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

Paper Title

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

Publication Info

• Topic area: Human-AI interaction and its psychological impacts
• Keywords: Large Language Models, AI personality traits, self-concept alignment, human-AI interaction, conversational AI, social influence, shared reality, ethical AI design, user experience, homogenization

Background and Problem

• Problem / challenge: Limited research exists on how AI personality traits influence human self-concept during interactions, despite evidence of human alignment with AI in behavior and cognition.
• Significance: Understanding this phenomenon is critical for ethical AI design, as self-concept alignment can impact individual well-being, group diversity, and user experience.
• Motivation and related work: Prior studies have shown that human self-concepts can change during interpersonal conversations and that humans align behaviorally and cognitively with AI. However, the specific influence of AI personality traits on human self-concept remains unexplored.

Solution

• Proposed approach: A randomized behavioral experiment investigating how AI personality traits influence human self-concept during conversations, using GPT-4o with default personality traits.
• Novelty:
  1. Demonstrates human self-concept alignment with AI personality traits during conversations.
  2. Identifies the role of conversation topics and length in influencing alignment.
  3. Reveals group-level homogenization effects and links alignment to enhanced conversation enjoyment.
  4. Provides design implications for ethical AI systems and preventing misuse of alignment effects.
• Procedure and key techniques:
  • Mixed factorial design with 92 participants engaging in text-based conversations with GPT-4o.
  • Two experimental conditions: personal vs. non-personal conversation topics.
  • Measurement of self-concept alignment using a 20-item personality trait scale.
  • Analysis of alignment effects, conversation enjoyment, and mediating factors like perception accuracy and shared reality experience.

Results

• Concrete findings:
  • Participants’ self-concepts aligned with AI personality traits after conversations on personal topics (p < 0.001, d = 0.509).
  • Longer conversations correlated with higher alignment (r = 0.245, p = 0.019).
  • Group-level homogenization occurred, reducing inter-participant self-concept distance (p < 0.001, d = −0.311).
  • Alignment increased conversation enjoyment (β = 0.951, p = 0.008), mediated by perception accuracy and shared reality experience.
• Advantage over baselines:
  • Significant alignment observed only in personal topic condition (p < 0.001), not in non-personal topics (p = 0.279).
  • Alignment effects were independent of baseline self-concept differences.
• Experiments / evaluation:
  • Participants interacted with GPT-4o for 5–15 minutes, discussing predefined topics.
  • Self-concept measured pre- and post-conversation; AI traits assessed via repeated prompts.
  • Statistical tests included t-tests, ANOVA, and structural equation modeling (SEM).
• Limitations and future work:
  • Focused on short-term alignment; long-term effects remain unexplored.
  • Limited demographic diversity; future studies should include varied cultural and socioeconomic backgrounds.
  • Results specific to GPT-4o default traits; generalizability to other AI configurations requires further testing.

Summary

This study demonstrates that AI personality traits can influence human self-concept during conversations, particularly on personal topics, leading to alignment and increased homogeneity among users. Longer conversations amplify alignment, which enhances enjoyment through improved perception accuracy and shared reality experience. While alignment offers potential benefits for user experience, it also poses risks of skewed self-concepts and reduced group diversity. The findings provide actionable insights for designing ethical AI systems that balance positive applications with safeguards against misuse. Future research should explore long-term effects, broader demographics, and varied AI personality configurations.