As large language model (LLM)-enhanced chatbots become increasingly expressive and socially responsive, many users begin forming companionship-like bonds with them. This study investigates how using AI companions relates to psychological well-being. We collected self-reported data from 1,131 U.S. adults who use CharacterAI, including survey responses and 4,664 chat sessions (464,687 messages) from 237 participants. By triangulating self-reported usage, relationship descriptions, and real chat histories, we identify patterns of engagement and associated outcomes. Smaller social networks were associated with reporting companionship as the primary chatbot use (beta = -0.03, 95% confidence interval (CI) [-0.05, -0.01]), which in turn was associated with lower well-being (beta = -0.48, 95% CI [-0.70, -0.25]). For self-reported companionship usage, this association was stronger when interactions were intensive (beta = -0.31, 95% CI [-0.56, -0.06]) and highly disclosive (beta = -0.38, 95% CI [-0.63, -0.14]). These results suggest that the association between AI companionship and well-being is not uniform and depends on how chatbots are used and users' offline social environments.
Our findings suggest that socially oriented chatbot use is consistently associated with lower well-being, particularly among users who engage more intensively or exhibit higher levels of emotional self-disclosure.
cross-sectional evidence supports an association, but moderation appears only in the self-reported primary-use model
participants who self-reported using chatbots primarily for companionship reported significantly lower well-being compared to those with other motivations
the primary-use regression supports the negative association, but cross-sectional measurement leaves temporal direction and residual confounding unresolved
Individuals with fewer close social ties were more likely to form companionship-oriented relationships with chatbots
cross-sectional primary-use data show a small association, without establishing whether limited social ties precede companionship-seeking
the negative association between companionship use of chatbots and well-being was stronger among users who used the chatbots more intensively
the self-reported model shows an interaction, but cross-sectional measurement cannot establish direction and other operationalizations do not reproduce it
These results indicate no evidence of a meaningful moderation effect of companionship-oriented chatbot use on the association between offline social network size and well-being.
cross-sectional models detect no interaction, but the reported interval and approximate equivalence check do not establish meaningful absence
Derived from the full evaluation — not a separate score.
Strengths
The study combines survey responses with donated conversations and distinguishes primary use, relationship descriptions, and session-level companionship. This provides more behavioral coverage than a single self-report indicator.
↳ Methods §§4.1–4.2; Results §2.1
The paper reports internal consistency for survey scales and human agreement checks for model-generated classifications and summaries. These checks provide evidence about measurement execution, although they do not resolve the conflicting classification instructions.
↳ Methods §§4.2.1–4.2.4
The Introduction acknowledges both benefits and risks of AI companionship. The limitations explicitly connect cross-sectional inference, donation selection, and population scope to the need for stronger future studies.
↳ Introduction; Limitations §3.1
Limitations
Methods permits multiple usage labels per relationship description, but the reproduced prompt permits only one. The donation narrative also says higher intensity characterized donors, whereas Table 4 reports a negative intensity coefficient of −0.57.
↳ Methods §4.2.1; Supplementary §§6.3 and 6.11; Table 4
The regressions lack baseline well-being and a strategy addressing dominant common causes of companionship-seeking and distress. The findings cannot distinguish an effect of companionship from selection into it by users with lower well-being.
↳ Methods §4.3; Tables 2–3; Limitations §3.1
Intensity and disclosure interactions are significant only for self-reported primary companionship use. The conclusion drops this qualifier, making the moderation pattern appear more consistent than the results show.
↳ Table 2; Results §§2.4–2.5; Conclusion §3.2
The study's main contribution is combining survey and conversation evidence while distinguishing companionship orientation from overall chatbot intensity. Table 2 supports negative companionship–well-being associations, but its significant intensity and disclosure interactions are confined to the self-reported primary-use measure. Cross-sectional measurement and limited confounding control constrain interpretation despite the measurement-validation and selection-analysis work. Conflicting classification instructions, the reversed donation description, and the unresolved equivalence calculation require clarification, while the broader literature discussion and explicit sampling limitations support the contextual assessment.
Nabu’s assessment, alongside the field’s view.
Are you an author of this paper?
Sound3.1
Confidence mediumThe survey and donated conversations extend evidence on companionship-oriented use through three operationalizations. The advance is incremental over closely related studies and does not resolve temporal direction.
Measurement validation and donation-selection analysis support the execution, but contemporaneous measurement and limited confounding control bind the score at 2.5. The classification-procedure conflict and reversed donation description are surfaced separately through the reliability screen.
The structure and principal association statements are followable. The score includes a 0.4 deduction because the conclusion generalizes moderation beyond the operationalization that supports it.
The paper engages mixed prior findings and explicitly traces design, sampling, and platform limitations into future research needs. It does not deeply reconcile its estimates with the opposing longitudinal and loneliness evidence.
Caveats4 of 4 checks
The standardized-coefficient label is insufficiently explained relative to the raw well-being composite. Ethics approval and consent are declared, while no data, code, preregistration, or conflict-of-interest statement is supplied.
Conflict of interest: not stated · Ethics approval: declared
114 of 120 checkable references verified
132 references in manuscript 12 are books, websites or datasets — counted, but not index-checkable
References not verified automatically are resolved manually within 24 hours.
No retraction or expression of concern found in Crossref (includes Retraction Watch).
Sources: Crossref (incl. Retraction Watch) ✓
Where this paper’s evidence sits on the path from initial observation to real-world use.
The proposed interface cues, distress detection, and referral mechanisms are not evaluated. Real-world conversation data do not establish that implementing those mechanisms would improve outcomes.
AI-generated, human-governed. Something look off? Contact us to request a review.
Keep reading
Other papers Nabu has evaluated blind, starting with the same field.
Richard Wiseman, Caroline Watt, Paul Stevens, et al. · 2003 · Psychology
Results revealed significantly more reports of unusual experiences in areas that had a reputation for being haunted.
SupportedDaniël Lakens · 2010 · Psychology
stick figures waving in synchrony were found to be rated higher on entitativity than stick figures waving in different rhythms
Well-supportedSpires-Jones, Tulloch · 2026 · Computer and information sciences
our data from Scotland show that the more expensive teaspoons were stolen more than the cheaper spoons, contrary to the findings in Melbourne
Tentative