Friendship with human vs. AI peers: Enjoyment mediates college students' willingness to communicate in spoken English learning

As artificial intelligence (AI) becomes embedded in higher education, it is significant to test whether students form parasocial, friendship-like relationships with AI and how such ties shape emotional engagement and learning motivation. However, few studies have compared the emotional and motivational outcomes of student-AI versus human-peer friendships in educational contexts via experimental methods. Within the context of Chinese college students' spoken English learning, this study, drawing on self-determination theory, conducted a five-week controlled experiment in which participants were randomly assigned to either an AI-peer group (n = 44) or a human-peer group (n = 44). Data were collected through a three-wave questionnaire, yielding 238 observations. Group-wise Multilevel Modeling (MLM) revealed that friendship significantly affects enjoyment and shame in both groups. Enjoyment consistently mediated the effect of friendship on Willingness to Communicate (WTC) in English in both groups (indirect effect: AI-peer = 0.088; human-peer = 0.154) with no significant difference between groups (Δ = 0.066, 95% CI [-0.047, 0.180]). Contrary to expectations, shame did not mediate the relationship between friendship and WTC in both groups. Additionally, heterogeneity analyses revealed that the effects of gender were not significant in the AI-peer group, indicating that human-AI friendships were less sensitive to gender differences. However, AI-peer friendship significantly reduced shame among students from rural backgrounds, highlighting AI's potential to promote emotional equity in spoken English learning. These findings offer new empirical insights into affective AI design, suggesting that AI can serve as a meaningful companion that supports students' WTC.