Exploring the Integrative Heuristic Processing in Human-AI Relationships and Their Impact on Users’ Undersociality: A Perspective of Expectancy Violation and Confirmation

Background: Heuristic processing plays a crucial role as an anchor for understanding human-AI relationships (Chaiken & Maheswaran, 1994; Chaiken & Ledgerwood, 2012). While the Computers are Social Actor (CASA) paradigm suggests individuals tend to perceive chatbots as humanlike heuristically due to social cues (Nass et al., 1994; Brandtzaeg et al., 2022; Xu et al., 2022), the machine heuristic argues individuals will treat AI and human differently (Sundar & Kim, 2019). However, whether from the perspective of CASA or machine heuristics, the exploration of how users perceive AI only illuminates limited facets of this complicated dynamic. Additionally, a onesided understanding of the heuristic process in human-AI relationships will mislead its influences beyond dyadic communication, because human-AI relationships are examined to lead to contradictory dual outcomes such as impacting users’ undersociality (Bickmore & Picard, 2005; Epley et al., 2022). Undersociality is induced by miscalibrated social expectations and asymmetric learning (Epley et al., 2022). While miscalibrated social expectations involve individuals’ inferences regarding how others may respond to their social outreach, asymmetric learning refers to the phenomenon where individuals gain different levels of knowledge or understanding from their social experiences based on the nature of those experiences and the feedback they receive (Epley et al., 2022). Despite existing explorations that highlight the differences between human-human interactions and human-AI interactions, the precise impact of these AI engagements on undersociality remains largely unexplored. Hence, this study aims to address (1) How do users perceive AI in relationships based on the integration of CASA and machine heuristics? (2) How does relational interaction with AI influence users’ undersociality?

Methods: Employing a qualitative methodology, the study conducted in-depth interviews with 26 participants recruited from online AI user communities on platforms like Sina Weibo, Xiaohongshu, and Douban. The purposive sampling criteria included individuals with experience in interacting with AI, those who typically feel anxious about socializing, and participants aged over 18 years. Interview transcripts were analyzed using open coding, followed by axial and selective coding to identify key themes and subthemes.

Findings: This study revealed that users develop their expectations about AI based on different heuristics and train AI to fulfill their expectations, but positive expectancy violations make users more pleasant than successful expectancy confirmation (RQ1). Users predominantly perceive AI chatbots as objective, unbiased, and rational entities. Thus, they instruct their chatbots to provide negative feedback to prevent perceived bias. However, unexpected warmth and empathy from chatbots are perceived positively by users, enhancing their satisfaction and reinforcing social norms within interactions. This study also captures a situation where users passively or actively reject expectation development. Despite understanding AI’s programmable nature, participants find it unnecessary to constantly remind themselves of its machine status, especially during emotionally need-oriented conversations. In terms of influences beyond dyadic communication, this study reveals that expectancy in how users perceive AI can buffer users’ undersociality led by biased social expectations (RQ2). AI provides a non-judgmental platform for users to express themselves freely, alleviating fears of misunderstanding or harsh judgment. This safe space reduces social anxiety and encourages more open self-disclosure, particularly during moments of extreme emotion. Furthermore, AI interactions alleviate concerns about inconveniencing others, allowing users to express emotions without fear of disturbing or affecting others. Moreover, interactions with AI enhance users’ self-efficacy and confidence, promoting greater engagement with human counterparts. Users adopt effective communication strategies observed in AI interactions, improving their social skills and conversational tactics for real-world scenarios. Frequent exchanges with AI also subtly reshape users’ thought processes and conversational styles, contributing to improved personal expression and social competence.

Conclusion: Users perceive AI in relationships based on integrative heuristics, which not only shapes their understanding of AI but also positively influences their social behaviors and cognitive frameworks, reducing feelings of undersociality. The study underscores the need for a comprehensive theoretical framework that integrates CASA, machine heuristics, expectancy violation theory, and confirmation theory to fully capture the nuances of human-AI relationships and their social implications. The results provide valuable insights for fostering balanced and healthy human-AI relationships that support social well-being.