Source-linked AI summary
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
TL;DR
Prior research leaves unclear how AI-companion engagement develops over time and relates to well-being, including through users’ interactions with other people. This two-wave study of Character.AI users found that sustained engagement was associated with lower well-being through continued engagement and lower in-person interaction.
Problem
Prior work leaves the longer-term development of AI-companion engagement and its relationship to well-being poorly understood, including how it relates to human social interaction.
Source: §1 · p. 3Method
A two-wave longitudinal study followed Character.AI users and assessed interaction intensity, companionship use, self-disclosure, well-being, and in-person interaction.
Source: §1 · p. 3Results
Sustained engagement across all three dimensions was associated with lower follow-up well-being through continued engagement and lower in-person interaction.
Source: §1 · p. 3, §2.3 · p. 8Takeaways & Limitations
The findings suggest evaluating engagement patterns over time and whether AI companions support human interaction rather than maximize interaction with AI.
Source: §3 · p. 10With only two measurement waves, the study cannot fully characterize dynamic and potentially reciprocal relationships between AI companionship and well-being over longer periods.
Source: §3.1 · p. 11
Abstract
from arXiv · showhide
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 CharacterAI 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
1 Introduction
Prior findings about AI companions and well-being conflict, while the longer-term development of engagement and its relationship to human interaction remain poorly understood. This two-wave study of Character.AI users finds that sustained engagement across three dimensions was associated with lower well-being through continued engagement and lower in-person interaction.
1 Introduction
- Prior studies report both potential benefits, such as reduced loneliness and emotional support, and associations with loneliness, reliance, and reduced offline engagement.The introduction argues that this mixed picture calls for examining how engagement shapes well-being rather than asking only whether companionship is beneficial or harmful.Source: §1 · p. 1
- The study examines how AI-companion engagement develops over time and how sustained engagement and human social interaction relate to well-being.It characterizes engagement as interaction intensity, companionship use, and self-disclosure.Source: §1 · pp. 2–3
- The study followed Character.AI users from baseline to a follow-up about 12 months later, assessing engagement across three dimensions.The baseline survey included 1,182 participants and the follow-up included 439.Source: Figure 1 · p. 2, §1 · p. 3
- Greater baseline interaction intensity predicted stronger subsequent engagement, and sustained engagement was associated with lower well-being through both engagement over time and lower in-person interaction.The authors conclude that understanding these associations requires considering engagement trajectories and users’ broader social lives.Source: §1 · p. 3
2 Results
Across the four results subsections, chatbot relationships included mixed and companion-oriented uses, and engagement persisted and developed across dimensions. Sustained engagement was associated with lower well-being through both follow-up engagement and lower in-person interaction.
2.1 Chatbot engagement sustains relational and mixed forms over time
- More than half of continuing users described multiple types of chatbot use at both waves, and supportive-companion descriptions became the most common category at follow-up.Supportive-companion descriptions increased from 40.2% at T1 to 42.3% at T2; tool, assistant, and system descriptions remained prevalent, changing from 43.8% to 40.5%.Source: §2.1 · p. 4, Figure 2 · p. 5
2.2 Social chatbot engagement exhibits reinforcing dynamics over time
- Baseline interaction intensity predicted all three engagement dimensions at follow-up, while companionship use and self-disclosure also persisted over time.Baseline companionship use predicted subsequent interaction intensity, but its association with subsequent self-disclosure was not significant; higher baseline interaction intensity also predicted continued chatbot use (OR = 1.95, p < 0.001). Baseline self-disclosure was not significantly associated with subsequent interaction intensity or companionship use.Source: §2.2 · pp. 4, 6, 7
2.3 Sustained social chatbot engagement was consistently associated with lower well-being over time across all three dimensions
- Across interaction intensity, companionship use, and self-disclosure, sustained engagement had significant negative indirect associations with follow-up well-being, while baseline engagement had no significant direct associations.The models assess indirect associations through follow-up engagement while accounting for baseline well-being and engagement.Source: Figure 3 · p. 7, §2.3 · p. 8
2.4 Social chatbot engagement was associated with lower well-being through displacement of human social interaction
- Greater follow-up engagement across all three dimensions was associated with less in-person interaction, which in turn was associated with higher well-being.Each dimension showed a significant negative indirect association with follow-up well-being through in-person interaction; direct associations were not statistically significant after accounting for it.Source: §2.4 · p. 8
3 Discussion
The discussion interprets sustained AI engagement and lower human interaction as pathways associated with lower well-being, and argues that companion design should support human relationships. These conclusions are bounded by limits in the study’s waves, measures, follow-up, and setting.
3 Discussion
- Stronger earlier engagement predicted stronger subsequent engagement, which was associated with lower well-being, including among users who did not initially seek companionship.The discussion notes that companion-like roles may emerge through repeated interaction, including in general-purpose chatbots.Source: §3 · pp. 8–9
- General-purpose chatbots may acquire companion-like roles through repeated interaction, so safety evaluations should consider actual use rather than intended roles alone.The discussion also cautions that trust from ordinary interactions may carry into sensitive contexts such as mental health support.Source: §3 · pp. 9–10
- Displacement may affect users’ social circles too: redirecting attention to AI companions can reduce others’ opportunities for interaction and, if widespread, weaken social participation.The discussion connects this conditional possibility to the networks, reciprocity, and trust through which social capital is maintained.Source: §3 · p. 10
- The authors argue that companion design should assess sustained engagement and whether it displaces human interaction, and consider facilitating shared activities with family, partners, or friends.This would position AI companions as support for relationships between people rather than competitors for social attention.Source: §3 · pp. 10–11
3.1 Limitations
- With only two waves, the study cannot fully characterize dynamic, potentially reciprocal relationships between AI companionship and well-being; sensitivity analyses cannot replace denser longitudinal data.The authors suggest three or more waves, ecological momentary assessment, or continuous behavioral logging for richer evidence.Source: §3.1 · p. 11
- Companionship-use measures changed between waves, limiting direct comparability, while self-reports may reflect recall, reporting, and interpretation differences; unobserved platform changes also remain possible.The study lacked access to Character.AI’s internal model updates, moderation policies, and design changes, so their occurrence and effects could not be assessed.Source: §3.1 · p. 11
- Attrition analyses cannot eliminate bias from unobserved factors, and the Character.AI, English-speaking US Prolific sample may not generalize across platforms or populations.Only 439 of 1,182 baseline participants returned at follow-up; many others were no longer active on Prolific.Source: §3.1 · pp. 11–12
- The longitudinal process may vary across users; which individual and contextual factors shape complementary support versus reduced human engagement remains unanswered.The authors identify loneliness, offline social support, attachment style, prior mental health, and usage motivations as factors for future study.Source: §3.1 · p. 12
4 Methods
The study followed Character.AI users across two waves and used longitudinal models to examine engagement, well-being, and in-person interaction. Measures covered three dimensions of chatbot engagement, well-being, human interaction, and perceived anthropomorphism.
4.1 Participants and data collection
- The two-wave study followed 1,182 baseline respondents, with 439 returning after about a year; 108 who had stopped chatbot use were coded zero on follow-up engagement measures.The analyses included all returning participants and applied a Heckman selection correction for potential non-random attrition.Source: §4 · p. 12, §4.1 · p. 13
4.2 Selection and attrition analysis
- A Heckman selection framework modeled follow-up completion and supplied an inverse Mills ratio for the well-being models to address potential non-random attrition.The selection model used baseline well-being, engagement, demographics, chatbot tenure, and prior Prolific approvals.Source: §4.2 · p. 13
4.3 Longitudinal analysis overview
- Longitudinal models tested whether baseline engagement predicted each engagement dimension at follow-up, and whether sustained engagement and in-person interaction were associated with follow-up well-being.Models adjusted for baseline chatbot tenure, age, gender, and relationship status; attrition correction was incorporated.Source: §4.3 · pp. 13–14
4.4.1 Identifying companion interaction
- Companionship use combined participants’ reported chatbot-use purposes with free-text descriptions classified into companionship, entertainment, productivity, and curiosity categories.Two annotators validated the GPT-5-mini classifications, with 95.8% average label-level agreement and Fleiss’ κ = 0.89.Source: §4.4.1 · p. 15
4.4.2 Measuring subjective well-being
- Well-being was a six-item mean adapted from the Comprehensive Inventory of Thriving, spanning life satisfaction, feelings, loneliness, social support, and belonging.Negative feelings and loneliness were reverse-coded; internal consistency was α=0.88 at T1 and 0.89 at T2.Source: Table 3 · p. 15, §4.4.2 · p. 16
4.4.3 Measuring chatbot interaction intensity
- Chatbot interaction intensity combined standardized behavioral-engagement and emotional-integration items into a composite score.The scale had seven items and internal consistency of α = 0.86 at T1 and 0.82 at T2.Source: §4.4.3 · p. 16, Table 4 · p. 16
4.4.4 Measuring self-disclosure levels
- Self-disclosure captured willingness to share personal thoughts, feelings, and experiences with chatbots using three selected MOCA items.The seven-point items had α=0.89 at T1 and 0.93 at T2.Source: §4.4.4 · p. 17, Table 5 · p. 17
4.4.5 Measuring human social interaction; 4.4.6 Measuring anthropomorphism
- Human interaction was measured as daily in-person time on a six-point scale, while anthropomorphism was a single seven-point rating of chatbot humanlikeness.Source: §4.4.5 · p. 17, §4.4.6 · p. 17
A.1 Robustness check: sensitivity analysis
The sensitivity analysis replaced observed follow-up chatbot usage with simulated intermediate usage measures constructed from baseline and follow-up scores. Negative associations with follow-up well-being persisted across tested mixing weights and measurement-noise levels.
A.1 Robustness check: sensitivity analysis
- Simulated intermediate usage was a weighted combination of standardized baseline and follow-up usage, with Gaussian noise representing measurement uncertainty.The analysis varied w from 0 to 1 in 0.25 increments and σ across 0, 0.10, 0.25, and 0.50.Source: §A.1 · p. 23
- For each of three usage constructs, the simulated measure replaced observed follow-up usage in the six longitudinal models.The analysis tested whether the estimated usage–well-being association remained consistent under this alternative measurement.Source: §A.1 · p. 23
- Across all six models, chatbot usage remained negatively associated with follow-up well-being across mixing weights and measurement-noise levels.Associations weakened as simulated usage placed more weight on baseline usage or added more measurement noise, but their direction did not change.Source: §A.1 · p. 23
A.2 Full results for the main longitudinal analyses (N=439)
The reported full longitudinal results include all 439 participants with valid follow-up responses, including those who had stopped using chatbots. Figure 5 describes a sensitivity analysis of simulated intermediate-time engagement measures.
A.2 Full results for the main longitudinal analyses (N=439)
- The primary longitudinal analyses included all 439 valid follow-up respondents, including 108 who had discontinued chatbot use; their follow-up engagement scores were coded as zero.Source: §A.2 · p. 23
- Figure 5 plots standardized coefficients for simulated intermediate-time engagement predicting follow-up well-being across mixing weights.The weight ranges from the follow-up measure at w=0 to the baseline measure at w=1.Source: Figure 5 · p. 24
- Figure 5’s shaded bands indicate ±1 standard error, and σ marks the simulated measurement-noise level.Source: Figure 5 · p. 24
A.2.1 Longitudinal development of social chatbot engagement using free-text-derived companionship use
This subsection reports longitudinal analyses using free-text-derived companionship use as a complementary measure to participants’ self-reported companionship use.
A.2.1 Longitudinal development of social chatbot engagement using free-text-derived companionship use
- Free-text descriptions of chatbot relationships provided a complementary operationalization of companionship use.Source: §A.2.1 · p. 24
- Table 6 reports longitudinal analyses using the free-text-derived companionship measure.Source: §A.2.1 · p. 24
- The free-text-derived analyses complement those based on self-reported companionship use in Section 2.2.Source: §A.2.1 · p. 24
A.2.2 Results for the longitudinal engagement pathway to well-being
The appendix identifies Table 7 as the full set of longitudinal engagement-pathway models relating three dimensions of chatbot engagement to subsequent well-being. Table 6 describes models of later engagement and continued use.
A.2.2 Results for the longitudinal engagement pathway to well-being
- Table 7 reports longitudinal associations between all three dimensions of social chatbot engagement and subsequent psychological well-being.The appendix identifies these as the full results for the analyses in Section 2.3.Source: §A.2.2 · p. 24
- Table 6 predicts follow-up interaction intensity, companionship use, and self-disclosure from baseline engagement.The table also tests whether baseline engagement predicts continued chatbot use at follow-up.Source: Table 6 · p. 25
- Table 6 operationalizes companionship using participants’ free-text relationship descriptions.Its models report standardized coefficients with 95% confidence intervals and two-sided p-values.Source: Table 6 · p. 25
A.2.3 Results for the social displacement pathway to well-being
The appendix describes models examining whether in-person social interaction accounts for links between chatbot engagement and subsequent well-being. It also identifies the longitudinal engagement models used to examine sustained engagement.
A.2.3 Results for the social displacement pathway to well-being
- Table 8 examines whether human social interaction accounts for associations between all three engagement dimensions and subsequent well-being.The appendix identifies these as the social displacement pathway analyses in Section 2.4.Source: §A.2.3 · p. 25
- Table 7 examines sustained interaction intensity, companionship use, and self-disclosure, with indirect effects through follow-up engagement.Companionship is operationalized using self-report and free-text measures; indirect effects use 5,000 bootstrap resamples.Source: Table 7 · p. 26
- Table 8 models interaction intensity, free-text companionship use, and self-disclosure in relation to follow-up engagement, in-person social interaction, and well-being.Indirect effects through in-person social interaction use percentile bootstrap confidence intervals from 5,000 resamples.Source: Table 8 · p. 27
A.3 Sensitivity analysis excluding discontinued chatbot users (N=331)
The sensitivity analyses excluding discontinued chatbot users retained the primary analyses’ overall conclusions. The primary analyses coded discontinued users’ follow-up engagement as zero, while the sensitivity sample included continuing users only.
A.3 Sensitivity analysis excluding discontinued chatbot users (N=331)
- The primary analyses included 439 follow-up respondents, coding follow-up engagement as zero for 108 who had discontinued chatbot use.The sensitivity analysis excluded those 108 participants, leaving 331 continued users.Source: §A.3 · p. 28
- Direct effects of interaction intensity and self-disclosure on well-being were significant in the restricted sample but not in the primary analyses.The indirect and total effects remained significant and directionally consistent across analyses.Source: §A.3 · p. 28
- The overall conclusions remained unchanged when discontinued users were excluded rather than retained with follow-up engagement coded as zero.The restricted-sample models showed the same overall pattern for the displacement pathway and consistent indirect and total effects.Source: §A.3 · p. 28
- Tables 9–11 report restricted-sample models of later engagement, sustained engagement and well-being, and the displacement pathway.The tables use participants who continued chatbot use at follow-up (N = 331).Source: Table 9 · p. 29, Table 10 · p. 30, Table 11 · p. 31
A.4 Relationship self-description classification prompt {
The prompt classifies free-text descriptions of chatbot use into four usage types. It requires independent binary labels, allowing multiple types to apply to one response.
A.4 Relationship self-description classification prompt {
- The prompt asks the classifier to assess productivity, entertainment, social/relational use, and novelty/curiosity in each response.It defines these respectively as assistance or information, fun or passing time, relating to the chatbot as a counterpart or strengthening human ties, and curiosity or capability exploration.Source: §A.4 · p. 32
- Each usage type is labeled 1 when the response contains any indication of that use, otherwise 0.The prompt directs the classifier to judge each type independently.Source: §A.4 · p. 32
- Examples distinguish instrumental assistance from relational uses, entertainment, and curiosity, while allowing categories to overlap.For example, a response describing a companion for venting and homework can be labeled both productivity and social/relational.Source: §A.4 · p. 32
- The output must be a JSON object with the four named keys and integer values of 0 or 1.The example responses show that a response may receive multiple positive labels.Source: §A.4 · p. 32