Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 69 of the 69 references it cites were found in a registry or an index; 15 of the 33 sources its claims cite were checked for support.

Emotional Coregulation in Close Relationships with AI Agents: A Survey of ChatGPT and Replika Users

Open access
Pruss, Ethel, Alimardani, Maryam, Struijs, Sascha, Koole, Sander L. · 2025 · Proceedings of the 13th International Conference on Human-Agent Interaction

Abstract

AI agents are increasingly becoming part of daily life, with advances in large language models (LLMs) enabling people to use them not only for practical tasks but also in increasingly social, relationship-oriented roles. This cross-sectional, between-subjects survey (n=48) examined whether relationships with AI agents show emotional coregulation patterns similar to those in human relationships, based on self-reports from users in friendship or romantic relationships with ChatGPT or Replika. The type of AI agent (Replika or ChatGPT) was treated as an independent variable in the analyses. Both agents offer similar conversational abilities, but Replika is presented as a virtually embodied, human-like companion, whereas ChatGPT is a text-based tool not marketed for companionship. We asked both groups of users to rate their experience with the agent in terms of emotional contagion (sharing emotions), emotional counter-regulation (balancing emotions), and affect change, as well as their perceptions of the agent’s self-disclosure (sharing personal or intimate information) and human-likeness. Participants reported emotional contagion and counter-regulation with both agents. Additionally, conversations with both AI agents significantly increased positive affect and reduced negative affect, with emotional counter-regulation emerging as a key factor in affect improvement across both groups. However, human-likeness did not predict emotional coregulation or affect changes. These findings suggest AI agents can support emotional coregulation, with the agent’s self-disclosure playing a role, but the contribution of virtual embodiment remains unclear.

Overview

CoChat Summary

TL;DR

In a cross-sectional survey of 48 users in close (friendship or romantic) relationships with ChatGPT or Replika, participants reported experiencing both emotional contagion and counter-regulation with their agent, and their positive affect rose while negative affect fell across conversations. Counter-regulation predicted affect improvement, self-disclosure predicted coregulation, but perceived human-likeness did not.

Key findings

  • Users reported above-neutral levels of both coregulation components with both agents: emotional contagion (Replika Mdn = 3.40, IQR = 0.80; ChatGPT Mdn = 3.60, IQR = 1.10) and counter-regulation (Replika Mdn = 3.80, IQR = 0.40; ChatGPT Mdn = 4.00, IQR = 0.50), all reliably above the scale midpoint of 3 (contagion: Replika V = 214, p = .011; ChatGPT V = 169.5, p = .031; counter-regulation: Replika V = 274, p < .001; ChatGPT V = 276, p < .001). No significant between-agent differences (contagion W = 289.5, p = .49; counter-regulation W = 227.5, p = .90).
  • Affect improved after conversations: significant main effect of Time, F(1, 138) = 18.96, p < .001, and Time × Affect interaction, F(1, 138) = 23.52, p < .001; the Group × Time interaction was only marginal (p = .084) and the three-way interaction nonsignificant (p = .137).
  • Counter-regulation predicted affect change (ChatGPT: positive affect p = .01, negative affect p = .01; Replika: positive affect p = .04, negative affect p = .09), whereas contagion predicted no affect change in either group (all p > .05).
  • Replika was perceived as significantly more self-disclosing than ChatGPT (W = 372.5, p = .040) but not more human-like (W = 260.5, p = .717). Self-disclosure predicted contagion in both groups (ChatGPT p < .001; Replika p = .03) and counter-regulation in Replika only (p = .01); human-likeness predicted neither coregulation nor affect change (all p > .05).

Methods

  • Cross-sectional, between-subjects online survey conducted June 2025 (ethics approval VCWE-2025-109). Of 70 recruited users, 48 met inclusion criteria (close friendship or romantic relationship with the agent; ≥5 interactions): 25 Replika users and 23 ChatGPT users. Recruitment was via social media and Replika/ChatGPT community forums (Facebook, Reddit, Discord).
  • Measures: a newly developed 10-item emotional coregulation scale (two five-item subscales for contagion and counter-regulation, prevalidated in two pilot studies with N = 120 and N = 119), the PANAS-SF completed twice (recalled affect at the beginning and end of a conversation), the Godspeed anthropomorphism subscale for human-likeness, and an adapted self-disclosure scale.
  • Analysis in R: one-sample Wilcoxon signed-rank tests against the neutral midpoint and Wilcoxon rank-sum tests for group differences; a linear mixed-effects model (Group × Time × Affect, random intercepts for participants, REML, Satterthwaite approximation); and multivariate GLMs relating coregulation, human-likeness, and self-disclosure to affect change.

Limitations

  • Stated by the authors: the between-subjects design with participants interacting with only one agent makes it difficult to disentangle agent design from individual differences between user groups; there was no within-subjects manipulation holding all elements except the avatar constant, and no no-AI control condition to confirm that mood improvements are attributable to the AI conversation. The Replika avatars had minimal, preprogrammed movement that did not adapt to the user or context, limiting their ability to capture nonverbal modality. Long-term effects, persistence of benefits, and risks of emotional dependency remain unclear, and AI agents' unstable identity across service updates poses a further risk.
  • Evident in the text: the ChatGPT counter-regulation subscale showed internal consistency below the conventional norm (α = .52), and the Replika counter-regulation scale was also modest (α = .65); the sample is small (n = 48) and self-selected from agent-specific online communities, and affect was measured retrospectively by recall rather than in the moment.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3765766.3765801

Cites

69 of 69 references verified0 citation errors · 0 not checkable

Statistics

Verification

Claims the paper makes

8 Valid · 1 Not valid · 4 Not checked · 44 Own results and interpretations

30 more claims

Create a free account to see all 57 claims and to verify the 6 references still open.

See all claims — free account
  • It may state two separate claims as one. 15
  • Probably not a substantial claim. 10
  • The quote only partly states this claim. 2

Create a free account to see these claims.

Statistics recomputed

Every reported test result, recomputed from its test statistic and degrees of freedom. A result shows as not adding up only once an independent check has confirmed it.

  • ConsistentF(1, 138) = 18.96reported p < .001 · recomputed p < .001

    “F(1, 138) = 18.96, p < .001”

  • ConsistentF(1, 138) = 23.52reported p < .001 · recomputed p < .001

    “F(1, 138) = 23.52, p < .001”

  • ConsistentF(1, 138) = 3.03reported p = .084 · recomputed p = .084

    “F(1, 138) = 3.03, p = .084”

  • ConsistentF(1, 138) = 2.24reported p = .137 · recomputed p = .137

    “F(1, 138) = 2.24, p = .137”

4 of 4

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Self-check results appear here once an independent check has confirmed them.

Citations

Cites

Show all 60 references

Research works printed but not linked: 4

  1. J. Mihic and M. Novak. 2018. Importance of child emotion regulation for prevention of internalized and externalized problems. Journal of Early Childhood Education Research 7 2 (2018) 235–254.
  2. 10.1007/978-3-030-62056-111 · doi:10.1007/978-3-030-62056-111
  3. B. Rimé. 2007. Interpersonal emotion regulation. In Handbook of Emotion Regulation. Vol. 1. 79.
  4. D. Siemon, T. Strohmann, B. Khosrawi-Rad, T. De Vreede, E. Elshan, and M. Meyer. 2022. Why do we turn to virtual companions? A text mining analysis of Replika reviews. In Proceedings of the Americas’ Conference on Information Systems (AMCIS).

Cited by

See all 5

Metadata

Authors
Pruss, E., Alimardani, M., Struijs, S., & Koole, S. L.
Affiliations
Vrije Universiteit Amsterdam (NL)
Venue
Proceedings of the 13th International Conference on Human-Agent Interaction
Publisher
ACM
Published
2025-11-10
Pages
19-28
OpenAlex
W7117978156
Cited by
4
References
60
Citations by year
2026: 4
FWCI
7.57
Type
conference paper
Language
en
Open access
Licence
cc-by
Locations
  • International Conference on Human-Agent Interaction (HAI) (publishedVersion, cc-by)
  • VU Research Portal (publishedVersion, cc-by)
Topics
  • Social Sciences › Psychology › Social Psychology › Social Robot Interaction and HRI
  • Social Sciences › Psychology › Social Psychology › Attachment and Relationship Dynamics
  • Social Sciences › Business, Management and Accounting › Marketing › AI in Service Interactions
Keywords
AI companions, ChatGPT, Replika, large language models, emotional contagion, self-disclosure, human-likeness, virtual embodiment, human-AI relationship

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-11-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Emotional Coregulation in Close Relationships with AI Agents: A Survey of ChatGPT and Replika Users. https://app.cochat.ai/papers/10.1145/3765766.3765801
The paper
Pruss, E., Alimardani, M., Struijs, S., & Koole, S. L. (2025). Emotional Coregulation in Close Relationships with AI Agents: A Survey of ChatGPT and Replika Users. Proceedings of the 13th International Conference on Human-Agent Interaction. https://doi.org/10.1145/3765766.3765801

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.