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

Diyi Yang; Dora Zhao; Jeffrey T. Hancock; Rebecca Anselmetti; Robert Kraut; Yixin Wang; Yutong Zhang

REVIEW

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

Human-Computer Interaction · Submitted 7 Sep 2026

Pith's verdict: Longitudinal engagement patterns are well documented, but the displacement claim rests on concurrent T2 measures and shared item content—causal interpretation is weak.

T0 · AI review

How this review was made: An AI referee (deepseek/deepseek-v4-flash) read the paper on September 10, 2026 and answered a fixed list of questions about it. T0 is the starting level every paper gets. The levels describe how deeply Pith has checked the paper, never the authors.

T0 AI review An AI read the paper and answered a fixed list of questions.
T1 Audit Automatic checks that the references and files the paper relies on exist and match.
T2 Compute The paper's code or calculations re-run with the same results.
T3 Formal statement The claims written out in a formal math language a computer can read.
T4 Proof Those claims proved, and the proof checked by computer.

How Pith checks papers →

· 3 major objections · 5 minor · Ask about this paper · The compiled record · Read the full review →

Challenge this review Comment (0) Read on arXiv ↗

Cite ▾

arXiv ID2609.07243 (version 3)

Pith IDHE3BEBUQ, Pith's permanent record of this paper and its review

CitationDiyi Yang, Dora Zhao, Jeffrey T. Hancock, Rebecca Anselmetti, Robert Kraut, Yixin Wang, Yutong Zhang. "Living with AI Companions: Sustained AI Companionship Predicts Lower Well-Being Through Lower Human Interaction." arXiv:2609.07243 (2026). https://pith.science/paper/HE3BEBUQ

Cite this reviewBibTeX for Pith's review of this paper

Claim paper

Why Pith says this

Highlight any text to comment on it.

Editor's note

read the letter

The paper has real strengths: it is the first multi-wave longitudinal study (12 months) of AI companion engagement and well-being, with thoughtful measurement across three dimensions, Heckman selection correction, and transparent discussion of limitations. The descriptive patterns—that initial interaction intensity predicts later companionship and self-disclosure, and that sustained engagement across all three dimensions is concurrently associated with lower well-being—are useful and new.

But the stress-test concern lands. The paper’s distinctive claim—that AI companionship lowers well-being through displacement of in-person interaction—is identified almost entirely from contemporaneous T2 associations. The mediator (in-person time) was only measured at T2, at the same time as the outcome. The genuinely lagged path (T1 engagement → T2 well-being) is uniformly non-significant. The sensitivity analysis confirms that the effect is driven by the T2 wave. On top of that, the six-item well-being composite includes loneliness, belonging, and social support—items conceptually proximal to the in-person-interaction mediator. So part of the mediator–outcome correlation may be mechanical overlap.

The authors are upfront about the single-item mediator and the measurement change, but they do not flag the concurrent timing and shared content as a load-bearing issue. I think the paper’s causal language should be dialed back: it documents concurrent associations, not evidence that AI companionship displaces human interaction. The longitudinal engagement pathway (sustained engagement → lower well-being) is also concurrent, but at least there the T1→T2 indirect effect is significant, albeit small.

Who is this for? Researchers in human-AI interaction and social computing who want longitudinal descriptive data. The paper is worth serious peer review—it’s a well-executed observational study on an important question—but it needs major revisions to temper the causal claims and to more clearly distinguish concurrent from lagged evidence. I would not cite the displacement result as evidence of a causal mechanism.

Referee report

Major objections (3)

[Section 2.4 / Table 8] The displacement pathway rests on a mediator (time spent interacting with people in person) measured only at T2, concurrently with the outcome (well-being at T2). This means the critical b-path (in-person time → well-being) is a cross-sectional correlation within the T2 wave, adjusted for T1 well-being. Temporal precedence of the mediator over the outcome is never established, because in-person interaction was not measured at T1. The paper's central claim about displacement is therefore identified from contemporaneous associations, not from longitudinal data. The authors should either acknowledge this limitation explicitly in the main text (it is mentioned only indirectly in Section 3.1) or provide a stronger causal identification strategy (e.g., instrumental variables, sensitivity analysis for unmeasured confounders that differ from the existing simulation). As it stands, the displacement claim is overstated relative to the evidence.
[Section 2.3 / Table 7] The direct path from T1 engagement to T2 well-being is non-significant for all three dimensions (β = -0.03 to -0.04, all p > 0.27), and the significant indirect effects all run through T2 engagement, which is measured concurrently with the outcome. This pattern is consistent with an alternative account in which unmeasured third variables at T2 (e.g., declining mental health) drive both increased chatbot engagement and reduced well-being, without any causal effect of AI companionship. The authors should discuss reverse causality and omitted variable bias more explicitly as rival explanations, and consider additional robustness checks (e.g., using T1 engagement as an instrument, or testing whether the T2 engagement–well-being association holds when controlling for T1 engagement and T1 well-being with stronger covariate sets).
[Appendix A.1 / Figure 5] The simulation confirming that the association attenuates toward zero as the simulated usage weight moves toward T1 (w → 1) further supports the concern that the effect is driven by the concurrent T2 measures. This result should be noted in the main text as a limitation of the longitudinal design, not relegated to a supplementary robustness check. The authors should clarify what additional information beyond a strong T1→T2 stability path is provided by the two-wave design, given that the key associations are cross-sectional.

Minor objections (5)

[Section 4.4.5] Human social interaction is measured with a single self-report item about time per day spent talking face-to-face. This is a narrow operationalization that may miss other forms of human interaction (e.g., phone calls, video chats, text messaging) that could also be displaced. The authors should discuss this limitation and consider whether broader measurement would change the displacement estimates.
[Section 4.4.1] Companionship use was measured with a forced-choice question at T1 and a Likert scale at T2, as the authors transparently note. This change limits comparability across waves and may affect the lagged estimates. The free-text-derived measure (Table 6) is a helpful complement, but the two measures show somewhat different cross-lagged patterns (e.g., T1 companionship does not predict T2 interaction intensity in Table 6 but does in Table 1). The authors should discuss this discrepancy more directly.
[Section 3 / Abstract] The abstract states that the findings 'further support the social displacement pathway,' but the evidence is correlational and temporally ambiguous. Consider softening this causal language to reflect the observational design (e.g., 'are consistent with' or 'correlational evidence for').
[Table 1 and Table 6] The Wald χ² values and selection correlations are reported but not explained in the main text. A brief parenthetical explanation of the selection correlation (ρ) would help readers interpret the Heckman correction.
[General] The paper would benefit from a directed acyclic graph (DAG) of the hypothesized causal structure, clarifying the assumed temporal ordering and potential confounders. This would help the reader evaluate the identification strategy.

How Pith checks papers → · Full audit of this paper →

The paper

The paper claims

Sustained AI companionship predicts lower well-being through reduced human interaction.

The reading

This two-wave longitudinal study of 439 Character.AI users over approximately 12 months argues that sustained social engagement with AI companions—across three dimensions of interaction intensity, companionship use, and self-disclosure—is consistently associated with lower psychological well-being. The paper finds that these dimensions show substantial continuity over time and that greater initial engagement predicts deeper subsequent engagement. Crucially, the negative association with well-being is partly explained by a social displacement pathway: users who engage more with AI companions spend less time interacting with other people in person, which in turn predicts lower well-being. These results suggest that AI companions can reshape users' social lives in ways that carry unintended costs, and that safety evaluations should look beyond single-session harms.

What carries the argument

The core mechanism is the social displacement pathway, whereby sustained engagement with AI companions redirects time and attention away from in-person human interaction, consistent with the social displacement hypothesis of Hall and Liu. The longitudinal engagement pathway—whereby earlier engagement shapes later engagement—is also a key component of the argument.

Extended reading notes

Core claim

The central claim is that social engagement with AI companions develops through longitudinal reinforcing dynamics—baseline interaction intensity predicts later intensity, companionship use, and self-disclosure—and that this sustained engagement is consistently associated with lower psychological well-being over time. Further, this association is mediated by reduced in-person human interaction, providing evidence for the social displacement hypothesis in the context of human-AI relationships.

Load-bearing premise

The study assumes that the change in companionship use measurement between waves (forced choice at T1 vs. Likert scale at T2) does not systematically distort the longitudinal estimates.

Editorial extensions

If this is right

AI companion designers should evaluate patterns of sustained engagement over time, not just content safety in individual sessions.
Evaluation of AI companion safety may need to include measures of users' human social interaction and well-being.
The findings support the idea that general-purpose chatbots can take on companion-like roles, widening the scope of systems that need safety oversight.
Policies that limit particularly long or frequent interaction sessions might mitigate displacement effects.
Practitioners should consider the potential second-order effects on users' social networks when their social attention shifts to AI.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If AI companions continue to improve in conversational ability, the displacement effect may grow stronger as they become more compelling substitutes for human interaction.
The study's reliance on a single self-report item for human social interaction may understate the true displacement; behavioral tracking could reveal larger effects.
Users with pre-existing low social support may be most vulnerable, while those with rich social networks may experience less displacement.

Original 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 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.

The compiled record

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 6,240 input tokens.

What the abstract itself states, as probabilities: claims first 0.04 · states limitation 0.09 · human participants 0.98 · controlled comparison 0.13 · randomized assignment 0.02 · states quantitative result 0.82 · compares with prior results 0.15 · mentions code or data availability 0.02.

The abstract, sentence by sentence

arXiv listing abstract v1. Results and the main claim are underlined; show every role.

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

background contribution method result interpretation main claim

What the classifier reports for each sentence. The number is the probability it gave the label.
#	sentence	role	the authors' own work	states the main conclusion	strength
s0	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.	gap 0.65	0.04	0.02	asserted 0.67
s1	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.	method 0.92	0.97	0.02	asserted 0.97
s2	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.	contribution 0.63	0.59	0.04	hedged possibility 0.95
s3	We found that interaction intensity, companionship use, and self-disclosure all showed substantial continuity over time.	result 1.00	0.96	0.04	asserted 1.00
s4	Greater interaction intensity at baseline predicted greater subsequent interaction intensity, companionship use, and self-disclosure.	result 1.00	0.96	0.05	asserted 1.00
s5	Consistent with the longitudinal engagement pathway, sustained social engagement across these dimensions was consistently associated with lower well-being.	result 0.99	0.96	0.51	asserted 0.82
s6	Results further support the social displacement pathway, indicating that these links were mainly explained by lower in-person social interaction.	result 0.53	0.95	0.28	hedged possibility 0.87
s7	These findings highlight the importance of designing AI companions that support human social relationships without displacing them	interpretation 0.70	0.14	0.53	asserted 0.72

What kind of paper the classifier reports this is: empirical study 1.00 · evidence from observation or experiment 0.99.

What was examined, and what was not

sources read

arXiv listing abstract v1

not examined

full text
figures and tables
references and citing sentences
relations to other papers

levels run

P abstract

not run

S sections
T body sentences
C citations
R cross-paper relations

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. Every number in the measurements table is the authors', checked by code. Record as JSON

Figures

**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…

**Figure 2.** Distribution of chatbot relationship descriptors at T1 and T2. (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 at T1 and T2, together with definitions and example descriptions for each category. Free-text relationship descriptors were extracted using GPT-5 mini, embedded using text-embedding-3-small, grouped through…

**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 between T1 engagement and T2 well-being, and panel (c) shows the corresponding indirect associations through T2 engagement. Estimates…

**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 followup 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…

**Figure 5.** Sensitivity analysis for the timing of engagement measurement. Standardized coefficients of the simulated intermediate-time engagement measure X(Ti) predicting follow-up well-being across mixing weights w (Equation 7), where w = 0 corresponds to the follow-up (T2) measure and w = 1 to the baseline (T1) measure. Shaded bands indicate ±1 standard error, and σ denotes the simulated measurement-noise level. A.2.1 Longitudinal development of social chatbot engagement using free-text-derived companionship use As a complementary operationalization of companionship use, we identified…

Figures reproduced from arXiv:2609.07243 by the authors.

References

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Open Benchmark of AI Impact on Humans (ImpactBench): an expert-guided, open platform for the holistic evaluation of AI impact on humans
cs.HC 2026-10 accept novelty 6.0 of 10

ImpactBench is an open, expert-guided benchmark platform that evaluates AI models on 809 behavioral metrics across psychological, physical, and social well-being, finding that models are better at fostering positive o...
"I Felt Very Seen, But Still Very Alone": Longitudinal Trajectories of General-Purpose LLM Use for Socioemotional Support
cs.HC 2026-09 conditional novelty 6.0 of 10

Through interviews, diaries, and focus groups over eight months, the study shows that socioemotional chatbot use follows trajectories shaped by users' care ecologies, with disruptions from system updates and personal ...

Discussion

Disagree with the review, or know something it missed? Leave a comment, highlight any passage above to comment on it, or challenge this review formally.