Evolution of Human–AI Relationships in Real-World Interactions

Determine how relationships between users and social chatbots (such as Replika and Character.AI) evolve during real-world interactions, beyond interviews, surveys, and self-reports, in order to characterize the development and dynamics of emotional attachment and relational behavior at scale.

Background

The paper highlights that emotionally responsive AI companions can foster perceived intimacy and attachment, yet most prior work relies on interviews, surveys, or self-reports rather than direct analysis of real conversational behavior. This creates uncertainty about how these relationships actually unfold over time in natural settings.

By assembling a large corpus of user-shared dialogue screenshots and conducting computational analysis, the authors aim to bridge this gap, but they explicitly note that understanding the longitudinal evolution of human–AI relationships in real-world interactions remains an unresolved question that future studies should address comprehensively.

References

However, most existing research is based on interviews, surveys, or self-reports, leaving open questions about how these relationships evolve in real-world interactions.

— Illusions of Intimacy: Emotional Attachment and Emerging Psychological Risks in Human-AI Relationships  (2505.11649 - Chu et al., 16 May 2025) in Introduction (Section 1)

Whether such reciprocation and co-regulation is required in human--robot dyads, is consequential to the formation of a social bond or attachment, and what it may afford our bonding or attachment criteria above, all remain open questions.

— A Procedure for Classifying Attachments and Affective Social Bonds in Human-Robot Dyads  (2609.29063 - Khan et al., 24 Sep 2026) in Section 5, “The Cases of Embodiment and the Reciprocation Loop,” Section 5.2

Repeated exposure could help users recognize and resist personalization. Longitudinal studies across models, prompt strategies, decision domains, and interaction durations could establish whether contextual influence persists and whether stylistic similarity becomes more consequential as relational bonds develop.

— Annie, Are You Okay? How Style- and Context-Based Personalization Shape AI-Assisted Decision-Making  (2609.24644 - Rahman et al., 21 Sep 2026) in Section Discussion, subsection “Limitations and Future Work”

We do not assume that this progression is irreversible, i.e., the stage can also return to a lower level, for instance after protective replies. This assumption is necessary for any intervention to be meaningful, and its validity is an empirical question that our evaluation addresses only in simulation.

— Sampling Safe Futures: Multimodal Trajectory Planning for Personalized Safety in Anthropomorphic AI  (2609.30780 - Picano et al., 25 Sep 2026) in Section 2, System Model

Ten people used one application over four months. They chose that product, so they are not a sample of companion users and still less of people, and the application frames the relationship in a particular way; how much of what is reported here survives a different product is unknown.

— RealCompanion: Benchmarking Human Understanding from Reasoning over Longitudinal Real-World Conversations  (2610.01780 - Behnam et al., 1 Oct 2026) in Appendix, Section “Discussion,” subsection “Limitations”