Companion-oriented AI chatbots are a distinct category of application built on large language models. Unlike tool-oriented AI used to retrieve information or complete tasks, companion AI products such as Xingye, Maoxiang, Zhumengdao, Maopaoya and Feichuan are used mainly for everyday chatting, emotional exchange and companionship-style conversation. With their anthropomorphic style, emotionally responsive replies and round-the-clock availability, they are becoming a new kind of social partner for many young people. The study’s author, Taorong Xiao of the School of Media and Communication at Yunnan Normal University, treated this kind of use as a form of media use and asked whether it is associated with the displacement of offline interaction, rather than framing the phenomenon as addiction or problematic use.

Emerging adulthood, roughly the ages of 18 to 29, is a developmental stage marked by identity exploration, relational instability and self-focus. People in this stage rely heavily on peer relationships to construct a sense of self while their offline social networks are being reorganized, which makes them susceptible to loneliness — and they are also among the primary users of companion AI. Two opposing theoretical predictions frame the question. The media-displacement hypothesis holds that time and psychological resources invested in media substitute for face-to-face interaction, weakening real-world ties and impairing wellbeing. The social-compensation hypothesis holds that media use can strengthen existing relationships through self-disclosure or make up for deficits in real-world interaction. Evidence for compensation comes largely from older media such as instant messaging and social-networking sites, and whether it extends to generative companion AI remained untested.

The survey was conducted online through the platform Wenjuanxing, with eligibility restricted to people aged 18 to 29 who had used a companion AI chatbot for chatting, emotional exchange or companionship in the preceding three months. Of 940 questionnaires collected, 97 were removed after attention-check failures or unusually fast completion times, leaving 843 valid responses, a valid response rate of 89.68 percent. The sample was predominantly female at 55.5 percent, mostly aged 18 to 22 at 53.0 percent, and most participants held a bachelor’s degree. The final moderated mediation analysis was based on 783 complete cases after control variables were included. The study received ethics approval from Yunnan Normal University.

Because no established Chinese scales existed for two of the key constructs, the study used newly drafted preliminary item sets. Companion AI use was measured with seven items, such as agreement with statements like using AI chatbots simply for chatting and companionship beyond looking up information. Self-attributed offline social displacement was also measured with seven items, including statements such as spending time once reserved for friends or family on chatting with AI instead, and reducing offline meetings because of frequent AI chatting. Loneliness was measured with the six-item UCLA Loneliness Scale, and relational-interdependent self-construal — the tendency to define the self through close relationships — was measured with the nine-item Chinese version of the Relational-Interdependent Self-Construal Scale. All four scales showed good internal consistency, with reliability coefficients around 0.81 to 0.87.

The statistical analysis proceeded in two stages. The sample was randomly split in half, with one half used for exploratory factor analysis of the two new item sets and the other, 422 respondents, used for confirmatory factor analysis. The four-factor measurement model fit the validation half-sample well, and discriminant validity checks confirmed that self-attributed displacement and loneliness were distinguishable constructs rather than two names for the same experience: the heterotrait-monotrait ratio between them was 0.447, well below the 0.85 threshold, and a combined one-factor model fit markedly worse than a two-factor model. Common method bias was addressed through procedural remedies such as different response anchors for each scale block, and a single-factor model fit poorly compared with the four-factor model.

The mediation results showed a clear pattern. Companion AI use was positively associated with self-attributed offline social displacement, and self-attributed displacement was in turn positively associated with loneliness. The indirect effect through displacement was statistically significant, and it accounted for roughly one-third of the total effect — approximately 33.4 percent — indicating partial mediation. The direct effect of companion AI use on loneliness also remained significant, meaning displacement is only one of the paths linking companion AI use to loneliness and other, non-mediated associations remain. All analyses controlled for age, education, living arrangement, weekly frequency of offline face-to-face contact, average daily AI-use duration and length of AI-use history, and sensitivity analyses excluding individual covariates produced the same substantive pattern.

The most striking finding concerned who was most affected. Relational-interdependent self-construal moderated the second stage of the mediation: the indirect effect through displacement was significant among individuals with moderate-to-high interdependence but non-significant among those with low interdependence. Simple-slope tests showed the positive association between displacement and loneliness strengthening as interdependence rose, from a slope of 0.080 at one standard deviation below the mean to 0.343 at one standard deviation above. The simple-slope plot revealed two crossing lines with a crossover point: when self-attributed displacement was low, highly interdependent individuals reported slightly lower loneliness, consistent with a protective role for close relationships; once displacement was perceived, the pattern reversed, and their loneliness was markedly higher than that of less interdependent peers.

This pattern suggests that relational interdependence operates as a boundary condition rather than a simple amplifier. For people whose sense of self and belonging is anchored in real-world close relationships, perceiving that those relationships have been displaced by AI appears to translate more strongly into loneliness, in line with belongingness theory, which holds that the need to belong is satisfied through frequent positive interaction embedded in stable, mutually caring, ongoing relationships. AI can provide frequent, immediate responses, but it lacks genuine reciprocity and social validation from a real person, and so cannot on balance substitute for what offline interaction supplies. Conversely, for individuals low in relational interdependence, whose self-worth depends less on close relationships, reported displacement posed a weaker threat to belongingness and was not significantly linked to loneliness through the mediation pathway.

The author emphasizes that the findings are preliminary and correlational, and several limitations matter for interpretation. The cross-sectional design cannot establish causal or temporal relations; the direction of the proposed pathway is set by theory rather than tested by the data. The displacement measure explicitly asked respondents to attribute reduced offline interaction to AI use, so the link between AI use and displacement may be partly strengthened by shared item content and should not be read as proof that AI use objectively displaced offline interaction. An external-criterion check supported this reading: companion AI use was weakly associated with lower offline social frequency, but that association vanished once self-attributed displacement was included, while displacement itself remained strongly associated with lower offline frequency. The two new measures also require independent validation across samples, and the data came from a single cultural context, so the moderation reflects within-culture individual variation rather than genuine differences between cultures. Within those limits, the study suggests that loneliness-related risk may be tied less to companion AI use itself than to users’ perception that such use has replaced offline interaction with real people — a perception that may matter most for those who define themselves through their closest relationships.

Subject of Research: Companion AI chatbot use, perceived offline social displacement and loneliness among Chinese emerging adults

Article Title: Companion-oriented AI chatbot use and loneliness among Chinese emerging adults: the mediating role of self-attributed offline social displacement and the moderating role of relational-interdependent self-construal

Article References: Xiao, T. (2026). Companion-oriented AI chatbot use and loneliness among Chinese emerging adults: the mediating role of self-attributed offline social displacement and the moderating role of relational-interdependent self-construal. Frontiers in Psychology, 17, Article 1933530. https://doi.org/10.3389/fpsyg.2026.1933530

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1933530

Keywords: companion AI chatbots, loneliness, emerging adulthood, social displacement, self-construal, media effects, large language models, belongingness, media psychology, cross-sectional survey, China, relational interdependence