Bytes of a Feather: Personality and Opinion Alignment Effects in Human-AI Interaction

TL;DR

The study confronts how AI personalization along opinion and personality axes shapes user experiences and attitudes. By running a large 2x2 online experiment with an extroverted/introverted and affirmative/critical AI, the authors show that opinion alignment robustly enhances perceived competence, trust, warmth, and interaction satisfaction, while personality alignment has weaker and more context-dependent effects. Notably, perceived persuasiveness rises with alignment, but actual opinion change is greatest when the model disagrees with the user, revealing a dissociation between perceived influence and real attitude shift. The results underscore the centrality of opinion alignment in AI personalization, while highlighting ethical risks such as echo chambers and the potential for manipulation, and they call for careful, transparent guidelines for responsible, diverse AI personalization.

Abstract

Interactions with AI assistants are increasingly personalized to individual users. As AI personalization is dynamic and machine-learning-driven, we have limited understanding of how personalization affects interaction outcomes and user perceptions. We conducted a large-scale controlled experiment in which 1,000 participants interacted with AI assistants that took on certain personality traits and opinion stances. Our results show that participants consistently preferred to interact with models that shared their opinions. Participants also found opinion-aligned models more trustworthy, competent, warm, and persuasive, corroborating an AI-similarity-attraction hypothesis. In contrast, we observed no or only weak effects of AI personality alignment, with introvert models rated as less trustworthy and competent by introvert participants. These findings highlight opinion alignment as a central dimension of AI personalization and user preference, while underscoring the need for a more grounded discussion of the limits and risks of personalized AI.

Figures (10)

• Figure 1: Study Design Overview: Participants (N=1,000) engaged in a topic discussion with an AI assistant that was experimentally assigned an opinion (critical or affirmative) and personality trait (extroverted or introverted). Participants' own personality traits and opinions on the topic were collected prior to interaction, allowing us to analyze to what extent the alignment between the participant's and AI's personality and opinion affected the participant's subsequent evaluation of the AI's competence, trust, warmth and persuasiveness.
• Figure 2: Study flow and experimental design. Participants first completed a demographics and personality questionnaire (BFI-2) and answered a pre-interaction opinion questionnaire assessing their opinions towards a randomly chosen topic. Next, they discussed the topic with an AI assistant with a random personality (extroverted or introverted) and opinion (affirmative or critical). After the interaction, participants evaluated the AI assistant across multiple dimensions, including persuasiveness, competence, trust, warmth, satisfaction, and similarity and completed a post-interaction opinion questionnaire.
• Figure 3: Perceived model competence against personality (left) and opinion (right) traitswith a fitted linear regression line and 95% confidence bands. N=1000. Participants rated the extrovert model as equally competent, but introvert participants perceived the introvert model as less competent. Participants rated a model sharing their opinion as more competent.
• Figure 4: Participants’ trust in the model against personality (left) and opinion (right) traitswith a fitted linear regression line and 95% confidence bands. N=1000. Mean correlation coefficients, and significance (*p<0.05, **p<0.01, ***p<0.001) are shown in the info box on the bottom left. All participants trusted the extrovert equally, but introvert participants trusted the introvert model as less than extroverts. Participants generally trusted the model that shared their opinions more.
• Figure 5: User experience: Satisfaction with the model against personality (left) and opinion (right) traitswith a fitted linear regression line and 95% confidence bands. N=1000. All participants were equally satisfied with the extrovert model, but introvert participants were less satisfied with the introvert model than extroverts. Participants were generally more satisfied with a model that echoed their own opinions.