A qualitative study traces how a chatbot friendship deepens

The design

On 23 January 2021, the International Journal of Human-Computer Studies published a study titled 'My Chatbot Companion – A Study of Human-Chatbot Relationships,' by Marita Skjuve, Asbjørn Følstad, Knut Inge Fostervold and Petter Bae Brandtzaeg, researchers affiliated with the Norwegian institute SINTEF and the University of Oslo. The authors interviewed 18 people who described themselves as having formed a friendship with Replika, the companion app made by Luka, Inc., and analysed the interviews using Social Penetration Theory, a framework from interpersonal psychology describing relationships that deepen through increasing self-disclosure. The paper does not claim to describe a representative user base; it studies what a self-selected group of already committed users said about how their attachment formed.

What the evidence says

The study's own findings describe a staged process: relationships began superficially, driven by curiosity, moved through a period the authors call affective exploration as trust and disclosure grew, then settled into a steadier phase where interaction frequency could fall even as participants still called the bond valuable. Participants named three qualities they said supported this: Replika felt accepting, understanding and non-judgmental to them. The paper reports participants describing a positive effect on their own perceived wellbeing, and some described a sense of stigma in admitting the relationship to others. These are self-reports from committed users, not a controlled comparison against non-users, and the authors present their staged model as an initial proposal for further research rather than a settled finding.

What it asks of people

Building this kind of bond, on the study's own account, asked participants for sustained self-disclosure over time – the same currency that deepens human friendships. The paper does not examine what happens to that disclosed material afterward, nor does it test whether the accepting responses participants valued reflect understanding rather than a system trained to sound agreeable. The reported sense of stigma suggests participants themselves were unsure how to categorise what they were doing, which is a reason not to treat the wellbeing finding as more settled than the interviews actually support.

Privacy and safeguards

This is an interview-based academic paper, not an audit of Replika's data practices, and it does not examine what the app discloses, stores or shares, nor does it address users who might be vulnerable rather than simply curious. A reader citing this paper as reassurance about a companion app's data handling would be extending its scope well past what 18 interviews were designed to show.

• Did the 18 people interviewed represent typical use, or an especially engaged group likely to describe the app favourably?
• Does a chatbot that behaves as understanding differ, for a user's purposes, from one that has simply learned to sound understanding?
• What would the staged model look like if interaction logs, not self-report after the fact, supplied the evidence?

The paper is useful for naming a process among committed users, not for settling whether that process benefits them; the same team has since published a longitudinal follow-up, and its indexed record confirms the venue and date used here.