Article Summary

1. Title of the Paper

The Heterogeneous Effects of AI Companionship: An Empirical Model of Chatbot Usage and Loneliness and a Typology of User Archetypes

2. Publication Details

Authors: Auren R. Liu, Pat Pataranutaporn, and Pattie Maes
Affiliation: MIT Media Lab
Publication Date: October 2025
Publication: Proceedings of the Eighth AAAI/ACM Conference on AI, Ethics, and Society (AIES 2025), 8(2), 1585–1597
DOI: https://doi.org/10.1609/aies.v8i2.36658
Publication Type: Peer-reviewed conference proceedings (AIES 2025); no journal impact factor.

3. Research Background and Objectives

The study addresses a specific gap in research on AI companionship: it remains unclear which user characteristics determine whether interactions with companion chatbots are associated with reduced or increased loneliness. The authors investigate how individual psychological characteristics and social circumstances may help explain these different outcomes (Liu et al., 2025).

4. Research Questions

The study focused on the following questions:

• What motivates individuals to use companion chatbots, and what purposes do these systems serve?

• How is companion chatbot use associated with loneliness, and which psychological or social factors may help explain this relationship?

• Can user profiles be identified based on chatbot-use patterns and psychosocial characteristics?

5. Methodology

Research Design and Participants

The researchers conducted a cross-sectional survey involving 404 adults recruited through CloudResearch Connect. Participants were required to be at least 18 years old, fluent in English, and to have used a chatbot at least once a week for a minimum of one month.

The sample was predominantly from the United States (93.6%), with 66.3% identifying as White and 59.7% as men. The most frequently reported platforms were Snapchat My AI (31.6%), Character.ai (30.4%), and Replika (22.6%). The inclusion of Snapchat My AI is noteworthy because it is not exclusively designed as a companion chatbot, which raises a question about how companion chatbot use was defined and represented in the study.

Measures

The researchers used several psychological and social measures. Loneliness was assessed using the eight-item UCLA Loneliness Scale (ULS-8), while social networks were measured using the six-item Lubben Social Network Scale (LSNS-6). Personality characteristics were assessed using the Big Five Inventory–10 (BFI-10), the 10-item short form of the Big Five Inventory.

Problematic chatbot use was measured using an adapted version of the Generalized Problematic Internet Use Scale 2 (GPIUS2). This is an important methodological consideration because the instrument was adapted from a measure of problematic internet use rather than being a validated chatbot-specific scale.

Data Analysis

The researchers used ordinary least squares (OLS) regression with exploratory backward elimination to examine factors associated with loneliness. Mediation analysis was conducted to investigate whether problematic use and social-network size helped explain the relationship between chatbot session length and loneliness. K-means clustering was used to identify seven user profiles based on chatbot engagement and psychosocial characteristics.

The researchers also conducted qualitative analysis of participants' responses. They manually coded a stratified sample of 105 responses, selecting 15 responses from each of the seven clusters. In addition, GPT-4 was used to extract topics from responses across all 404 participants. The researchers reviewed the AI-generated output and compared it with the manually identified themes.

6. Major Findings

Chatbot Use and Loneliness

The findings showed that chatbot usage frequency was not significantly associated with loneliness. Although longer chatbot sessions initially showed a small positive correlation with loneliness (r = .10), this relationship became statistically non-significant after other variables were controlled. This suggests that session length alone did not independently predict loneliness.

Problematic chatbot use, however, was significantly associated with higher loneliness. The mediation analysis identified a significant specific indirect effect through problematic use, but the total indirect effect was not statistically significant. Therefore, the findings do not establish an overall significant mediation effect.

Psychological and Social Factors

Personality and social characteristics were important predictors of loneliness. Neuroticism and problematic chatbot use were associated with higher loneliness, whereas larger social networks and agreeableness were associated with lower loneliness. The final regression model explained approximately 49.6% of the variance in loneliness, with neuroticism and social network size showing the largest standardized coefficients.

The researchers also found an interaction between chatbot session length and attraction to a close person. Longer sessions were associated with slightly higher loneliness among participants who reported stronger attraction to someone close.

Seven User Profiles

The K-means analysis identified seven user profiles with different patterns of chatbot engagement and psychosocial characteristics. Two profiles illustrate that problematic use does not necessarily correspond to greater loneliness:

• Fulfilled Dependent Users: These users had high problematic chatbot-use scores but below-average loneliness. They also reported adequate social connections and positive real-world social interactions.

• Lonely Moderate Users: These users also had high problematic chatbot-use scores but experienced high loneliness. They reported lower socialization and greater interest in emotional support and companionship

The authors identified three profiles—Lonely Moderate Users, Lonely Light Users, and Socially Challenged Frequent Users—as particularly vulnerable because of their high loneliness.

However, the clusters should be interpreted cautiously. The average silhouette score was −0.04, indicating substantial overlap between the identified groups. Box's M-test also indicated unequal covariance matrices, further limiting confidence in the separation of the profiles.

Reasons for Chatbot Use

Technology exploration (30.9%) and entertainment (26.4%) were the most frequently reported initial reasons for chatbot use. Companionship accounted for only 12.2% of initial motivations. This indicates that emotional companionship was not the primary reason most participants began using these chatbots.

7. Discussion and Limitations

The authors interpret the relationship between problematic chatbot use and loneliness through compensatory internet use theory. People experiencing social deficits may turn to online technologies to meet unmet social or emotional needs. However, when this use becomes problematic, it may reinforce the difficulties it was intended to relieve, potentially contributing to greater loneliness. This interpretation helps explain why chatbot engagement may be associated with different loneliness outcomes across users.

Limitations

Several limitations should be considered when interpreting the findings. First, the study relied on self-reported measures, which may be affected by recall bias. Its cross-sectional design also prevents conclusions about causal relationships. The sample was predominantly from the United States (93.6%), White (66.3%), and male (59.7%), which may limit the generalizability of the findings to other populations. Additional limitations included the use of exploratory backward elimination in regression analyses, shortened measurement scales, and coarse self-reported chatbot-use items. The adapted GPIUS2 measure was not specifically validated for chatbot use, and the clustering results showed considerable overlap between profiles. Furthermore, GPT-4-based theme extraction may have introduced interpretive bias despite researcher review. Finally, participants' mental health treatment status was not assessed, limiting the interpretation of their psychological characteristics and experiences.

8. Conclusion

The study found that chatbot usage frequency was not significantly associated with loneliness, while session length showed only a small initial association that disappeared after statistical adjustment. Problematic chatbot use and users' psychological and social characteristics were more informative, and the contrast between Fulfilled Dependent Users and Lonely Moderate Users shows that similarly high problematic-use scores can coexist with very different levels of loneliness. Because the study was cross-sectional, these findings describe associations rather than causal effects.