AI-exhibited Personality Traits Can Shape Human Self-concept through Conversations

TL;DR

This study investigates whether AI-exhibited personality traits can shape human self-concept during conversations with an LLM-based chatbot. Using a randomized online design, participants interacted with GPT-4o under default trait settings, with pre- and post-conversation self-concept measures and a between-subjects manipulation of personal versus non-personal topics. Findings show that self-concepts align with the AI's measured traits in personal-topic conversations, with greater alignment as conversation length increases, and that this alignment increases cross-participant homogeneity. A serial mediation through perception accuracy and shared reality fully explains the link between alignment and conversation enjoyment, highlighting design risks (loss of diversity, manipulation potential) and opportunities (enhanced user experience) for AI systems. These results offer actionable guidance for responsible AI design and future research on human-AI self-concept dynamics.

Abstract

Recent Large Language Model (LLM) based AI can exhibit recognizable and measurable personality traits during conversations to improve user experience. However, as human understandings of their personality traits can be affected by their interaction partners' traits, a potential risk is that AI traits may shape and bias users' self-concept of their own traits. To explore the possibility, we conducted a randomized behavioral experiment. Our results indicate that after conversations about personal topics with an LLM-based AI chatbot using GPT-4o default personality traits, users' self-concepts aligned with the AI's measured personality traits. The longer the conversation, the greater the alignment. This alignment led to increased homogeneity in self-concepts among users. We also observed that the degree of self-concept alignment was positively associated with users' conversation enjoyment. Our findings uncover how AI personality traits can shape users' self-concepts through human-AI conversation, highlighting both risks and opportunities. We provide important design implications for developing more responsible and ethical AI systems.

Figures (4)

• Figure 1: Experimental overview. The left side is a flowchart of the experimental procedure. The right side shows the interface of the conversation in Stage 2 (Personal topic condition).
• Figure 2: a) Raincloud plot shows the degree of human-AI alignment significantly fell above zero. The solid line indicates the mean, and the dashed line represents an origin ($x = 0$). b) Interaction plot visualizes participants’ baseline human-AI self-concept distance and post-conversation human-AI self-concept distance across two conditions (personal vs. non-personal). We found a significant interaction effect between within-subject factor and between-subject factor. Points represent mean values, and error bars indicate ± one standard error. *** indicates significant results ($p < .001$).
• Figure 3: Distribution plots of baseline and post-conversation inter-participant self-concept distance among 46 participants in the personal topic condition (1035 pairs). We found the post-conversation self-concept distance was significantly smaller than the baseline. The two solid lines indicate the mean values. *** indicates significant results ($p < .001$).
• Figure 4: Path diagram of the SEM testing H4 and H5 reveals all direct paths from the degree of alignment to conversation enjoyment, through the sequential mediation of perception accuracy and shared reality experience. Solid lines indicate paths with significant direct effects, and dashed lines indicate non-significant paths. Levels of significant results are marked as follows: $p < .05$: *, $p < .01$: **, and $p < .001$: ***.