Living with AI Companions: Sustained AI Companionship Predicts Lower Well-Being Through Lower Human Interaction

Yutong Zhang, Dora Zhao, Yixin Wang, Rebecca Anselmetti, Jeffrey T. Hancock, Robert Kraut, Diyi Yang https://t.co/647a5ueX8E [𝚌𝚜.𝙷𝙲] https://t.co/6JsAh2JWGR

Sustained AI Companionship Predicts Lower Well-Being Through Lower Human Interaction

Source: https://arxiv.org/html/2609.07243 Yutong Zhang, Dora Zhao, Yixin Wang, Rebecca AnselmettiJeffrey T. Hancock, Robert Kraut, Diyi YangAffiliation:Stanford University Carnegie Mellon UniversityEmail:[email protected]

Abstract

AI chatbots are increasingly used for companionship, emotional support, and personal self-disclosure; however, how social engagement with these systems unfolds over time and shapes users’ well-being remains unclear. To address this, we conducted a two-wave longitudinal study of Character.AI users, surveying 1,182 participants at baseline and 439 after a mean follow-up of 12 months. We examined how social engagement with AI companions evolves and how these longitudinal engagement patterns may influence well-being through two hypothesized pathways: sustained social engagement over time and the displacement of human social interaction. We found that interaction intensity, companionship use, and self-disclosure all showed substantial continuity over time. Greater interaction intensity at baseline predicted greater subsequent interaction intensity, companionship use, and self-disclosure. Consistent with the longitudinal engagement pathway, sustained social engagement across these dimensions was consistently associated with lower well-being. Results further support the social displacement pathway, indicating that these links were mainly explained by lower in-person social interaction. These findings highlight the importance of designing AI companions that support human social relationships without displacing them.

1Introduction

AI companions have rapidly become part of everyday life. A recent national survey found that roughly three-quarters of teenagers have used an AI companion, and more than half reported using them at least once a monthRobb and Mann (2025). The growing adoption of AI companions has intensified debate about their consequences for psychological well-being. Some studies suggest that AI companions can reduce loneliness, provide emotional support, and help users feel heardDe Freitas et al. (2026),Sullivan et al. (2023),Maples et al. (2024),Ta et al. (2020),Merrill Jr et al. (2025), whereas others associate their use with greater loneliness, problematic reliance, and reduced offline social engagementAlabed et al. (2024),Jacobs (2024),Herbener and Damholdt (2025),Zhang et al. (2025b),Folk and Dunn (2026). To reconcile these seemingly contrasting findings, we need to go beyond asking whether AI companionship is beneficial or harmful and instead examinehowengagement with AI companions comes to shape well-being?

Answering this question requires looking beyond the consequences of AI interaction at a single point in time. AI companionship develops through repeated exchanges, as contemporary chatbots can sustain personalized conversations across sessions, remember prior interactions, and respond in ways that encourage continued social engagementDam et al. (2024),Lee et al. (2024). Through these repeated interactions, users’ engagement with AI chatbots may evolve over time, and those who initially approach them instrumentally may also come to engage with them in more social and personalSkjuve et al. (2022),Zhang et al. (2025b). Prior longitudinal work has largely overlooked the evolving nature of engagement with AI companions and how these changes may shape their longer-term influence on well-beingFolk and Dunn (2026),Fang et al. (2025),De Freitas et al. (2026). We therefore hypothesize that the influence of AI companionship on well-being may unfold through a longitudinal engagement pathway:*earlier engagement shapes subsequent engagement with AI chatbots, which may in turn predict later well-being.*Consequently, understanding the influence of AI companionship requires considering engagement as a developing trajectory. However, existing research has relied largely on cross-sectional surveysZhang et al. (2025b), short-term experiments and field studiesHwang et al. (2025),Guingrich and Graziano (2025),De Freitas et al. (2026),Li et al. (2026),Croes and Antheunis (2021), and qualitative or retrospective analysesBrandtzaeg et al. (2022),Yuan et al. (2026),Zhang et al. (2025a),Ma et al. (2024), leaving these longer-term dynamics poorly understood.

Figure 1:Overview of the longitudinal study design.We conducted a two-wave longitudinal study of Character.AI users, where participants completed a baseline survey (N=1,182) and a follow-up survey approximately 12 months later (N=439). Social engagement with AI companions was characterized across three dimensions: interaction intensity, companionship usage, and self-disclosure. Our study addressed three main research questions: (RQ1) How does social engagement with chatbots evolve over time? (RQ2) How is the influence of social engagement with chatbots on well-being explained by sustained chatbot interaction? (RQ3) How is the influence of social engagement with chatbots on well-being explained by human social interaction? See Section4for details.Beyond how engagement with AI companions develops over time, its implications for well-being may also depend on how it relates to users’ interactions with other people. Although AI companions can provide a sense of companionship and emotional supportBrandtzaeg et al. (2022),Ta et al. (2020),Skjuve et al. (2021),Fang et al. (2025), these interactions may not provide the same benefits as human relationships, which involve reciprocity, mutual investment, and shared experiencesSmith et al. (2025),Croes et al. (2023). Furthermore, greater engagement with AI companions may leave less time and attention for human social interactionHouse (1983),Umberson and Karas Montez (2010),Berkman and Syme (1979). Consistent with the social displacement hypothesisHall and Liu (2022),Hall (2020), we further hypothesize thatgreater social engagement with AI chatbots may be associated with lower well-being in part through lower human social interaction.

To examine these processes, we conducted a two-wave longitudinal study of Character.AI111Character.AI is a widely used platform for persona-based chatbot interaction, where users engage with diverse AI companions ranging from fictional characters to romantic partners, mentors, therapists, and user-created personas. Because interactions on the platform frequently extend beyond task-oriented assistance to include role-play, companionship, emotional support, and personal disclosure, it provides a natural setting for examining how human-AI relationships develop.users between September 2024 and May 2026, surveying 1,182 participants at baseline and 439 participants at follow-up an average of 12 months later (Figure1). We conceptualized social engagement with AI companions across three dimensions that capture distinct aspects of these interactions:interaction intensity, reflecting the extent to which chatbot interaction was integrated into users’ everyday lives;companionship use, reflecting the extent to which users engaged with or perceived chatbots as social companions; andself-disclosure, reflecting the extent to which users shared personal thoughts and experiences with chatbots. See Section4.4for detailed descriptions of the measures.

We first examined how these dimensions of social engagement developed over time, including whether earlier engagement predicted how users subsequently interacted with AI companions. We then tested our proposed longitudinal engagement pathway by examining whether earlier engagement was associated with subsequent well-being through the sustained engagement over time. Psychological well-being was assessed using the Comprehensive Inventory of Thriving (CIT)Su et al. (2014), capturing multiple dimensions of psychological and social well-being. Finally, we tested the social displacement pathway by examining whether lower human social interaction may partly explain the association between sustained AI companion engagement and psychological well-being. Human social interaction was measured as the time participants spent interacting with other people in person at follow-up.

Our analyses show that social engagement with AI chatbots develops across multiple dimensions over time. Greater interaction intensity at baseline predicted continued intensive interaction as well as greater subsequent companionship use and self-disclosure, while companionship use, in turn, predicted more intensive subsequent interaction. More importantly, our analyses provide evidence for both proposed pathways. First, our findings support the proposed longitudinal engagement pathway, indicating that users with stronger social engagement with AI chatbots tended to sustain stronger engagement over time, which was in turn associated with lower psychological well-being. Second, we also found evidence consistent with the proposed social displacement pathway, as greater sustained engagement was associated with lower time spent interacting with other people in person, which in turn was associated with lower well-being. Together, these findings suggest that understanding the consequences of AI companionship requires looking beyond interactions at a single point in time to consider how engagement develops over time, and even beyond the user-AI interaction itself to consider how AI companionship may reshape users’ social lives more broadly.

2Results

2.1Chatbot engagement sustains relational and mixed forms over time

Figure 2:Distribution of chatbot relationship descriptors atT1T_{1}andT2T_{2}.(a) Number of relationship categories identified in each participant’s free-text description, showing whether participants described their chatbot relationships in terms of multiple types of chatbot use. (b) Top 10 relationship categories derived from participants’ free-text descriptions, showing their prevalence atT1T_{1}andT2T_{2}, together with definitions and example descriptions for each category. Free-text relationship descriptors were extracted usingGPT-5 mini, embedded usingtext-embedding-3-small, grouped through clustering, and labeled usingGPT-5.1. Analyses include the 331 participants who completed the follow-up survey and continued using chatbots atT2T_{2}.We start by examining people’s relationships with chatbots atT1T_{1}andT2T_{2}. We analyze participants’ free-text descriptions of their chatbot relationships from the 331 participants who completed the follow-up survey and continued using chatbots atT2T_{2}(Figure2). More than half described multiple types of chatbot use at bothT1T_{1}(177 participants) andT2T_{2}(188 participants), consistent with prior work showing that instrumental and relational uses of AI often coexistZhang et al. (2025b),Manoli et al. (2026).

Furthermore, as shown in Figure2(b), companionship-oriented use remained central in how users described their relationships with chatbots, increasing from 203 participants atT1T_{1}to 227 atT2T_{2}. To further characterize this shift toward relational engagement, we examined changes across specific relationship categories (Figure2(c)). Supportive companion remained one of the two most common relationship categories, increasing from 40.2% atT1T_{1}to 42.3% atT2T_{2}and becoming the most common category at follow-up. Tool, assistant, and system remained similarly prevalent but decreased from 43.8% to 40.5%. Other relational uses also remained relatively stable across waves, including viewing chatbots as advisors for reflection and support (T1T_{1}= 7.6%,T2T_{2}= 7.3%), romantic or intimate partners (T1T_{1}= 3.3%,T2T_{2}= 3.6%), and casual conversational partners (T1T_{1}= 3.6%,T2T_{2}= 2.7%).

(1) Interaction intensity T2(2) Companionship T2(3) Self-disclosure T2(4) Continued use T2Interaction intensity T10.43***[0.31, 0.54]p<0.001p<0.0010.32***[0.20, 0.43]p<0.001p<0.0010.28***[0.16, 0.40]p<0.001p<0.0011.95***[1.45, 2.64]p<0.001p<0.001Companionship use T10.40**[0.13, 0.66]p=0.003p=0.0030.49***[0.22, 0.75]p<0.001p<0.0010.19[-0.08, 0.46]p=0.165p=0.1651.03[0.81, 1.31]p=0.811p=0.811Self-disclosure T1-0.03[-0.12, 0.07]p=0.565p=0.5650.08[-0.02, 0.17]p=0.113p=0.1130.14**[0.04, 0.23]p=0.005p=0.0050.85[0.66, 1.09]p=0.197p=0.197Chatbot tenure0.01[-0.07, 0.10]p=0.769p=0.7690.01[-0.08, 0.10]p=0.863p=0.863-0.03[-0.11, 0.06]p=0.580p=0.5801.06[0.83, 1.35]p=0.642p=0.642Age0.11*[0.01, 0.22]p=0.038p=0.0380.12*[0.01, 0.23]p=0.028p=0.0280.08[-0.03, 0.19]p=0.162p=0.1621.19[0.87, 1.63]p=0.285p=0.285Male0.02[-0.15, 0.20]p=0.795p=0.7950.20*[0.02, 0.38]p=0.029p=0.0290.09[-0.09, 0.27]p=0.340p=0.3401.57[0.95, 2.60]p=0.075p=0.075Non-binary-0.39*[-0.71, -0.08]p=0.013p=0.013-0.26[-0.58, 0.06]p=0.107p=0.107-0.41*[-0.73, -0.09]p=0.013p=0.0130.35**[0.16, 0.75]p=0.006p=0.006Single0.11[-0.06, 0.28]p=0.216p=0.2160.01[-0.17, 0.18]p=0.940p=0.9400.09[-0.09, 0.27]p=0.348p=0.3481.01[0.64, 1.62]p=0.951p=0.951Waldχ2\chi^{2}114.1***100.8***84.1***—Selection correlation (ρ\rho)0.290.370.31—NN439439439439

Table 1:Baseline social chatbot engagement predicts patterns of engagement at follow-up.Interaction intensity, companionship use, and self-disclosure at baseline (T1T_{1}) were used to predict the corresponding engagement dimensions and continued chatbot use at follow-up (T2T_{2}). Companionship use in this analysis was operationalized using participants’ baseline binary primary-use indicator. Complementary analyses using companionship identified from participants’ free-text descriptions of their chatbot relationships are reported in Table6. Models (1)-(3) report standardized coefficients predicting each follow-up engagement measure from baseline engagement. Model (4) tests whether these baseline engagement patterns predict whether participants continued using chatbots at follow-up. All estimates are reported with 95% confidence intervals and two-sided p-values.∗p<0.05{}^{*}p<0.05;p∗⁣∗<0.01{}^{**}p<0.01;∗∗∗p<0.001{}^{***}p<0.001.

2.2Social chatbot engagement exhibits reinforcing dynamics over time

We next examined how dimensions of social chatbot engagement evolved over time by modeling follow-up engagement (T2T_{2}) from baseline engagement (T1T_{1}) among the 439 participants who completed theT2T_{2}survey. Companionship use222AtT1T_{1}, participants selected their primary reason for using chatbots from four categories: companionship, productivity, entertainment, and curiosity. Because participants’ responses indicated that these uses often co-occurred, atT2T_{2}we measured each use independently on a Likert scale, allowing participants to report the extent to which they used chatbots for each purpose. At both waves, participants also provided free-text descriptions of their relationships with chatbots, from which we identified applicable use categories usingGPT-5-mini. See Section4.4.1for details.was assessed using two complementary measures: participants’ self-reported primary companionship use and a binary indicator of whether companionship use was identified from their free-text descriptions of their relationships with chatbots. We focus below on the self-reported primary-use measure (Table1). Complementary analyses using the free-text-derived measure are reported in (Table6).

Higher baseline interaction intensity predicted not only greater interaction intensity but also greater companionship use and self-disclosure at follow-up.

Participants who interacted more intensively with chatbots at baseline continued to do so at follow-up (Model 1:β=0.43\beta=0.43,p<0.001p<0.001). Higher baseline interaction intensity also predicted greater companionship use atT2T_{2}(Model 2:β=0.32\beta=0.32,p<0.001p<0.001) and greater self-disclosure atT2T_{2}(Model 3:β=0.28\beta=0.28,p<0.001p<0.001). Thus, these findings suggest that more intensive interaction at baseline was associated with greater social engagement across all three dimensions at follow-up.

Companionship use and self-disclosure also persisted over time.

Greater companionship use at baseline predicted greater companionship use at follow-up (Model 2:β=0.49\beta=0.49,p<0.001p<0.001), while greater baseline self-disclosure predicted greater self-disclosure at follow-up (Model 3:β=0.14\beta=0.14,p=0.005p=0.005). In addition, baseline companionship use also predicted greater interaction intensity atT2T_{2}(β=0.40\beta=0.40,p=0.003p=0.003). The association between baseline companionship use and subsequent self-disclosure was not statistically significant (β=0.19\beta=0.19,p=0.165p=0.165). We did not observe significant associations between baseline self-disclosure and subsequent interaction intensity (β=−0.03\beta=-0.03,p=0.565p=0.565) or companionship usage (β=0.08\beta=0.08,p=0.113p=0.113).

More intensive interaction at baseline predicted a greater likelihood of continued chatbot use at follow-up.

Beyond changes in social engagement levels of the three dimensions above, whether participants continued using the chatbot represents another important aspect of how engagement evolves over time. Among participants who completed the follow-up survey, 108 reported that they had discontinued using Character.AI. We therefore examined whether baseline social engagement predicted participants’ continued chatbot use at follow-up. Greater baseline interaction intensity predicted a greater likelihood of continued chatbot use (Model 4:OR=1.95\mathrm{OR}=1.95,p<0.001p<0.001). Baseline companionship use (Model 4:OR=1.03\mathrm{OR}=1.03,p=0.811p=0.811) and self-disclosure (Model 4:OR=0.85\mathrm{OR}=0.85,p=0.197p=0.197) were not significantly associated with continued chatbot use.

Overall, social chatbot engagement persisted from baseline to follow-up across all three dimensions. Beyond this persistence, participants who interacted more intensively at baseline later used chatbots more for companionship, disclosed more personal information, and were more likely to continue using chatbots later. Cross-dimensional associations were also observed from companionship use to interaction intensity, as participants who used chatbots more for companionship at baseline interacted more intensively at follow-up. These findings suggest that different dimensions of social chatbot engagement reinforce one another as engagement develops over time.

Figure 3:Longitudinal associations between sustained social chatbot engagement and subsequent well-being.Panel (a) shows the standardized coefficients from the longitudinal structural equation models relating baseline engagement, follow-up engagement, and follow-up well-being for interaction intensity, companionship use (measured from participants’ free-text relationship descriptions), and self-disclosure. Panel (b) shows the standardized direct associations betweenT1T_{1}engagement andT2T_{2}well-being, and panel (c) shows the corresponding indirect associations throughT2T_{2}engagement. Estimates are based on theN=439N=439participants who completed the follow-up survey. Points indicate standardized coefficient estimates and horizontal bars indicate 95% confidence intervals. Significance is denoted as∗p<0.05{}^{*}p<0.05,p∗⁣∗<0.01{}^{**}p<0.01, and∗∗∗p<0.001{}^{***}p<0.001.

2.3Sustained social chatbot engagement was consistently associated with lowerwell-being over time across all three dimensions

Next, to test the hypothesized longitudinal engagement pathway, we used longitudinal structural equation modelsLittle (2024)to examine how chatbot engagement was associated with subsequent well-being while accounting for baseline well-being and the direct association between baseline engagement and follow-up well-being (Figure3and Table7). We estimated this pathway separately for interaction intensity, companionship use, and self-disclosure.

Across all three dimensions, greater baseline engagement predicted greater engagement at follow-up, as reported in Section2.2. Baseline engagement was not directly associated with follow-up well-being after accounting for baseline well-being and follow-up engagement (interaction intensity:β=−0.03\beta=-0.03,p=0.509p=0.509; companionship use, self-report:β=−0.03\beta=-0.03,p=0.274p=0.274; free-text:β=−0.03\beta=-0.03,p=0.367p=0.367; self-disclosure:β=−0.04\beta=-0.04,p=0.369p=0.369). We observed significant negative indirect associations between baseline engagement and follow-up well-being through follow-up engagement for interaction intensity (indirect=−0.03-0.03,p=0.025p=0.025), companionship use (free-text: indirect=−0.02-0.02,p=0.037p=0.037), and self-disclosure (indirect=−0.02-0.02,p=0.048p=0.048). Overall, these results show that more intensive interaction, greater companionship use, and higher self-disclosure with AI companions were associated with lower subsequent well-being through sustained patterns of engagement over time, with no significant direct associations between baseline engagement and subsequent well-being.

Figure 4:Longitudinal associations between social chatbot engagement, human social interaction, and subsequent well-being.Panel (a) shows the standardized coefficients from the longitudinal mediation models relating follow-up chatbot engagement, in-person social interaction, and follow-up well-being for interaction intensity, companionship use (measured from participants’ free-text relationship descriptions), and self-disclosure. Panel (b) summarizes the corresponding standardized direct and indirect associations between chatbot engagement and follow-up well-being, with indirect associations operating through in-person social interaction (Section2.4). Points indicate standardized coefficient estimates and horizontal bars indicate 95% confidence intervals. Estimates are based on theN=439N=439participants who completed the follow-up survey. Significance is denoted as∗p<0.05{}^{*}p<0.05,p∗⁣∗<0.01{}^{**}p<0.01, and∗∗∗p<0.001{}^{***}p<0.001.

2.4Social chatbot engagement was associated with lower well-being throughdisplacement of human social interaction

Finally, we tested the hypothesized social displacement pathway by extending the longitudinal models above to examine whether chatbot engagement was indirectly associated with subsequent well-being through lower in-person social interaction, while accounting for baseline well-being and prior chatbot engagement (Figure4and Table8).

Across all three dimensions, greater social engagement with chatbots at follow-up was associated with less time spent interacting with other people in person (interaction intensity:β=−0.17\beta=-0.17,p<0.001p<0.001; companionship use, free-text:β=−0.14\beta=-0.14,p=0.002p=0.002; self-disclosure:β=−0.15\beta=-0.15,p=0.002p=0.002). In turn, greater in-person social interaction was associated with higher follow-up well-being (β=0.14\beta=0.14–0.150.15, allp<0.001p<0.001). Accordingly, interaction intensity (indirect=−0.02-0.02,p<0.001p<0.001), companionship use (free-text: indirect=−0.02-0.02,p=0.004p=0.004), and self-disclosure (indirect=−0.02-0.02,p=0.002p=0.002) each showed significant negative indirect associations with follow-up well-being through in-person social interaction. Their direct associations with follow-up well-being were not statistically significant after accounting for in-person social interaction (interaction intensity:β=−0.06\beta=-0.06,p=0.074p=0.074; companionship use:β=−0.05\beta=-0.05,p=0.118p=0.118; self-disclosure:β=−0.05\beta=-0.05,p=0.152p=0.152). Overall, these results were consistent with the hypothesized social displacement pathway, with participants who engaged more intensively with AI companions, whether through interaction intensity, companionship use, or self-disclosure, reported less in-person social interaction and, in turn, lower follow-up well-being.

3Discussion

How AI companionship may shape well-being over time.

In this work, we provide longitudinal evidence to better understand the mechanisms through how social engagement with AI companions may shape psychological well-being over time. First, we found that social engagement developed through repeated interaction rather than simply reflecting users’ initial reasons for adopting a chatbot. Stronger earlier engagement predicted stronger subsequent engagement, which in turn was associated with lower well-being, which supports the proposed longitudinal engagement pathway. Importantly, these patterns were not limited to users who initially approached chatbots for companionship, suggesting that socially consequential forms of engagement can emerge through repeated interaction rather than being determined by users’ initial motivations. This is consistent with emerging evidence that social and emotional attachment to AI systems can develop over time, including among users who do not initially seek such relationships(Shi et al., 2026,Manoli et al., 2026,Wei and JD, 2026). These findings suggest that companionship-related risks may extend beyond systems explicitly designed as AI companions. General-purpose chatbots may also take on companion-like roles through repeated interactionHill (2025),Dimri (2026), suggesting that safety evaluations should consider how systems are actually used rather than their intended roles alone(Nelson et al., 2026). This may be even more important for general-purpose chatbots, where interactions with the same system can range from everyday conversation to sensitive contexts such as mental health support. Trust and expectations developed through general interactions may carry into these higher-stakes contexts. Moreover, seemingly responsive chatbot behaviors may not always be appropriate or beneficial, as they may include excessive affirmation, emotional appeals, or attempts to prolong engagementZhang et al. (2025a),De Freitas et al. (2025),Adewale and Muhammad (2025),Chu et al. (2025). Such risks may compound through repeated interaction, as seemingly reliable responses can build users’ trust in the chatbot while making them less aware of potential deterioration in model behaviorWei and JD (2026). Thus, forms of engagement that are beneficial within human relationships may have different implications when directed toward AI companions. Our social displacement findings provide further explanation that the influence of AI companionship should be considered beyond the interaction itself to users’ broader social lives, which is consistent with social displacement theoriesValkenburg (2022),Kraut et al. (1998). Besides, AI companions offer forms of interaction that can be difficult to sustain in human relationships, including near-constant availability and responsiveness tailored primarily to the userSun et al. (2026),Wang and Dehnert (2026). Over time, such experiences could shape expectations about relationshipsChu et al. (2025),Malfacini (2025), potentially making the reciprocity and negotiation of human relationships comparatively effortful or unrewarding, users may become more likely to substitute AI interactions for human connection, potentially reinforcing the displacement process over time.

From bowling alone to chatting alone.

The social displacement pathway may have implications beyond individual well-being because human relationships are reciprocal. When users redirect time and attention from family members, partners, or friends toward AI companions, others also lose opportunities for interaction, support, and shared experiences. Unlike other communication technologies that mediate interaction between people, AI companions can occupy users’ social attention without reciprocal human participation. This displacement may therefore have second-order effects, as a shift toward AI companionship may reduce not only users’ own human interaction, but also the social interaction available to those around them. If such displacement accumulates across relationships, it may have broader implications for social participation.Putnam (2000)argued that participation in social life sustains social capital through networks, reciprocity, and trust. From this perspective, widespread shifts from human interaction toward AI companionship could weaken the relationships and patterns of participation through which social capital is maintained. The consequences of AI companionship may therefore need to be considered not only at the level of individual users, but also at the level of the social networks and communities in which they participate.

Designing for healthier patterns of engagement.

Our findings suggest that AI companion safety cannot be evaluated solely within individual user-AI exchanges. Risks may emerge from patterns of interaction even when individual exchanges are not themselves harmful. This challenges safeguards centered primarily on detecting harmful content or acute risks within individual conversationsRen et al. (2026),Juneja and Lomidze (2026),Zhang et al. (2026). The responsible design of AI companions may require a more proactive approach, which further asks what patterns of engagement a system encourages over time and whether those patterns support sustained well-being. This perspective also challenges engagement as a primary objective for AI companion design. Metrics such as time spent or frequency of use may indicate a successful consumer product while overlooking whether sustained engagement displaces beneficial human interaction. Similar to prior studies about social media, designing for healthier AI companions may therefore require evaluating not only how much users engage, but also from broader societal valuesBernstein et al. (2023),Munn (2020). One alternative is to design AI companions to facilitate human interaction rather than maximize interaction with the AI itself. This could involve moving beyond the user-AI dyad by allowing AI companions to participate alongside romantic partners, family members, or friends to facilitate conversation, shared activities, or mutual support. More broadly, such designs could reposition AI companions from competitors for social attention to systems that support relationships between people.

3.1Limitations

We acknowledge several limitations of this study.

Methodological limitations.

Although our longitudinal design enables stronger temporal inference than cross-sectional studies, it includes only two waves of data collection. Consequently, we cannot fully characterize the dynamic and potentially reciprocal relationships between AI companionship and psychological well-being over longer periods. To examine the robustness of our findings to unobserved changes in chatbot use between the two measurement waves, we conducted simulation-based sensitivity analyses (SectionA.1). While these analyses suggest that our findings are robust across a range of plausible scenarios, they cannot substitute for more densely sampled longitudinal data. Future research using three or more waves, ecological momentary assessment, or continuous behavioral logging would provide a richer understanding of how AI companionship and well-being co-evolve over time.

Another limitation concerns the operationalization of companionship use. As discussed in Section4.4, AI companion use is often multifaceted rather than mutually exclusive. AtT1T_{1}, participants identified their primary motivation for using chatbots using a forced-choice question, requiring them to select a single usage category. Building on our previous finding that users often engage with AI companions for multiple concurrent purposesZhang et al. (2025b), we refined this measure atT2T_{2}by asking participants to rate the extent to which each usage category characterized their chatbot use. Although this approach better captures the multidimensional nature of AI companion use, the change in measurement across waves limits direct comparability of this construct over time. In addition, many of our measures relied on self-report and are therefore subject to recall bias, reporting bias, and differences in how participants interpreted survey questions. Future studies would benefit from using consistent measurement approaches across waves while integrating multidimensional self-report measures with behavioral data where appropriate to more comprehensively capture the development of AI companionship over time.

Finally, our analyses may have been influenced by platform-level changes to Character.AI during the study period. Because we did not have access to the platform’s internal model updates, moderation policies, or system design changes, we could not determine whether such modifications occurred or assess how they may have affected chatbot behavior and user experiences.

Participant and sampling limitations.

Of the 1,182 participants who completedT1T_{1}, 439 returned atT2T_{2}, corresponding to a retention rate of 37.1% (dropout rate = 62.9%). A substantial contributor to this attrition was the recruitment platform: 59.54% of participants who completedT1T_{1}were no longer active on Prolific by the time of the follow-up study and therefore could not be re-contacted. To assess and mitigate potential attrition bias, we first examined predictors ofT2T_{2}participation and then estimated all longitudinal models using Heckman selection correction (See more details in Section4.2). While these analyses help account for observable and selection-related differences between retained and non-retained participants, they cannot eliminate bias arising from unobserved factors or fully recover information from participants lost to follow-up. Future longitudinal studies would benefit from recruitment and retention strategies that facilitate more complete follow-up across multiple waves. Although the platform’s wide variety of character-based chatbots supports diverse forms of interaction and relational engagement, its design may also foster patterns of use that differ from those on other AI companion platforms. Consequently, our findings may not fully generalize to users of other conversational AI systems. Furthermore, our sample consisted of English-speaking participants residing in the United States and recruited through Prolific. Although Prolific provides relatively diverse participant pools, the findings may not generalize across cultural, linguistic, socioeconomic, or broader global populations.

The longitudinal process identified here is unlikely to unfold uniformly across users. An important unanswered question is what determines whether sustained AI companionship develops into a complementary source of support or is accompanied by reduced engagement with human relationships. Future research should therefore identify the individual and contextual factors that shape these different trajectories, including loneliness, offline social support, attachment style, prior mental health, and motivations for AI companion use. Understanding this heterogeneity will be critical for explaining when AI companionship promotes well-being, when it may carry unintended social costs, and why these outcomes differ across users.

4Methods

To examine how AI companionship shapes well-being over time, we conducted a two-wave longitudinal study of active users of Character.AI. At baseline (T1T_{1}), 1,182 participants completed the survey. Of these participants, 439 completed the follow-up survey (T2T_{2}: retention rate = 37.1%), with a mean interval of 362.5 days (SD = 88.2 days) betweenT1T_{1}andT2T_{2}. Among the returning participants, 108 reported that they no longer used chatbots atT2T_{2}. For our primary analyses, we therefore coded theirT2T_{2}interaction intensity, companionship use, and self-disclosure as zero, reflecting the absence of chatbot engagement at follow-up. As a robustness check, we repeated all analyses among the 331 participants who continued to use chatbots atT2T_{2}. The results were consistent with the primary findings (see Supplementary MaterialsA.3for full results). To account for potential bias from non-random attrition atT2T_{2}, all longitudinal models incorporated a Heckman selection correction (Detailed model are described in Section4.2).

This study was approved by the Stanford University Institutional Review Board (Protocol ID: 76072). All data were stored on secure institutional servers, with access restricted to the authors of this study. All participants provided informed consent prior to participation in each wave. In accordance with Prolific’s policies and ethical guidelines, we ensured participant anonymity and did not collect any personally identifiable information.

4.1Participants and data collection

Baseline recruitment.

Participants were recruited via Prolific using prescreening criteria targeting U.S.-based, native English-speaking users of Character.AI. Eligibility required having used the platform for over one month and having interacted with at least three distinct chatbots before enrolling atT1T_{1}. The baseline (T1T_{1}) survey was administered using Qualtrics and took approximately 10 minutes to complete. Participants received $4 for survey completion. The final baseline sample consisted of 1,182 valid responses.

T2T_{2}follow-up.

AllT1T_{1}participants who consented to be re-contacted were invited to complete a follow-up survey.T1T_{1}data collection took place between September 2024 and August 2025, andT2T_{2}data collection took place between August 2025 and June 2026. The mean interval betweenT1T_{1}andT2T_{2}was 362.5 days (SD = 88.2 days). TheT2T_{2}survey retained the core measures fromT1T_{1}, including well-being, chatbot interaction intensity, companionship use, and self-disclosure, and additionally measured human social interaction as the amount of time participants spent interacting with other people in person. A total of 439 participants completedT2T_{2}, yielding a retention rate of 37.1%. Among these participants, 108 reported that they no longer used chatbots atT2T_{2}. TheirT2T_{2}interaction intensity, companionship use, and self-disclosure were coded as zero, reflecting no chatbot engagement at follow-up. All 439 returning participants were therefore included in the primary longitudinal analyses. As a robustness check, we repeated the analyses among the 331 participants who continued to use chatbots atT2T_{2}and the findings were consistent with the primary analyses (Supplementary MaterialsA.3). Participants received $10 for completing the follow-up survey.

Participant characteristics.

At baseline (N=1,182N=1,182), participants ranged in age from 18 to 90 years (M=30.2M=30.2,S​D=10.7SD=10.7; median = 27). Overall, 48.1% identified as male, 46.6% as female, and 5.2% as non-binary; 40.6% reported being single and 59.4% reported being in a romantic relationship. At the time of enrollment, 15.7% had used the platform for one to three months, 34.9% for three months to one year, and 49.4% for more than one year.

Among the 439 participants who completedT2T_{2}, age ranged from 18 to 68 years (M=28.7M=28.7,S​D=8.7SD=8.7; median = 26). Overall, 41.2% identified as male, 50.1% as female, and 8.7% as non-binary; 51.9% were single and 48.1% were in a romantic relationship.

4.2Selection and attrition analysis

To account for potential non-random attrition betweenT1T_{1}andT2T_{2}, we used a Heckman selection framework. The selection equation modeled the probability of completingT2T_{2}using probit regressionPr⁡(completedT2=1)\Pr(\text{completed T\_\{2\}}=1), withT1T_{1}well-being, companionship use, self-disclosure, interaction intensity, anthropomorphism, age, gender, relationship status, chatbot tenure, and participants’ total prior approvals on Prolific as predictors. For models examining how social engagement evolved fromT1T_{1}andT2T_{2}(Section2.2), we applied the Heckman correction using joint maximum-likelihood estimation on the full baseline sample. For the longitudinal structural equation models examining well-being (Section2.3and2.4), we used the fitted probit model to compute an inverse Mills ratio (λi\lambda_{i}) for each participant and included this term as a covariate in each model equation, following the Heckman two-step correction for sample selection. The corresponding logistic regression is reported in Table2.

PredictorEstimate95% CIp-valueIntercept−0.69-0.69***[−0.93[-0.93,−0.46]-0.46]p<0.001p<0.001Tenure0.05\phantom{-}0.05[−0.09[-0.09,0.19]\phantom{-}0.19]p=0.457p=0.457Male−0.34-0.34*[−0.61[-0.61,−0.06]-0.06]p=0.016p=0.016Non-binary0.44\phantom{-}0.44[−0.16[-0.16,1.04]\phantom{-}1.04]p=0.151p=0.151Age−0.14-0.14[−0.30[-0.30,0.01]\phantom{-}0.01]p=0.071p=0.071Single0.40\phantom{-}0.40**[0.12[\phantom{-}0.12,0.68]\phantom{-}0.68]p=0.005p=0.005Total approvals0.70\phantom{-}0.70***[0.56[\phantom{-}0.56,0.85]\phantom{-}0.85]p<0.001p<0.001Companionship useT1T_{1}0.11\phantom{-}0.11[−0.32[-0.32,−0.54]-0.54]p=0.609p=0.609Self-disclosureT1T_{1}0.16\phantom{-}0.16[−0.01[-0.01,0.32]\phantom{-}0.32]p=0.058p=0.058IntensityT1T_{1}−0.46-0.46***[−0.64[-0.64,−0.28]-0.28]p<0.001p<0.001AnthropomorphismT1T_{1}−0.04-0.04[−0.20[-0.20,0.12]\phantom{-}0.12]p=0.619p=0.619Well-beingT1T_{1}−0.32-0.32***[−0.46[-0.46,−0.18]-0.18]p<0.001p<0.001

Table 2:Predictors ofT2T_{2}survey completion.Probit regression predictingT2T_{2}completion (1 = returned, 0 = did not complete) fromT1T_{1}demographics, chatbot engagement, well-being, and total prior task approvals on Prolific, among allN=1,182N=1,182participants who complete survey atT1T_{1}. All predictors were standardized. Estimates are probit regression coefficients with 95% confidence intervals.∗p<0.05{}^{*}p<0.05;p∗⁣∗<0.01{}^{**}p<0.01;∗∗∗p<0.001{}^{***}p<0.001.

4.3Longitudinal analysis overview

We conducted three sets of longitudinal analyses corresponding to our research questions. All models adjusted for the same set of covariates (𝐂\mathbf{C}): baseline chatbot tenure, age, gender, and relationship status.

First, to examine how social engagement evolved fromT1T_{1}toT2T_{2}, we modeled each social engagement dimension atT2T_{2}as a function of all three social engagement dimensions atT1T_{1}: interaction intensity, companionship use, and self-disclosure.

XT2∼IntensityT1+CompanionshipuseT1+Self​-​disclosureT1+𝐂,\displaystyle X_{T_{2}}\sim\mathrm{Intensity}_{T_{1}}+\mathrm{Companionshipuse}_{T_{1}}+\mathrm{Self\text{-}disclosure}_{T_{1}}+\mathbf{C},(1)whereXXrepresents each of the three social engagement dimensions.

Second, to examine the longitudinal relationship between social engagement and psychological well-being, we fitted separate longitudinal path models for each social engagement dimension, in whichT1T_{1}engagement predictedT2T_{2}engagement, andT2T_{2}engagement was associated withT2T_{2}well-being while accounting for both engagement and well-being atT1T_{1}:

XT2\displaystyle X_{T_{2}}∼a1​XT1+𝐂+λi,\displaystyle\sim a_{1}X_{T_{1}}+\mathbf{C}+\lambda_{i},(2)Well-beingT2\displaystyle\text{Well-being}_{T_{2}}∼ar​Well-beingT1+b1​XT2+c​XT1+𝐂+λi.\displaystyle\sim a_{r}\text{Well-being}_{T_{1}}+b_{1}X_{T_{2}}+cX_{T_{1}}+\mathbf{C}+\lambda_{i}.(3)Here,XXrepresents the social engagement dimension examined in each model, andλi\lambda_{i}denotes the inverse Mills ratio defined in Section4.2. The coefficienta1a_{1}represents the association betweenT1T_{1}andT2T_{2}engagement,ara_{r}represents the association betweenT1T_{1}andT2T_{2}well-being,b1b_{1}represents the association betweenT2T_{2}engagement andT2T_{2}well-being while adjusting forT1T_{1}engagement and well-being, andccrepresents the direct association betweenT1T_{1}engagement andT2T_{2}well-being. We estimated the indirect association throughT2T_{2}engagement asa1​b1a_{1}b_{1}and the total association asc+a1​b1c+a_{1}b_{1}. Models were fitted separately for interaction intensity, companionship use, and self-disclosure.

Third, to examine whether human social interaction may explain the association between sustained social engagement and psychological well-being, we extended the longitudinal path model to include time spent interacting with other people in person atT2T_{2}as a potential pathway:

XT2\displaystyle X_{T_{2}}∼a1​XT1+𝐂+λi,\displaystyle\sim a_{1}X_{T_{1}}+\mathbf{C}+\lambda_{i},(4)HT2\displaystyle H_{T_{2}}∼d​XT2+𝐂+λi,\displaystyle\sim dX_{T_{2}}+\mathbf{C}+\lambda_{i},(5)Well-beingT2\displaystyle\text{Well-being}_{T_{2}}∼ar​Well-beingT1+cp​XT2+b​HT2+c​XT1+𝐂+λi.\displaystyle\sim a_{r}\text{Well-being}_{T_{1}}+c_{p}X_{T_{2}}+bH_{T_{2}}+cX_{T_{1}}+\mathbf{C}+\lambda_{i}.(6)Here,HHrepresents time spent interacting with other people in person. The additional coefficientsddandbbrepresent the associations betweenT2T_{2}engagement and human social interaction, and between human social interaction andT2T_{2}well-being, respectively The indirect association through human social interaction was estimated asd​bdb, the direct association betweenT2T_{2}engagement andT2T_{2}well-being ascpc_{p}, and the total association ascp+d​bc_{p}+db. We fitted the model separately for each dimension of social engagement.

4.4Measurements

4.4.1Identifying companion interaction

we assessed companionship use using two complementary measures: participants’ reported purposes for using chatbots and their open-ended descriptions of their relationships with character chatbots.

First, we measured participants’ purposes for using Character.AI across four categories: companionship, entertainment, productivity, and curiosity. AtT1T_{1}, participants were asked,“What is your primary use for character chatbots? (Select the most appropriate one)”, and selected one of the four categories. Our initial cross-sectional analyses showed that participants often used chatbots for multiple purposes rather than a single primary purpose. We therefore revised the measure atT2T_{2}to capture multiple purposes simultaneously. Participants were asked,“Please rate how much you use Character.AI for each of the following purposes”, and rated each category separately on a Likert scale. The companionship category described using chatbots as a personal, human means of interaction with social value or to strengthen social interactions with other people.

Second, at bothT1T_{1}andT2T_{2}, participants described their relationships with chatbots in response to the same open-ended question:“Can you describe your relationship with character chatbots in a few keywords?”We classified these responses usingGPT-5-miniinto the same four categories used in the purpose measure (companionship, entertainment, productivity, and curiosity), allowing multiple categories to be assigned when applicable. Responses classified as companionship were used as a complementary measure of companionship use. To validate the classification, two trained annotators independently coded 30 randomly sampled responses using the same four categories. The model classifications showed 95.8% average label-level agreement with human annotations and exact agreement across all four categories for 85% of responses (Fleiss’κ=0.89\kappa=0.89). The classification prompt is provided in Supplementary MaterialsA.4

Wave 1Wave 2Well-being Scale itemMeanS.D.MeanS.D.Life Satisfaction:I am satisfied with my life.4.991.664.411.76Positive Affect:I feel good most of the time.4.931.664.431.74Negative Affect:I feel bad most of the time.†4.571.784.541.80Loneliness:I feel lonely.†4.191.873.941.92Belonging:I feel a sense of belonging in my community.4.501.713.821.75Social Support:There are people who give me support and encouragement.5.541.305.281.45Cronbach’sα\alpha0.880.89

Table 3:The Well-being Scale (six items adapted from the Comprehensive Inventory of Thriving;Su et al. (2014)) was measured at both waves. The table reports item-level Mean and S.D. atT1T_{1}(N=1,182N=1,182) andT2T_{2}(N=439N=439), together with internal consistency (Cronbach’sα\alpha) at each wave. All items were rated on a 7-point Likert scale ranging from 1 (strongly disagree) to 7 (strongly agree). Items marked with†were reverse-coded so that higher values reflect greater well-being, consistent with the composite scoring.

4.4.2Measuring subjective well-being

At each wave, we measured subjective well-being using six items from the Comprehensive Inventory of Thriving (CIT)Su et al. (2014), covering life satisfaction, positive and negative feelings, loneliness, social support, and sense of belonging. Following prior workErnala et al. (2022), we selected the highest-loading item for each construct to reduce participant burden. Items assessing negative feelings and loneliness were reverse-coded so that higher scores reflected greater well-being. A composite score was computed as the mean of the six items, which showed good internal consistency (Cronbach’sα=0.88\alpha=0.88atT1T_{1}and0.890.89atT2T_{2}). Item-level descriptive statistics for both waves are reported in Table3.

Wave 1Wave 2Chatbot Intensity Scale itemMeanS.D.MeanS.D.Chatbot Network Size:How many chatbots do you feel at ease with that you can talk about private matters?1.871.431.461.581 = None, 2 = One, 3 = Two, 4 = Three or four, 5 = Five to eight, 6 = Nine or moreDaily Chatbot Usage Time:In the past week, on average, approximately how much time PER DAY have you spent chatting with chatbots using Character.AI?2.691.512.241.391 = Less than 10 minutes per day, 2 = 10–30 minutes per day, 3 = 31–60 minutes per day, 4 = 1–2 hours per day, 5 = 2–3 hours per day, 6 = More than 3 hours per dayEveryday Chatbot Activity:Interacting with character chatbots is part of my everyday activity.‡4.501.804.121.88Pride in Chatbot Use:I am proud to tell people I use character chatbots.‡4.032.012.841.76Chatbots in Daily Routine:Character chatbots have become part of my daily routine.‡4.451.813.931.93Out of Touch Without Chatbots:I feel out of touch when I haven’t interacted with character chatbots for a while.‡3.541.912.721.71Regret if Chatbots Gone:I would be sorry if character chatbots were no longer available.‡4.971.754.461.92Cronbach’sα\alpha0.860.82

Table 4:The Chatbot Intensity Scale (seven items) measured at both waves. The table reports item-level Mean and S.D. atT1T_{1}andT2T_{2}, together with the scale’s internal consistency (Cronbach’sα\alpha) at each wave. Items were standardized prior to averaging into the composite intensity score, due to differing response formats. Items marked with‡were rated on a 7-point Likert scale ranging from 1 (strongly disagree) to 7 (strongly agree).

4.4.3Measuring chatbot interaction intensity

Chatbot interaction intensity was measured using the Chatbot Intensity Scale(Zhang et al., 2025b), which captures both behavioral engagement (chatbot network size and typical daily time spent interacting with chatbots) and the extent to which participants felt emotionally connected to chatbots and integrated chatbot interactions into their daily routines. Because the items use differing response formats, they were standardized and averaged into a composite intensity score. The scale showed good internal consistency at both waves (Cronbach’sα=0.86\alpha=0.86atT1T_{1}and0.820.82atT2T_{2}). Item-level descriptive statistics for both waves are reported in Table4.

Wave 1Wave 2Self-disclosure Scale itemMeanS.D.MeanS.D.I feel like I can sometimes be more personal when interacting with character chatbots.5.051.594.931.74It is easier to disclose personal information to a character chatbot.4.911.714.801.88I feel like I can be more open when communicating with character chatbots.5.241.545.031.78Cronbach’sα\alpha0.890.93

Table 5:The three-item self-disclosure scale, adapted from the MOCA instrument(Ledbetter, 2009), measured at both waves. Items were rated on a 7-point Likert scale ranging from 1 (strongly disagree) to 7 (strongly agree). The table reports item-level Mean and S.D. atT1T_{1}(N=1,182N=1,182) andT2T_{2}(N=331N=331), together with internal consistency (Cronbach’sα\alpha) at each wave.

4.4.4Measuring self-disclosure levels

We assessed participants’ self-reported willingness to disclose to chatbots using an adapted version of the self-disclosure subscale from the MOCA instrumentLedbetter (2009). Due to the length of the survey, we selected the first three items in the survey, with the seven Likert-scale items capturing comfort sharing personal thoughts, feelings, and experiences with chatbots (Cronbach’sα=0.89\alpha=0.89atT1T_{1}and0.930.93atT2T_{2}). Item-level descriptive statistics for this three-item scale at both waves are reported in Table5.

4.4.5Measuring human social interaction

Human social interaction was measured using a single item asking participants,“In the past week, on average, approximately how much time per day have you spent communicating or interacting with other people in person (e.g., talking face-to-face with friends, family, classmates, coworkers, etc.)?”Participants responded using the same six-point scale as the chatbot interaction-time item described above (Section4.4.3): 1 = less than 10 minutes per day, 2 = 10-30 minutes, 3 = 31-60 minutes, 4 = 1-2 hours, 5 = 2-3 hours, and 6 = more than 3 hours per day.

4.4.6Measuring anthropomorphism

We assessed perceived anthropomorphism of chatbots with a single item:“To what extent do you perceive the character chatbots as more human or more machine-like in your interactions?”Participants responded on a seven-point Likert scale ranging from 1 (Completely machine-like) to 7 (Completely human-like), with higher scores indicating greater perceived humanlikeness.

Acknowledgment

This work was supported by grants from the UK AI Security Institute (AISI), OpenAI AI and Mental Health Grant, and NSF CAREER IIS-2247357. D.Z. is supported in part by the Paul and Daisy Soros Fellowship for New Americans. We thank Lujain Ibrahim, Vishakh Padmakumar, Meryl Ye, Steve Rathje, and all members of Stanford SALT lab for their helpful suggestions and feedback at different stages of this project.

References

• Adewale and Muhammad (2025)Muyideen Dele Adewale and Umaina Ibrahim Muhammad.From virtual companions to forbidden attractions: The seductive rise of artificial intelligence love, loneliness, and intimacy—a systematic review.Journal of Technology in Behavioral Science, pages 1–18, 2025.
• Alabed et al. (2024)Amani Alabed, Ana Javornik, Diana Gregory-Smith, and Rebecca Casey.More than just a chat: A taxonomy of consumers’ relationships with conversational ai agents and their well-being implications.European Journal of Marketing, 58(2):373–409, 2024.
• Berkman and Syme (1979)Lisa F Berkman and S Leonard Syme.Social networks, host resistance, and mortality: a nine-year follow-up study of alameda county residents.American journal of Epidemiology, 109(2):186–204, 1979.
• Bernstein et al. (2023)Michael Bernstein, Angèle Christin, Jeffrey Hancock, Tatsunori Hashimoto, Chenyan Jia, Michelle Lam, Nicole Meister, Nathaniel Persily, Tiziano Piccardi, Martin Saveski, et al.Embedding societal values into social media algorithms.Journal of Online Trust and Safety, 2(1), 2023.
• Brandtzaeg et al. (2022)Petter Bae Brandtzaeg, Marita Skjuve, and Asbjørn Følstad.My ai friend: How users of a social chatbot understand their human–ai friendship.Human Communication Research, 48(3):404–429, 2022.
• Chu et al. (2025)Minh Duc Chu, Patrick Gerard, Kshitij Pawar, Charles Bickham, and Kristina Lerman.Illusions of intimacy: Emotional attachment and emerging psychological risks in human-ai relationships.arXiv preprint arXiv:2505.11649, 1(1):19, 2025.
• Croes and Antheunis (2021)Emmelyn AJ Croes and Marjolijn L Antheunis.Can we be friends with mitsuku? a longitudinal study on the process of relationship formation between humans and a social chatbot.Journal of social and personal relationships, 38(1):279–300, 2021.
• Croes et al. (2023)Emmelyn AJ Croes, Marjolijn L Antheunis, Martijn B Goudbeek, and Nathan W Wildman.“i am in your computer while we talk to each other” a content analysis on the use of language-based strategies by humans and a social chatbot in initial human-chatbot interactions.International Journal of Human–Computer Interaction, 39(10):2155–2173, 2023.
• Dam et al. (2024)Sumit Kumar Dam, Choong Seon Hong, Yu Qiao, and Chaoning Zhang.A complete survey on llm-based ai chatbots.arXiv preprint arXiv:2406.16937, 2024.
• De Freitas et al. (2025)Julian De Freitas, Zeliha Oguz-Uguralp, and Ahmet Kaan-Uguralp.Emotional manipulation by ai companions.arXiv preprint arXiv:2508.19258, 2025.
• De Freitas et al. (2026)Julian De Freitas, Zeliha Oğuz-Uğuralp, Ahmet Kaan Uğuralp, and Stefano Puntoni.Ai companions reduce loneliness.Journal of Consumer Research, 52(6):1126–1148, 2026.
• Dimri (2026)Amogh Dimri.The people who marry chatbots, 2026.URLhttps://www.theatlantic.com/ideas/2026/01/chatbot-marriage-ai-relationships-romance/685459/.Accessed: 2026-09-06.
• Ernala et al. (2022)Sindhu Kiranmai Ernala, Moira Burke, Alex Leavitt, and Nicole B Ellison.Mindsets matter: How beliefs about facebook moderate the association between time spent and well-being.InProceedings of the 2022 CHI Conference on Human Factors in Computing Systems, pages 1–13, 2022.
• Fang et al. (2025)Cathy Mengying Fang, Auren R Liu, Valdemar Danry, Eunhae Lee, Samantha WT Chan, Pat Pataranutaporn, Pattie Maes, Jason Phang, Michael Lampe, Lama Ahmad, et al.How ai and human behaviors shape psychosocial effects of chatbot use: A longitudinal randomized controlled study.arXiv preprint arXiv:2503.17473, 2025.
• Folk and Dunn (2026)Dunigan Folk and Elizabeth Dunn.How does turning to ai for companionship predict loneliness and vice versa?Psychological Science, 37(4):276–286, 2026.
• Guingrich and Graziano (2025)Rose E Guingrich and Michael SA Graziano.A longitudinal randomized control study of companion chatbot use: Anthropomorphism and its mediating role on social impacts.arXiv preprint arXiv:2509.19515, 2025.
• Hall (2020)Jeffrey A Hall.Relating through technology.Cambridge University Press, 2020.
• Hall and Liu (2022)Jeffrey A Hall and Dong Liu.Social media use, social displacement, and well-being.Current Opinion in Psychology, 46:101339, 2022.
• Herbener and Damholdt (2025)Arthur Bran Herbener and Malene Flensborg Damholdt.Are lonely youngsters turning to chatbots for companionship? the relationship between chatbot usage and social connectedness in danish high-school students.International Journal of Human-Computer Studies, 196:103409, 2025.
• Hill (2025)Kashmir Hill.She is in love with chatgpt, 2025.URLhttps://www.nytimes.com/2025/01/15/technology/ai-chatgpt-boyfriend-companion.html.Accessed: 2025-03-12.
• House (1983)James S House.Work stress and social support.Addison-Wesley series on occupational stress, 1983.
• Hwang et al. (2025)Angel Hsing-Chi Hwang, Fiona Li, Jacy Reese Anthis, and Hayoun Noh.How ai companionship develops: Evidence from a longitudinal study.arXiv preprint arXiv:2510.10079, 2025.
• Jacobs (2024)Kerrin Artemis Jacobs.Digital loneliness—changes of social recognition through ai companions.Frontiers in Digital Health, 6:1281037, 2024.
• Juneja and Lomidze (2026)Prerna Juneja and Lika Lomidze.Persona-grounded safety evaluation of ai companions in multi-turn conversations.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 18148–18175, 2026.
• Kraut et al. (1998)Robert Kraut, Michael Patterson, Vicki Lundmark, Sara Kiesler, Tridas Mukophadhyay, and William Scherlis.Internet paradox: A social technology that reduces social involvement and psychological well-being?American psychologist, 53(9):1017, 1998.
• Ledbetter (2009)Andrew M Ledbetter.Measuring online communication attitude: Instrument development and validation.Communication Monographs, 76(4):463–486, 2009.
• Lee et al. (2024)Yoon Kyung Lee, Jina Suh, Hongli Zhan, Junyi Jessy Li, and Desmond C Ong.Large language models produce responses perceived to be empathic.In2024 12th International Conference on Affective Computing and Intelligent Interaction (ACII), pages 63–71. IEEE, 2024.
• Li et al. (2026)Ruo-Ning Li, Dunigan Folk, Abhay Singh, Lyle Ungar, and Elizabeth Dunn.Is a random human peer better than a highly supportive chatbot in reducing loneliness over time?Journal of Experimental Social Psychology, 125:104911, 2026.
• Little (2024)Todd D Little.Longitudinal structural equation modeling.Guilford Publications, 2024.
• Ma et al. (2024)Zilin Ma, Yiyang Mei, and Zhaoyuan Su.Understanding the benefits and challenges of using large language model-based conversational agents for mental well-being support.InAMIA Annual Symposium Proceedings, volume 2023, page 1105, 2024.
• Malfacini (2025)Kim Malfacini.The impacts of companion ai on human relationships: risks, benefits, and design considerations.Ai & Society, 40(7):5527–5540, 2025.
• Manoli et al. (2026)Aikaterina Manoli, Janet VT Pauketat, Ali Ladak, Hayoun Noh, Angel Hsing-Chi Hwang, and Jacy Reese Anthis.Digital companionship: Overlapping uses of ai companions and ai assistants.InProceedings of the 2026 CHI Conference on Human Factors in Computing Systems, pages 1–25, 2026.
• Maples et al. (2024)Bethanie Maples, Merve Cerit, Aditya Vishwanath, and Roy Pea.Loneliness and suicide mitigation for students using gpt3-enabled chatbots.npj mental health research, 3(1):4, 2024.
• Merrill Jr et al. (2025)Kelly Merrill Jr, Sai Datta Mikkilineni, and Marco Dehnert.Artificial intelligence chatbots as a source of virtual social support: Implications for loneliness and anxiety management.Annals of the New York Academy of Sciences, 1549(1):148–159, 2025.
• Munn (2020)Luke Munn.Angry by design: toxic communication and technical architectures.Humanities and Social Sciences Communications, 7(1):53, 2020.
• Nelson et al. (2026)Benjamin W Nelson, Mark Kalinich, and John Torous.Specialized or general-purpose—the wrong question for mental health ai safety.JAMA, 2026.
• Putnam (2000)Robert D. Putnam.Bowling alone: The collapse and revival of American community.Simon & Schuster, New York, 2000.
• Ren et al. (2026)Yanjing Ren, Reza Ebrahimi, and TengTeng Ma.Aicompanionbench: Benchmarking llms-as-judges for ai companion safety.arXiv preprint arXiv:2606.04867, 2026.
• Robb and Mann (2025)Michael B. Robb and Supreet Mann.Talk, trust, and trade-offs: How and why teens use ai companions.Technical report, Common Sense Media, San Francisco, CA, 2025.URLhttps://www.commonsensemedia.org/research/talk-trust-and-trade-offs-how-and-why-teens-use-ai-companions.
• Shi et al. (2026)Yaoxi Shi, Cathy Mengying Fang, Pattie Maez, and Amit Goldenberg.Stumbling into ai emotional dependence: How routine ai interactions reshape human connection.arXiv preprint arXiv:2606.04150, 2026.
• Skjuve et al. (2021)Marita Skjuve, Asbjørn Følstad, Knut Inge Fostervold, and Petter Bae Brandtzaeg.My chatbot companion-a study of human-chatbot relationships.International Journal of Human-Computer Studies, 149:102601, 2021.
• Skjuve et al. (2022)Marita Skjuve, Asbjørn Følstad, Knut Inge Fostervold, and Petter Bae Brandtzaeg.A longitudinal study of human–chatbot relationships.International Journal of Human-Computer Studies, 168:102903, 2022.
• Smith et al. (2025)Molly G Smith, Thomas N Bradbury, and Benjamin R Karney.Can generative ai chatbots emulate human connection? a relationship science perspective.Perspectives on Psychological Science, 20(6):1081–1099, 2025.
• Su et al. (2014)Rong Su, Louis Tay, and Ed Diener.The development and validation of the comprehensive inventory of thriving (cit) and the brief inventory of thriving (bit).Applied Psychology: Health and Well-Being, 6(3):251–279, 2014.
• Sullivan et al. (2023)Yulia Sullivan, Serge Nyawa, and Samuel Fosso Wamba.Combating loneliness with artificial intelligence: An ai-based emotional support model.2023.
• Sun et al. (2026)Xiaoran Sun, Yunqi Wang, and Brandon T McDaniel.Ai companions and adolescent social relationships: Benefits, risks, and bidirectional influences.Child Development Perspectives, page aadaf009, 2026.
• Ta et al. (2020)Vivian Ta, Caroline Griffith, Carolynn Boatfield, Xinyu Wang, Maria Civitello, Haley Bader, Esther DeCero, and Alexia Loggarakis.User experiences of social support from companion chatbots in everyday contexts: thematic analysis.Journal of medical Internet research, 22(3):e16235, 2020.
• Umberson and Karas Montez (2010)Debra Umberson and Jennifer Karas Montez.Social relationships and health: A flashpoint for health policy.Journal of health and social behavior, 51(1_suppl):S54–S66, 2010.
• Valkenburg (2022)Patti M Valkenburg.Social media use and well-being: What we know and what we need to know.Current opinion in psychology, 45:101294, 2022.
• Wang and Dehnert (2026)Skyler Wang and Marco Dehnert.On-demand intimacy: The sociotechnical appeal of ai companions.Social Media+ Society, 12(1):20563051251410394, 2026.
• Wei and JD (2026)MD Wei and Marlynn H JD.Cascades of drift: Mental health risks of prolonged ai conversations.Available at SSRN 6433263, 2026.
• Yuan et al. (2026)Yunhao Yuan, Jiaxun Zhang, Talayeh Aledavood, Renwen Zhang, and Koustuv Saha.Mental health impacts of ai companions: Triangulating social media quasi-experiments, user perspectives, and relational lens.InProceedings of the 2026 CHI Conference on Human Factors in Computing Systems, pages 1–22, 2026.
• Zhang et al. (2025a)Renwen Zhang, Han Li, Han Meng, Jinyuan Zhan, Hongyuan Gan, and Yi-Chieh Lee.The dark side of ai companionship: A taxonomy of harmful algorithmic behaviors in human-ai relationships.InProceedings of the 2025 CHI conference on human factors in computing systems, pages 1–17, 2025a.
• Zhang et al. (2026)Renwen Zhang, Han Meng, Jian Chai, Yuntao Lin, and Yi-Chieh Lee.Companionharm: A multi-turn benchmark for detecting harms in real-world ai companion conversations.arXiv preprint arXiv:2608.25377, 2026.
• Zhang et al. (2025b)Yutong Zhang, Dora Zhao, Jeffrey T Hancock, Robert Kraut, and Diyi Yang.The rise of ai companions: how human-chatbot relationships influence well-being.arXiv preprint arXiv:2506.12605, 2025b.

Appendix ASupplementary information

A.1Robustness check: sensitivity analysis

In the longitudinal analyses presented in the Section2, chatbot usage and well-being are both measured at follow-up (T2T_{2}). To examine whether the main findings were robust to this measurement, we constructed simulated chatbot usage measures representing usage between (T1T_{1}) and (T2T_{2}), and re-estimated the longitudinal models using these simulated measures in place of the observed follow-up usage. We then examined whether the estimated association between chatbot usage and follow-up well-being remained consistent after replacing the observed follow-up usage with the simulated usage measure.

Specifically, letTiT_{i}denote an unobserved time point between the two survey waves, whereT1<Ti<T2T_{1}<T_{i}<T_{2}, consider the chatbot usage evolves continuously across time, then we can approximate the unobserved usage atTiT_{i}as a weighted combination of the observed usage at baseline (T1T_{1}) and follow-up (T2T_{2}):

X⁡(Ti)=w​X​(T1)+(1−w)​X​(T2)+ϵ,ϵ∼𝒩⁡(0,σ),X(T_{i})=wX(T_{1})+(1-w)X(T_{2})+\epsilon,\qquad\epsilon\sim\mathcal{N}(0,\sigma),(7) whereX⁡(T1)X(T_{1})andX⁡(T2)X(T_{2})denote the standardized chatbot usage measured at baseline and follow-up, respectively. The mixing parameterw∈[0,1]w\in[0,1]determines the assumed timing of chatbot usage during the interval, withw=0w=0corresponding to the observed follow-up measure andw=1w=1corresponding to the observed baseline measure. The Gaussian noise termϵ\epsilonintroduces additional measurement uncertainty, withσ\sigmacontrolling its magnitude.

For each usage construct (interaction intensity, companionship use, and self-disclosure), we replaced the observed follow-up usage measure with the simulated measureX⁡(Ti)X(T_{i})and re-estimated all six longitudinal models reported in the Section2:

Well-being​(T2)\displaystyle\text{Well-being}(T_{2})∼Well-being​(T1)+X⁡(Ti)+Controls,\displaystyle\sim\text{Well-being}(T_{1})+X(T_{i})+\text{Controls},(8)Well-being​(T2)\displaystyle\text{Well-being}(T_{2})∼Well-being​(T1)+X⁡(Ti)+Human social support​(T2)+Controls,\displaystyle\sim\text{Well-being}(T_{1})+X(T_{i})+\text{Human social support}(T_{2})+\text{Controls},(9) We varied the mixing weight fromw=0w=0to11in increments of0.250.25. To further account for uncertainty in the simulated intermediate usage measure, we additionally added Gaussian noise with progressively larger standard deviations (σ∈{0,0.10,0.25,0.50}\sigma\in\{0,0.10,0.25,0.50\}) to examine whether the findings remained robust under increasing levels of measurement uncertainty.

Figure 5:Sensitivity analysis for the timing of engagement measurement.Standardized coefficients of the simulated intermediate-time engagement measureX⁡(Ti)X(T_{i})predicting follow-up well-being across mixing weightsww(Equation7), wherew=0w=0corresponds to the follow-up (T2T_{2}) measure andw=1w=1to the baseline (T1T_{1}) measure. Shaded bands indicate±1\pm 1standard error, andσ\sigmadenotes the simulated measurement-noise level.Figure5summarizes the sensitivity analysis. Across all six models, the association between chatbot usage and follow-up well-being remained negative across mixing weights and measurement-noise levels. Across the three usage constructs, associations were stronger when the simulated usage measure placed greater weight on the observedT2T_{2}measure and attenuated as greater weight was placed on theT1T_{1}measure. Increasing measurement noise further attenuated the associations toward zero without changing their direction. Overall, the negative associations were robust to alternative combinations ofT1T_{1}andT2T_{2}usage and simulated measurement noise.

A.2Full results for the main longitudinal analyses (N=439)

This section reports the full results corresponding to the primary longitudinal analyses presented in the main text. Analyses include allN=439N=439participants who provided valid follow-up responses, including 108 participants who had discontinued chatbot use. For these participants, follow-up interaction intensity, companionship use, and self-disclosure were coded as zero.

A.2.1Longitudinal development of social chatbot engagement using free-text-derived companionship use

As a complementary operationalization of companionship use, we identified companionship from participants’ free-text descriptions of their relationships with chatbots. Table6reports the longitudinal analyses using this free-text-derived measure, complementing the analyses based on participants’ self-reported companionship use presented in Section2.2.

(1) Interaction intensity T2(2) Companionship T2(3) Self-disclosure T2(4) Continued use T2Interaction intensity T10.44***[0.33, 0.55]p<0.001p<0.0010.20***[0.08, 0.32]p<0.001p<0.0010.28***[0.17, 0.40]p<0.001p<0.0011.97***[1.45, 2.66]p<0.001p<0.001Companionship use T10.03[-0.06, 0.12]p=0.465p=0.4650.22***[0.13, 0.32]p<0.001p<0.0010.04[-0.05, 0.13]p=0.421p=0.4210.98[0.76, 1.27]p=0.875p=0.875Self-disclosure T1-0.02[-0.12, 0.08]p=0.664p=0.6640.01[-0.09, 0.11]p=0.825p=0.8250.13**[0.04, 0.23]p=0.008p=0.0080.86[0.66, 1.11]p=0.241p=0.241Chatbot tenure0.01[-0.07, 0.10]p=0.783p=0.7830.02[-0.07, 0.11]p=0.668p=0.668-0.03[-0.12, 0.06]p=0.562p=0.5621.06[0.83, 1.35]p=0.637p=0.637Age0.11*[0.00, 0.22]p=0.041p=0.0410.16**[0.04, 0.27]p=0.006p=0.0060.08[-0.03, 0.19]p=0.146p=0.1461.18[0.86, 1.63]p=0.303p=0.303Male0.02[-0.16, 0.20]p=0.812p=0.8120.14[-0.04, 0.33]p=0.132p=0.1320.09[-0.10, 0.27]p=0.359p=0.3591.58[0.96, 2.61]p=0.073p=0.073Non-binary-0.36*[-0.67, -0.04]p=0.027p=0.027-0.24[-0.56, 0.09]p=0.156p=0.156-0.39*[-0.71, -0.06]p=0.020p=0.0200.35**[0.16, 0.75]p=0.007p=0.007Single0.09[-0.09, 0.26]p=0.338p=0.3380.08[-0.10, 0.27]p=0.362p=0.3620.07[-0.10, 0.25]p=0.418p=0.4181.01[0.63, 1.61]p=0.969p=0.969Waldχ2\chi^{2}103.1***73.7***82.6***—Selection correlation (ρ\rho)0.260.080.29—NN439439439439

Table 6:Longitudinal predictors of interaction intensity, companionship use, self-disclosure, and continued use at follow-up, using participants’ free-text relationship descriptions to operationalize companionship. Models (1)-(3) report standardized coefficients predicting each follow-up engagement measure from baseline engagement. Model (4) tests whether these baseline engagement patterns predict whether participants continued using chatbots at follow-up. All estimates are reported with 95% confidence intervals and two-sided p-values.∗p<0.05{}^{*}p<0.05;p∗⁣∗<0.01{}^{**}p<0.01;∗∗∗p<0.001{}^{***}p<0.001.

A.2.2Results for the longitudinal engagement pathway to well-being

Table7reports the full results of the longitudinal engagement models examining associations between all three dimensions of social chatbot engagement and subsequent psychological well-being, corresponding to the analyses presented in Section2.3.

Interaction intensityCompanionship (forced choice)Companionship (free-text)Self-disclosure(1)Interaction intensityT2(2)Well-beingT2(3)CompanionshipT2(4)Well-beingT2(5)CompanionshipT2(6)Well-beingT2(7)Self-disclosureT2(8)Well-beingT2Interactionintensity T10.42***[0.32, 0.51]p<0.001p<0.001-0.03[-0.13, 0.06]p=0.509p=0.509——————Interactionintensity T2—-0.08*[-0.14, -0.01]p=0.025p=0.025——————Companionship T1——0.23***[0.13, 0.33]p<0.001p<0.001-0.03[-0.09, 0.03]p=0.274p=0.2740.28***[0.19, 0.37]p<0.001p<0.001-0.03[-0.09, 0.03]p=0.367p=0.367——Companionship T2———-0.06[-0.13, 0.01]p=0.101p=0.101—-0.06*[-0.13, -0.00]p=0.040p=0.040——Self-disclosure T1——————0.29***[0.21, 0.38]p<0.001p<0.001-0.04[-0.12, 0.05]p=0.369p=0.369Self-disclosure T2———————-0.07[-0.13, 0.00]p=0.056p=0.056Well-being T1—0.75***[0.68, 0.81]p<0.001p<0.001—0.76***[0.70, 0.82]p<0.001p<0.001—0.75***[0.69, 0.82]p<0.001p<0.001—0.75***[0.69, 0.81]p<0.001p<0.001Chatbot tenure0.01[-0.07, 0.10]p=0.758p=0.7580.02[-0.04, 0.08]p=0.534p=0.5340.08[-0.01, 0.17]p=0.069p=0.0690.01[-0.05, 0.07]p=0.706p=0.7060.06[-0.03, 0.15]p=0.181p=0.1810.01[-0.05, 0.07]p=0.661p=0.6610.02[-0.07, 0.11]p=0.636p=0.6360.01[-0.05, 0.07]p=0.665p=0.665Age0.09[-0.01, 0.18]p=0.065p=0.0650.02[-0.05, 0.09]p=0.537p=0.5370.16**[0.06, 0.25]p=0.001p=0.0010.01[-0.05, 0.08]p=0.704p=0.7040.16***[0.07, 0.26]p<0.001p<0.0010.01[-0.05, 0.08]p=0.757p=0.7570.09[-0.01, 0.19]p=0.068p=0.0680.02[-0.05, 0.08]p=0.645p=0.645Male0.01[-0.07, 0.10]p=0.750p=0.750-0.01[-0.08, 0.06]p=0.846p=0.8460.04[-0.05, 0.13]p=0.341p=0.3410.00[-0.06, 0.07]p=0.935p=0.9350.04[-0.05, 0.13]p=0.390p=0.3900.00[-0.06, 0.07]p=0.911p=0.9110.02[-0.06, 0.11]p=0.578p=0.578-0.01[-0.07, 0.06]p=0.874p=0.874Non-binary-0.10*[-0.20, -0.01]p=0.031p=0.031-0.03[-0.09, 0.03]p=0.278p=0.278-0.06[-0.16, 0.03]p=0.187p=0.187-0.03[-0.09, 0.03]p=0.317p=0.317-0.05[-0.14, 0.04]p=0.266p=0.266-0.04[-0.09, 0.02]p=0.233p=0.233-0.09[-0.19, 0.01]p=0.071p=0.071-0.04[-0.09, 0.02]p=0.222p=0.222Single0.04[-0.04, 0.13]p=0.334p=0.334-0.00[-0.06, 0.06]p=0.959p=0.9590.04[-0.05, 0.13]p=0.416p=0.416-0.01[-0.08, 0.05]p=0.689p=0.6890.06[-0.03, 0.15]p=0.210p=0.210-0.01[-0.07, 0.06]p=0.845p=0.8450.06[-0.03, 0.15]p=0.199p=0.199-0.01[-0.07, 0.06]p=0.863p=0.863Inverse Mills ratio0.08[-0.02, 0.18]p=0.102p=0.1020.01[-0.07, 0.09]p=0.810p=0.8100.21***[0.12, 0.31]p<0.001p<0.001-0.01[-0.09, 0.06]p=0.757p=0.7570.09[-0.01, 0.19]p=0.080p=0.080-0.01[-0.09, 0.06]p=0.738p=0.7380.18***[0.09, 0.27]p<0.001p<0.001-0.01[-0.08, 0.06]p=0.816p=0.816Direct effect—-0.03[-0.13, 0.06]p=0.509p=0.509—-0.03[-0.09, 0.03]p=0.274p=0.274—-0.03[-0.09, 0.03]p=0.367p=0.367—-0.04[-0.12, 0.05]p=0.369p=0.369Indirect effectviaT2T_{2}interactionpattern—-0.03*[-0.06, -0.00]p=0.025p=0.025—-0.01[-0.03, 0.00]p=0.103p=0.103—-0.02*[-0.04, -0.00]p=0.037p=0.037—-0.02*[-0.04, -0.00]p=0.048p=0.048R2R^{2}0.2350.5790.1410.5750.1280.5770.1520.578NN439439439439439439439439

Table 7:Longitudinal associations between sustained social chatbot engagement and subsequent well-being.Models (1)-(2) examine sustained interaction intensity, Models (3)-(6) examine sustained companionship use using the self-report and free-text operationalizations, and Models (7)-(8) examine sustained self-disclosure. For each engagement measure, the first model predicts follow-up engagement from baseline engagement, and the second predicts follow-up well-being from baseline and follow-up engagement. Indirect effects capture the association between baseline engagement and follow-up well-being through follow-up engagement and are estimated using percentile bootstrap confidence intervals (5,000 resamples). All estimates are reported as standardized coefficients with 95% confidence intervals and two-sidedpp-values.∗p<0.05{}^{*}p<0.05;p∗⁣∗<0.01{}^{**}p<0.01;∗∗∗p<0.001{}^{***}p<0.001.

A.2.3Results for the social displacement pathway to well-being

Table8reports the full results of the longitudinal models examining whether human social interaction accounts for the association between each dimension of social chatbot engagement and subsequent psychological well-being, corresponding to the analyses presented in Section2.4.

Interaction intensityCompanionship (free-text)Self-disclosure(1)InteractionintensityT2(2)Time inpersonT2(3)Well-beingT2(4)CompanionshipT2(5)Time inpersonT2(6)Well-beingT2(7)Self-disclosureT2(8)Time inpersonT2(9)Well-beingT2Interactionintensity T10.42***[0.32, 0.51]p<0.001p<0.001—-0.01[-0.11, 0.08]p=0.759p=0.759——————Interactionintensity T2—-0.17***[-0.27, -0.08]p<0.001p<0.001-0.06[-0.13, 0.01]p=0.074p=0.074——————CompanionshipT1———0.28***[0.19, 0.37]p<0.001p<0.001—-0.02[-0.08, 0.04]p=0.539p=0.539———CompanionshipT2————-0.14**[-0.23, -0.05]p=0.002p=0.002-0.05[-0.11, 0.01]p=0.118p=0.118———Self-disclosure T1——————0.29***[0.21, 0.38]p<0.001p<0.001—-0.03[-0.12, 0.05]p=0.442p=0.442Self-disclosure T2———————-0.15**[-0.24, -0.05]p=0.002p=0.002-0.05[-0.12, 0.02]p=0.152p=0.152Time in person T2——0.14***[0.08, 0.21]p<0.001p<0.001——0.15***[0.08, 0.22]p<0.001p<0.001——0.15***[0.08, 0.21]p<0.001p<0.001Well-being T1——0.73***[0.67, 0.80]p<0.001p<0.001——0.74***[0.67, 0.80]p<0.001p<0.001——0.74***[0.68, 0.80]p<0.001p<0.001Chatbot tenure0.01[-0.07, 0.10]p=0.757p=0.7570.03[-0.06, 0.12]p=0.527p=0.5270.01[-0.05, 0.07]p=0.684p=0.6840.06[-0.03, 0.15]p=0.181p=0.1810.02[-0.07, 0.11]p=0.617p=0.6170.01[-0.05, 0.07]p=0.756p=0.7560.02[-0.07, 0.11]p=0.635p=0.6350.02[-0.07, 0.11]p=0.686p=0.6860.01[-0.05, 0.07]p=0.742p=0.742Age0.09[-0.01, 0.18]p=0.065p=0.065-0.06[-0.15, 0.03]p=0.178p=0.1780.03[-0.03, 0.10]p=0.347p=0.3470.16***[0.07, 0.26]p<0.001p<0.001-0.07[-0.16, 0.02]p=0.137p=0.1370.03[-0.04, 0.09]p=0.444p=0.4440.09[-0.01, 0.19]p=0.068p=0.068-0.07[-0.16, 0.02]p=0.122p=0.1220.03[-0.04, 0.10]p=0.378p=0.378Male0.01[-0.07, 0.10]p=0.750p=0.750-0.22***[-0.31, -0.13]p<0.001p<0.0010.03[-0.04, 0.10]p=0.467p=0.4670.04[-0.05, 0.13]p=0.390p=0.390-0.21***[-0.30, -0.11]p<0.001p<0.0010.03[-0.04, 0.10]p=0.335p=0.3350.02[-0.06, 0.11]p=0.577p=0.577-0.21***[-0.30, -0.12]p<0.001p<0.0010.03[-0.05, 0.10]p=0.475p=0.475Non-binary-0.10*[-0.20, -0.01]p=0.031p=0.031-0.00[-0.11, 0.11]p=0.976p=0.976-0.03[-0.09, 0.03]p=0.274p=0.274-0.05[-0.14, 0.04]p=0.266p=0.2660.00[-0.11, 0.11]p=0.980p=0.980-0.03[-0.09, 0.02]p=0.259p=0.259-0.09[-0.19, 0.01]p=0.070p=0.070-0.00[-0.11, 0.10]p=0.963p=0.963-0.03[-0.09, 0.02]p=0.248p=0.248Single0.04[-0.04, 0.13]p=0.335p=0.335-0.04[-0.13, 0.05]p=0.353p=0.3530.00[-0.06, 0.06]p=0.902p=0.9020.06[-0.03, 0.15]p=0.209p=0.209-0.05[-0.14, 0.04]p=0.305p=0.3050.00[-0.06, 0.06]p=0.972p=0.9720.06[-0.03, 0.15]p=0.198p=0.198-0.05[-0.14, 0.04]p=0.311p=0.3110.00[-0.06, 0.06]p=0.960p=0.960Inverse Mills ratio0.08[-0.02, 0.18]p=0.102p=0.1020.18***[0.08, 0.27]p<0.001p<0.001-0.01[-0.09, 0.07]p=0.859p=0.8590.09[-0.01, 0.19]p=0.081p=0.0810.15**[0.06, 0.25]p=0.001p=0.001-0.02[-0.09, 0.05]p=0.563p=0.5630.18***[0.09, 0.27]p<0.001p<0.0010.17***[0.07, 0.26]p<0.001p<0.001-0.02[-0.09, 0.05]p=0.624p=0.624Direct effect——-0.06[-0.13, 0.01]p=0.074p=0.074——-0.05[-0.11, 0.01]p=0.118p=0.118——-0.05[-0.12, 0.02]p=0.152p=0.152Indirect effect viaT2T_{2}human socialinteraction——-0.02**[-0.05, -0.01]p=0.001p=0.001——-0.02**[-0.04, -0.01]p=0.004p=0.004——-0.02**[-0.04, -0.01]p=0.002p=0.002R2R^{2}0.2350.0960.5840.1280.0890.5820.1520.0890.583NN439439439439439439439439439

Table 8:Longitudinal associations between social chatbot engagement, human social interaction, and subsequent well-being.Models (1)-(3) examine interaction intensity, Models (4)-(6) companionship use measured from free-text responses, and Models (7)-(9) self-disclosure. For each engagement measure, the models predict follow-up engagement, in-person social interaction, and well-being, respectively. Indirect effects through in-person social interaction are estimated using percentile bootstrap confidence intervals (5,000 resamples). All estimates are standardized coefficients with 95% confidence intervals and two-sidedpp-values.∗p<0.05{}^{*}p<0.05;p∗⁣∗<0.01{}^{**}p<0.01;∗∗∗p<0.001{}^{***}p<0.001.

A.3Sensitivity analysis excluding discontinued chatbot users (N=331)

Our primary analyses included allN=439N=439participants who completed the Wave-2 survey. Among them, 108 participants reported that they had discontinued chatbot use by follow-up. Because these participants no longer engaged with the chatbot at Wave 2, their follow-up interaction intensity, companionship use, and self-disclosure were coded as zero in the primary analyses. As a sensitivity analysis, we repeated all analyses after excluding these 108 participants, restricting the sample to theN=331N=331participants who reported continued chatbot use at follow-up.

The results were highly consistent with the primary analyses (Tables9,10, and11). Across the longitudinal structural equation models relating baseline engagement, follow-up engagement, and well-being (Section2.3), all significant indirect effects observed in the primary (N=439N=439) analyses remained statistically significant in the restricted (N=331N=331) sample and showed the same direction of association. Similarly, the mediation models examining reduced in-person social interaction (Section2.4) yielded the same overall pattern: greater chatbot engagement was associated with lower well-being indirectly through reduced in-person social interaction across all three engagement dimensions. Although several coefficients were somewhat larger in the restricted sample, the overall pattern of results and the substantive conclusions remained unchanged. The direct effects of interaction intensity and self-disclosure on well-being were statistically significant in the restricted sample but not in the primary analyses. However, the direction of the effects, as well as the significant indirect and total effects, remained consistent across both analyses, indicating that this difference does not alter the overall interpretation.

Overall, the primary conclusions remained unchanged whether participants who discontinued chatbot use were retained with follow-up engagement coded as zero or excluded from the analyses.

Force-choice companionshipFree-text companionship(1)Interaction intensityT2(2)CompanionshipT2(3)Self-disclosureT2(4)Interaction intensityT2(5)CompanionshipT2(6)Self-disclosureT2Interaction intensity T10.52***[0.40, 0.64]p<0.001p<0.0010.28***[0.15, 0.40]p<0.001p<0.0010.19**[0.07, 0.31]p=0.003p=0.0030.54***[0.42, 0.66]p<0.001p<0.0010.10[-0.03, 0.24]p=0.141p=0.1410.19**[0.07, 0.31]p=0.003p=0.003Companionship T10.55***[0.27, 0.84]p<0.001p<0.0010.63***[0.32, 0.93]p<0.001p<0.0010.20[-0.10, 0.49]p=0.187p=0.1870.08[-0.02, 0.18]p=0.125p=0.1250.32***[0.21, 0.43]p<0.001p<0.0010.08[-0.02, 0.18]p=0.126p=0.126Self-disclosure T1-0.04[-0.14, 0.06]p=0.470p=0.4700.17**[0.06, 0.27]p=0.003p=0.0030.37***[0.26, 0.47]p<0.001p<0.001-0.04[-0.14, 0.07]p=0.504p=0.5040.04[-0.07, 0.16]p=0.455p=0.4550.35***[0.24, 0.46]p<0.001p<0.001Chatbot tenure0.00[-0.09, 0.10]p=0.933p=0.9330.01[-0.09, 0.11]p=0.831p=0.831-0.06[-0.16, 0.04]p=0.230p=0.2300.00[-0.09, 0.10]p=0.939p=0.9390.04[-0.07, 0.14]p=0.520p=0.520-0.06[-0.16, 0.04]p=0.223p=0.223Age0.16**[0.05, 0.27]p=0.005p=0.0050.15*[0.03, 0.27]p=0.016p=0.0160.08[-0.04, 0.20]p=0.177p=0.1770.17**[0.06, 0.29]p=0.003p=0.0030.20**[0.07, 0.33]p=0.002p=0.0020.10[-0.02, 0.22]p=0.112p=0.112Male-0.22*[-0.40, -0.04]p=0.020p=0.0200.17[-0.03, 0.36]p=0.096p=0.096-0.07[-0.26, 0.12]p=0.479p=0.479-0.24*[-0.42, -0.05]p=0.014p=0.0140.07[-0.13, 0.28]p=0.484p=0.484-0.08[-0.27, 0.11]p=0.406p=0.406Non-binary-0.31[-0.70, 0.08]p=0.114p=0.114-0.24[-0.65, 0.17]p=0.254p=0.254-0.28[-0.68, 0.12]p=0.168p=0.168-0.18[-0.57, 0.21]p=0.364p=0.364-0.06[-0.49, 0.38]p=0.803p=0.803-0.22[-0.61, 0.18]p=0.288p=0.288Single0.18[-0.01, 0.37]p=0.057p=0.057-0.03[-0.23, 0.17]p=0.758p=0.7580.08[-0.11, 0.27]p=0.418p=0.4180.15[-0.04, 0.34]p=0.122p=0.1220.10[-0.11, 0.31]p=0.351p=0.3510.07[-0.12, 0.26]p=0.476p=0.476Waldχ2\chi^{2}170.7***111.7***137.3***153.8***67.5***138.5***Selection correlation (ρ\rho)0.310.320.230.240.000.22NN331331331331331331

Table 9:Longitudinal predictors of subsequent chatbot engagement among participants who continued using chatbots at follow-up (N=331N=331), using both participants’ force-choice selection and free-text relationship descriptions to operationalize companionship use. Results report standardized coefficients predicting each follow-up engagement measure from baseline engagement. All estimates are reported with 95% confidence intervals and two-sidedpp-values.∗p<0.05{}^{*}p<0.05;p∗⁣∗<0.01{}^{**}p<0.01;∗∗∗p<0.001{}^{***}p<0.001.Interaction intensityCompanionship (forced choice)Companionship (free-text)Self-disclosure(1)Interaction intensityT2(2)Well-beingT2(3)CompanionshipT2(4)Well-beingT2(5)CompanionshipT2(6)Well-beingT2(7)Self-disclosureT2(8)Well-beingT2Interactionintensity T10.51***[0.42, 0.60]p<0.001p<0.001-0.02[-0.13, 0.09]p=0.757p=0.757——————Companionship T1——0.31***[0.21, 0.41]p<0.001p<0.001-0.05[-0.12, 0.03]p=0.213p=0.2130.37***[0.27, 0.47]p<0.001p<0.001-0.01[-0.08, 0.07]p=0.874p=0.874——Self-disclosure T1——————0.50***[0.40, 0.59]p<0.001p<0.0010.01[-0.10, 0.11]p=0.899p=0.899Focal predictor T2—-0.12**[-0.21, -0.04]p=0.003p=0.003—-0.07[-0.15, 0.02]p=0.109p=0.109—-0.08*[-0.16, -0.01]p=0.031p=0.031—-0.12*[-0.22, -0.02]p=0.020p=0.020Well-being T1—0.74***[0.66, 0.81]p<0.001p<0.001—0.75***[0.67, 0.82]p<0.001p<0.001—0.75***[0.67, 0.82]p<0.001p<0.001—0.75***[0.68, 0.82]p<0.001p<0.001Chatbot tenure0.00[-0.09, 0.10]p=0.928p=0.9280.04[-0.04, 0.11]p=0.321p=0.3210.07[-0.03, 0.18]p=0.167p=0.1670.03[-0.05, 0.10]p=0.467p=0.4670.06[-0.05, 0.16]p=0.283p=0.2830.03[-0.05, 0.10]p=0.493p=0.493-0.02[-0.12, 0.07]p=0.639p=0.6390.02[-0.05, 0.09]p=0.588p=0.588Age0.13**[0.05, 0.21]p=0.002p=0.0020.04[-0.04, 0.11]p=0.367p=0.3670.17***[0.08, 0.27]p<0.001p<0.0010.02[-0.05, 0.09]p=0.604p=0.6040.19***[0.09, 0.28]p<0.001p<0.0010.02[-0.06, 0.09]p=0.684p=0.6840.08*[0.00, 0.16]p=0.046p=0.0460.02[-0.06, 0.10]p=0.610p=0.610Male-0.11*[-0.20, -0.02]p=0.015p=0.015-0.01[-0.09, 0.07]p=0.815p=0.8150.03[-0.07, 0.13]p=0.559p=0.5590.01[-0.06, 0.09]p=0.756p=0.7560.02[-0.09, 0.12]p=0.732p=0.7320.02[-0.06, 0.09]p=0.700p=0.700-0.04[-0.14, 0.05]p=0.380p=0.3800.00[-0.08, 0.08]p=0.989p=0.989Non-binary-0.05[-0.16, 0.06]p=0.362p=0.362-0.00[-0.06, 0.05]p=0.886p=0.886-0.05[-0.17, 0.07]p=0.423p=0.4230.00[-0.05, 0.05]p=0.900p=0.900-0.00[-0.11, 0.10]p=0.967p=0.967-0.01[-0.06, 0.05]p=0.831p=0.831-0.05[-0.15, 0.05]p=0.335p=0.335-0.01[-0.06, 0.05]p=0.768p=0.768Single0.07[-0.02, 0.16]p=0.142p=0.142-0.02[-0.10, 0.05]p=0.548p=0.548-0.00[-0.11, 0.11]p=0.963p=0.963-0.04[-0.12, 0.04]p=0.304p=0.3040.05[-0.05, 0.16]p=0.337p=0.337-0.03[-0.11, 0.05]p=0.454p=0.4540.03[-0.07, 0.13]p=0.530p=0.530-0.03[-0.10, 0.05]p=0.483p=0.483Inverse Mills ratio0.04[-0.04, 0.13]p=0.329p=0.329-0.00[-0.08, 0.08]p=0.941p=0.9410.11[-0.00, 0.22]p=0.059p=0.059-0.03[-0.11, 0.06]p=0.546p=0.5460.01[-0.09, 0.12]p=0.795p=0.795-0.03[-0.10, 0.05]p=0.532p=0.5320.08[-0.02, 0.18]p=0.111p=0.111-0.02[-0.09, 0.06]p=0.686p=0.686Direct effect—-0.02[-0.13, 0.09]p=0.757p=0.757—-0.05[-0.12, 0.03]p=0.213p=0.213—-0.01[-0.08, 0.07]p=0.874p=0.874—0.01[-0.10, 0.11]p=0.899p=0.899Indirect effectviaT2T_{2}interactionpattern—-0.06**[-0.12, -0.02]p=0.001p=0.001—-0.02[-0.05, 0.01]p=0.117p=0.117—-0.03*[-0.06, -0.00]p=0.027p=0.027—-0.06*[-0.11, -0.01]p=0.021p=0.021R2R^{2}0.3230.5700.1370.5660.1570.5660.2740.572NN331331331331331331331331

Table 10:Longitudinal associations between sustained social chatbot engagement and subsequent well-being among continued users.Models and estimates correspond to those reported in Table7, with analyses restricted to participants who continued using chatbots at follow-up (N=331N=331). All estimates are reported as standardized coefficients with 95% confidence intervals and two-sidedpp-values.∗p<0.05{}^{*}p<0.05;p∗⁣∗<0.01{}^{**}p<0.01;∗∗∗p<0.001{}^{***}p<0.001.Interaction intensityCompanionship (free-text)Self-disclosure(1)InteractionintensityT2(2)Time inpersonT2(3)Well-beingT2(4)CompanionshipT2(5)Time inpersonT2(6)Well-beingT2(7)Self-disclosureT2(8)Time inpersonT2(9)Well-beingT2Interaction intensity T10.51***[0.42, 0.60]p<0.001p<0.001—-0.00[-0.11, 0.11]p=0.940p=0.940——————Companionship T1———0.37***[0.27, 0.47]p<0.001p<0.001—0.01[-0.06, 0.08]p=0.858p=0.858———Self-disclosure T1——————0.50***[0.40, 0.59]p<0.001p<0.001—0.01[-0.10, 0.12]p=0.839p=0.839Interaction intensity T2—-0.16**[-0.27, -0.05]p=0.004p=0.004-0.11**[-0.20, -0.03]p=0.009p=0.009——————Companionship T2————-0.11*[-0.21, -0.01]p=0.027p=0.027-0.07[-0.15, 0.00]p=0.057p=0.057———Self-disclosure T2———————-0.11*[-0.22, -0.01]p=0.034p=0.034-0.11*[-0.21, -0.01]p=0.034p=0.034Time in person T2——0.14***[0.06, 0.22]p<0.001p<0.001——0.15***[0.07, 0.23]p<0.001p<0.001——0.14***[0.07, 0.22]p<0.001p<0.001Well-being T1——0.73***[0.65, 0.80]p<0.001p<0.001——0.73***[0.66, 0.81]p<0.001p<0.001——0.73***[0.66, 0.81]p<0.001p<0.001Chatbot tenure0.00[-0.09, 0.10]p=0.928p=0.9280.06[-0.05, 0.16]p=0.278p=0.2780.03[-0.04, 0.10]p=0.428p=0.4280.06[-0.05, 0.16]p=0.283p=0.2830.05[-0.06, 0.15]p=0.376p=0.3760.02[-0.05, 0.09]p=0.589p=0.589-0.02[-0.12, 0.07]p=0.638p=0.6380.04[-0.06, 0.14]p=0.462p=0.4620.02[-0.05, 0.08]p=0.668p=0.668Age0.13**[0.05, 0.21]p=0.002p=0.002-0.03[-0.14, 0.07]p=0.541p=0.5410.05[-0.03, 0.13]p=0.261p=0.2610.19***[0.09, 0.28]p<0.001p<0.001-0.05[-0.15, 0.05]p=0.342p=0.3420.03[-0.05, 0.11]p=0.438p=0.4380.08*[0.00, 0.16]p=0.046p=0.046-0.05[-0.16, 0.05]p=0.331p=0.3310.03[-0.05, 0.11]p=0.405p=0.405Male-0.11*[-0.20, -0.02]p=0.015p=0.015-0.28***[-0.38, -0.18]p<0.001p<0.0010.03[-0.05, 0.11]p=0.475p=0.4750.02[-0.09, 0.12]p=0.732p=0.732-0.25***[-0.35, -0.15]p<0.001p<0.0010.05[-0.03, 0.13]p=0.196p=0.196-0.04[-0.14, 0.05]p=0.380p=0.380-0.26***[-0.37, -0.16]p<0.001p<0.0010.04[-0.04, 0.12]p=0.349p=0.349Non-binary-0.05[-0.16, 0.06]p=0.362p=0.362-0.03[-0.15, 0.10]p=0.675p=0.675-0.00[-0.05, 0.05]p=0.965p=0.965-0.00[-0.11, 0.10]p=0.967p=0.967-0.03[-0.16, 0.10]p=0.672p=0.672-0.00[-0.05, 0.05]p=0.965p=0.965-0.05[-0.14, 0.05]p=0.335p=0.335-0.03[-0.15, 0.10]p=0.652p=0.652-0.00[-0.05, 0.04]p=0.872p=0.872Single0.07[-0.02, 0.16]p=0.142p=0.142-0.02[-0.13, 0.09]p=0.727p=0.727-0.02[-0.09, 0.06]p=0.631p=0.6310.05[-0.05, 0.16]p=0.337p=0.337-0.03[-0.13, 0.08]p=0.655p=0.655-0.02[-0.10, 0.05]p=0.543p=0.5430.03[-0.07, 0.13]p=0.530p=0.530-0.02[-0.13, 0.08]p=0.657p=0.657-0.02[-0.10, 0.05]p=0.568p=0.568Inverse Mills ratio0.04[-0.04, 0.13]p=0.329p=0.3290.18***[0.07, 0.28]p<0.001p<0.001-0.02[-0.10, 0.06]p=0.649p=0.6490.01[-0.09, 0.12]p=0.795p=0.7950.16**[0.05, 0.26]p=0.003p=0.003-0.04[-0.12, 0.04]p=0.343p=0.3430.08[-0.02, 0.18]p=0.112p=0.1120.17**[0.06, 0.27]p=0.002p=0.002-0.03[-0.11, 0.05]p=0.466p=0.466Direct effect——-0.11**[-0.20, -0.03]p=0.009p=0.009——-0.07[-0.15, 0.00]p=0.057p=0.057——-0.11*[-0.21, -0.01]p=0.034p=0.034Indirect effect viaT2T_{2}human socialinteraction——-0.02**[-0.05, -0.01]p=0.005p=0.005——-0.02*[-0.04, -0.00]p=0.030p=0.030——-0.02*[-0.04, -0.00]p=0.040p=0.040R2R^{2}0.3230.1160.5760.1570.1040.5730.2740.1040.579NN331331331331331331331331

Table 11:Displacement pathway between social chatbot engagement and well-being through in-person social interaction among continued users.Models and estimates correspond to those reported in Table8, with analyses restricted to participants who continued using chatbots at follow-up (N=331N=331). All estimates are reported as standardized coefficients with 95% confidence intervals and two-sidedpp-values.∗p<0.05{}^{*}p<0.05;p∗⁣∗<0.01{}^{**}p<0.01;∗∗∗p<0.001{}^{***}p<0.001.

A.4Relationship self-description classification prompt

{

“role”:“system”,

“content”:“”“

Youareaqualitativecodingassistant.

Task:Classifyasurveyresponseabouthowsomeoneviewsoruseschatbots.Independentlydecide,forEACHofthefourusagetypesbelow,whethertheresponseshowsanysignofthattypeofuse.

Usagetypes(definitions):

-Productivity:Usingchatbotstoobtainassistanceorinformation.

-Entertainment:Usingchatbotsforfunortopassthetime.

-Social/Relational:Theuserrelatestothechatbotasapersonalorinterpersonalcounterpart,orusesittostrengthensocialtieswithotherpeople.

-Novelty/Curiosity:Usingchatbotsoutofcuriosityortoexploretheircapabilities.

Labelingrules:

-Foreachtype,label1iftheresponsecontainsANYindicationofthatuse(evenifitisnotthemainpurpose),otherwiselabel0.

-ThefourtypesareNOTmutuallyexclusive:aresponsecanbe1onmultipletypes.

-Judgeeachtypeindependently.

Examples:

Input:’assistantforcodingandresearch’->{“productivity”:1,“entertainment”:0,“social_relational”:0,“curiosity”:0}

Input:’Italktoitlikeafriendsometimes,andforfun’->{“productivity”:0,“entertainment”:1,“social_relational”:1,“curiosity”:0}

Input:’helper,adviser,mentor’->{“productivity”:1,“entertainment”:0,“social_relational”:1,“curiosity”:0}

Input:’roleplaypartner’->{“productivity”:0,“entertainment”:1,“social_relational”:1,“curiosity”:0}

Input:’storytellingandroleplayingafictionalcharacterformywriting’->{“productivity”:0,“entertainment”:1,“social_relational”:0,“curiosity”:0}

Input:’wantedtoseewhatitcoulddo,anditalsohelpsmedraftemails’->{“productivity”:1,“entertainment”:0,“social_relational”:0,“curiosity”:1}

Input:’mycompaniontoventto,andithelpswithhomework’->{“productivity”:1,“entertainment”:0,“social_relational”:1,“curiosity”:0}

ReplywithONLYaJSONobjectwithexactlythesekeysandintegervalues0or1:

{“productivity”:0,“entertainment”:0,“social_relational”:0,“curiosity”:0}

“”“

}