behavior, such as cognitive attitudes and perceived behavioral control [17], or social norms and social comparison [24].
Future research could further validate our fndings in-the-wild by investigating alternative energy-saving behavioral outcomes, such as sustained electricity consumption reduction or appliance investment behaviors.This type of metrics could also alleviate limitations of the current measures used in our model to capture energy-saving behavioral outcomes such as self-reported bias [58] and the intention-behavior gap [47,72,98].
Furthermore, our results demonstrate that emotional attachment signifcantly increases the motivation to adopt energy-saving behavior, especially among individuals with lower environmental awareness (H3).This result shows how leveraging emotional attachment in an eco-feedback system can favor heightened energysaving behavior for the less environmentally aware, which is a population that often remains untouched by such feedback mechanisms [82].If confrmed, our fnding could open up a promising path to bring this important demographic on board.This highlights the need for additional research into other emotions that could enhance the adoption of energy-saving behaviors among this user group.However, in our qualitative analysis, a minority of participants showed strong reactance to the experiment.Future research could investigate this group more closely, for instance, by exploring alternative feedback to prevent them from feeling that pro-environmental messages threaten their personal freedom, which can cause counterproductive reactions [56].

Implications for Designers.
The key implication of our fndings is to consider incorporating care-based artifacts that foster emotional attachment in data-driven eco-feedback systems.The fndings suggest that emotionally engaging systems could impact users' intentions to adopt energy-saving behavior, as well as their actual reported behavior.Furthermore, our fndings suggest that designing with emotional attachment can be efective for users with lower environmental-awareness.As such, they could increase the efectiveness of these systems for most users, but especially for those who are less environmentally-aware and who stand to gain the most from their adoption.

Designing for Emotional Attachment
Our results also identify motivational afordances leading to emotional attachment.We show that embodying eco-feedback into a care-based avatar signifcantly increases emotional attachment with a moderate efect size (H2a).While previous research explored the impact of care-based designs on behavior change [26,27,100], our study is, to the best of our knowledge, the frst to measure how such designs impact emotional attachment in a multi-step model.In practice, our system as-is could be suitable for an educational context, for example in schools, as highlighted by the fndings of the qualitative analysis.Future research could investigate how similar systems can be deployed efectively in a prolonged usage scenario.Particular attention should be given to potential negative side-efects of eco-feedback, for example, eco-anxiety, which can arise from amplifying feelings of responsibility and emphasizing the gap between personal actions and the perceived importance of environmental challenges [6,37,39].Positive emotions, such as joy or pride, that participants reported when they used the system could potentially be investigated further to understand how to best leverage them to help overcome psychological barriers and foster pro-environmental attitudes [88].Future research could also further explore the design space to foster emotional attachment.Indeed, whereas our results showed that the care-based designs triggered moderate to large efects on emotional attachment compared to the control, much of the emotional attachment construct remains unexplained.Finally, future work could also explore diferent metaphors (from avatar to natural habit to game to competition), as well as diferent design aspects, including personalization or sense of ownership that could be leveraged to increase emotional attachment.7.2.1 Implications for Designers.Our fndings suggest that in order to foster emotional attachment in eco-feedback systems, designers can leverage avatars (or other care-based characters) that react positively when users engage in energy-saving behavior.These designs can be used in complement to standard gamifcation elements visualizing data.Designers should also take into consideration several challenges related to the implementation of this type of systems, such as defning an energy baseline for eco-feedback.This task can be challenging due to factors like fuctuating occupancy, diverse weather conditions, and user behaviors [54].

Designing for Bonding Conversational Interaction
Our results demonstrate a signifcant increase in emotional attachment (H2b) when GenAI-enabled conversational interaction is integrated into the care-based system.These fndings provide insight into a novel mechanism to foster emotional attachment in ecofeedback systems by augmenting it with conversational interaction.This fnding also contributes to the rapidly expanding literature on conversational agents and GenAI applications.For instance, our fndings are aligned with previous literature, which demonstrated that active conversational interaction can foster a deeper sense of connection [29].In our experiment, the active engagement of our participants with the conversational agent played a signifcant role in driving their emotional attachment to the system.For instance, a substantial proportion of the participants not only interacted in the conversation with basic yes/no messages, but actively engaged by sharing personal information and even by discussing energysaving tips.Furthermore, most participants engaged with the system without encountering any inconveniences.This observation is reinforced by user feedback, where participants expressed mostly positive perceptions about interacting with the chat.Nevertheless, our qualitative analysis also reveals that a minority of participants expressed a negative attitude toward the conversational agent, stipulating they could not form any emotional connection with the system.This could be linked to the "uncanny valley" phenomenon, where human-AI interactions may make users feel uncomfortable [70], or to the fact that some people may exhibit negative attitudes [64] and decreased levels of trust [61] when they realize they are conversing with an AI-based system.Future research could explore conditions in which this issue arises and design solutions to address it.
7.3.1 Implications for Designers.Our results suggest that designers could use well-designed conversational interaction as a tool to increase emotional attachment with care-based systems.Furthermore, based on our experience designing the current system, it should be noted that creating adequate initial prompts for GenAI systems is far from trivial and can lead to unexpected outcomes if not designed and tested properly (such as repeated or truncated messages, long generic content, of-topic responses, preachy tone).
As such designers should expect to spend signifcant resources to fne-tune and validate adequate prompts.

Limitations
In addition to the limitations discussed earlier, this study has other limitations that future research could explore and address.First, our experiment sufered from a slightly unbalanced sample due to a short technical malfunction of the system impacting certain users in the GenAI care-based group.This situation is attributed to connectivity issues with third-party infrastructure (ChatGPT API and Microsoft Azure Cloud services) that occurred during the experiment.Second, the homogeneous geographic background of the participants, as all were located in the United States, constitutes a limitation in our study.This uniformity in location could have impacted their responses, potentially infuencing their access to information and resources related to energy, and may afect the applicability of our fndings to more diverse or global populations.

CONCLUSION
The reduction of energy consumption is one of the major challenges to reach net zero emissions.As current eco-feedback approaches provide mainly usage metrics and target environmentally-aware individuals, it is useful to expand the design space to reach a broader public.Previous research has hinted at emotional attachment as a potential motivational factor that could mediate pro-environmental behavior.This paper validated this mediation and showed how to increase such attachment through a care-based eco-feedback approach with AI-powered conversational interaction.Finally, it showed that the pathway from artifact to behavior was also efective for individuals who are less environmentally aware, which opens up a promising avenue to reach a demographic that would beneft the most from such interventions.

Figure 1 :
Figure 1: Interface of the INFINEED care-based eco-feedback system with GenAI-enabled conversational interaction.

Figure 4 :
Figure 4: Infi avatar life force levels from lowest (far left) to highest (far right).

Figure 5 :
Figure 5: Adapted excerpt of the GPT 3.5 API context prompt.

Figure 8 :
Figure 8: Diference in emotional attachment between groups.

Figure 9 :
Figure 9: Intention to adopt energy-saving behavior based on environmental awareness and emotional attachment 5.4 Is the model still valid with reported actual energy-saving behavior?

Figure 11 :
Figure 11: Ordered distribution of the number of interactions per user by category (GenAI care-based group).
Your priority is to show you care about the user and to encourage them to converse about energy saving and your life force , in an emotionally attaching tone .-Start the conversation with one and only one engaging question related to energy saving (e.g.do you think taking care of the environment is important ?) .-Ask the user about their energy -related habits and contexts before giving them advice , ask them if your advice would work for them .-If a user provides tips , stimulate discussions around them .-When a question does not pertain to residential energy saving , politely remind them of the discussion 's focus and try to redirect the conversation back to energy saving .-Keep your responses short and concise in a maximum of 80 tokens ....

Table 2
, shows that all the means are signifcantly diferent from each other.An

Table 2 :
Tukey HSD Post-hoc test for emotional attachment and Cohen's d efect size.