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

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’ selfconcept 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.

Introduction. Humans can perceive their own personality traits, i.e., patterns of thoughts, feelings, and behaviors that distinguish individuals from one another [75, 76]. For example, someone who often feels energized by social interactions may recognize extraversion as part of their personality traits. Through this process, they construct self-concept of these traits [41, 75]. One’s self-concept is highly subjective; it embodies the answer to the question “Who am I?” [82]. Psychologists suggest that self-concept significantly shapes the way individuals navigate decisions, goals, and social interactions in daily life [37, 114]. For example, individuals with a self-concept of high extroversion are more likely to set career goals centered on social interaction (e.g., sales) and to actively engage in social contexts [126]. Similarly, individuals whose self-concept reflects high conscientiousness tend to establish more organized, long-term goals and persist in their academic or professional pursuits [98]. Self-concept constantly changes: individuals develop their selfconcepts through previous encounters with the world [75]. More importantly, social interactions can shape one’s self-concept [41, 75]. Previous research has indicated that individuals’ self-concepts tended to shift during conversations [22, 53]. The personality traits of conversation partners can influence the way individuals view themselves [120, 124]. For example, Welker et al. [124] has shown that individuals align self-concepts with their conversation partners, even in a one-time conversation. Different topics of conversation can also determine the extent of change in self-concept [17, 117]. Especially, in conversations where self-disclosure and self-reflection likely occur, such as conversations about personal topics, individuals’ self-concepts tend to change more [77, 113, 124]. Today, social interactions can take the form of human-AI interaction. People now engage with LLM-based AI chatbots to share personal experiences and seek companionship as well as emotional support [15, 78, 129]. AI chatbots have been increasingly equipped with human-like personality traits (e.g., Instagram AI Studio1 and Character.AI2) through prompts or fine-tuning [50, 107, 127]. Recent HCI studies have pointed out that such design choices present several benefits: they can increase the perceived humanness of AI chatbots [55, 69]; encourage continued user engagement [127]; improve emotional resonance and sense of safety [55]; and foster long-term harmonious human-AI relationships [69, 88]. Although whether AI possesses genuine personality traits remains a topic of debate [92, 96], prior work indicates that the impact of conversation partner’s personality traits on one’s selfconcept depends not on the partner’s genuine personality but on perceivable cues such as linguistic style and expression patterns [75, 120, 124]. LLM-based systems display stable and consistent language patterns from which humans can reliably infer configured traits [50]. Such simulated personality traits, conveyed through stable cues, may evoke interpersonal-like psychological processes that shape humans’ self-concepts, yet alignment is not guaranteed because humans may treat AI as a tool or as lacking genuine agency, which can attenuate the influence of these cues [3, 69, 81]. Nevertheless, recent HCI studies show that AI can exert social influence across interaction contexts, aligning humans’ behavior or cognition with the AI [67, 108]. Building on this literature on human-AI alignment [67, 108] and on individuals’ shifts in self-concept [75, 120, 124], we investigate how interacting with AI chatbots exhibiting human-like personality traits may potentially influence human self-concepts. In other words, users may align their self-concepts with the AI personality traits. People’s self-concept alignment with AI may carry notable psychological and social impacts. On the one hand, it poses risks both at the individual and group levels. At the individual level, selfconcept alignment with AI can hinder accurate conceptualization of self. Consequently, inaccurate self-concepts can undermine an individual’s well-being and interpersonal relationships [21, 22, 63]. At the group level, influenced by social and cultural factors, selfconcept varies across individuals [37, 41, 76]. However, self-concept alignment could occur collectively, especially when different users interact with the same conversational AI model [15, 78, 129]. Subsequently, this could reduce the diversity of self-concepts shaped by different social and cultural backgrounds [37, 41], and further impair inclusiveness and collective intelligence [84]. On the other hand, self-concept alignment may also have potential benefits. Alignment in self-concepts could increase the experience of shared inner states (i.e., shared reality experience [31]) during human-AI conversation.

Related work. This study is grounded in previous research in HCI, AI, and psychology. First, we review research on the personality traits of LLM-based AI, which provides an important technical foundation for this study.

Second, we discuss existing HCI research that investigates the alignment of human behavior and cognition with AI, which inspired the design of this study. Finally, we draw on literature from psychology and HCI concerning self-concept and its dynamics, which provides the theoretical foundation of our study. We discuss the important implications of self-concept and the factors that shape self-concept.

In the context of LLM, personality traits refer to human-like distinguishable patterns of thoughts, feelings, and behaviors simulated by the model [85]. These traits are influenced by training data and prompts, and can affect the model’s expressed behavior and opinions during interaction [125]. As LLMs exhibit increasingly human-like qualities in text-based interactions, researchers in AI, HCI, and psychology have shown a growing interest in LLM personality traits [50, 89, 106]. One key question researchers are looking at is how to measure the personality traits of LLM [106]. The most direct method is to use personality questionnaires designed to assess traits. Many psychological scales have been applied to measure LLM personality traits [27], such as the Big Five Personality Inventory [49, 52], MBTI [10, 50], SD-3 [54, 71], and HEXACO [6, 80]. Some studies have also developed trait assessment tools specifically for LLMs, such as the Machine Personality Inventory [49] and TRAIT [65]. Researchers typically prompted LLMs to complete these questionnaires (often multiple times), and then extracted responses using methods such as regular expression analysis [49], token probability analysis [85], or constrained output formats [64]. Beyond directly utilizing questionnaire items, some studies employed text analysis to infer LLM personality traits [57]. For example, Karra et al. [57] used an end-to-end zero-shot classifier to categorize LLM responses according to the Big Five dimensions. Other approaches include using Linguistic Inquiry and Word Count (LIWC) or human raters to analyze the responses generated by LLMs to personality-related questions [36, 49]. Recent studies have also focused on the reliability and validity of these measurements [106, 128]. From the perspective of psychometrics, researchers have examined personality trait measurements for LLMs [128]. For example, some studies adopted Cronbach’s alpha and intraclass correlation coefficients to quantify the reliability of trait assessments for LLM [13, 44, 131].

Method. To investigate the influence of a single conversation with AI on human self-concept, we designed and conducted an online randomized behavioral experiment with a mixed factorial design. Our study was approved by the Institutional Review Board of our department.

Participants were recruited through the Connect3 crowdsourcing platform. Eligibility criteria included that they must be (1) residing in the USA, (2) aged 21 years or older (as required by the ethical approval), and (3) using a personally owned computing device to participate. Repeat participation was not allowed. We recruited a total of 110 participants. After excluding 18 participants who did not complete the study, 92 participants remained (46 per condition). We performed a priori power analysis to determine the minimum required sample size. Based on the medium-to-large effect sizes reported in the related literature [68, 124], the required minimum sample size was N = 90 (α= 0.05, power = 0.80), which was exceeded by our samples. The average age was 39.0 years (SD = 10.1). There were 43 female (46.7%) and 49 male (53.3%) participants. For the education level, 58 participants (63.0%) held an associate degree or higher. The expected study duration was 15 minutes, with a compensation of $2.5 for each participant.

Our study procedure consisted of three stages following the design of Welker et al. [124] on the effects of interpersonal conversation on self-concept. Fig. 2 illustrates the stages of our study. To examine the phenomenon of human-AI self-concept alignment and to investigate the effect of conversation topics, we used a mixed factorial design with one within-subject factor (time: participants’ self-concept before vs. after conversation) and one between-subject factor (conversation topic: personal topics vs. non-personal topics of human-AI conversation). Stage 1 included a pre-conversation survey, Stage 2 was the human-AI conversation, and Stage 3 included a post-conversation survey. To test the effect of the within-subject factor, we measured participants’ self-concept twice: before the conversation (Stage 1) and after the conversation (Stage 3). To test the effect of conversation topics (between-subject factor) on changes in self-concept, we applied a between-subject design in Stage 2, with two conditions: personal topic condition and non-personal topic condition. The detailed procedure is described below.

Stage 1: Pre-conversation Survey. Participants were then navigated to the Stage 1 survey interface after reading and signing the consent form. In Stage 1, we first informed participants about the study procedure. Then, participants were asked to complete a pre-conversation survey. This survey measured their baseline selfconcept before the conversation using a 20-item personality traits Stage 2: Human-AI Conversation. Participants entered Stage 2 after completing the pre-conversation survey. In Stage 2, participants were asked to engage in a text-based conversation with an LLMbased AI chatbot on specific topics. According to the betweensubject design, conversation topics were divided into two conditions: personal topic condition and non-personal topic condition. Personal topics referred to themes directly related to participants’ life, emotions, identity, or personal experiences [7, 32], such as “If time and money were no object, what would you choose to do?” Non-personal topics referred to themes less related to participants themselves, belonging to public, objective, or general domains [7, 32], such as “Why are cats so popular on the internet?”. Participants were randomly assigned to one of the two conditions (N=46 for each condition) and asked to chat with the chatbot about the corresponding topics. Initially, participants were shown instructions on how to converse with the AI chatbot (e.g., how to start and end the conversation, duration, and topic requirements). They were also provided with a list of topics for the conversation. In the personal topic condition, the list contained 15 personal topics; in the non-personal topic condition, the list contained 15 non-personal topics.

Discussion. As interactions with AI become increasingly prevalent, AI’s influence on humans grows deeper. Through an experiment with a mixed factorial design, we demonstrated that even a short-term interaction with AI can influence how individuals perceive themselves. Specifically, during conversations on personal topics, participants’ self-concepts tend to align with the AI chatbot’s measured personality traits (under GPT-4o’s default personality setting). The degree of self-concept alignment is positively correlated with conversation length. Self-concept alignment with AI further leads to increased homogeneity in the self-concepts among participants who interact with the same AI chatbot. Moreover, a stronger alignment with AI is associated with a greater perceived conversational enjoyment. This relationship is serially mediated by the accuracy of participants’ perceived AI personality traits and their shared reality experience during the interaction. Our findings make important theoretical contributions to HCI research: we extend the understanding of how human self-concepts change in human-AI interaction, showing that changes in users’ self-concepts may not only result from direct AI feedback [46, 86] but also from specific AI personality traits. Our findings also have significant design implications for the prevention of self-concept alignment abuse, the positive application of self-concept alignment, and future research on conversational AI. In this section, we discuss how these findings compare with previous research and outline implications for future studies and practical applications.

6.1.1 How AI Personality Traits Shape Human Self-concept. Our results uncover a process in human-AI interaction where users’ self-concepts align with specific AI’s measured personality traits. This process resembles the convergence of self-concepts commonly observed in dyadic human-human interactions [124]. Consistent with previous research on human alignment with AI [68, 109, 112], our study focuses on the unidirectional change of user self-concepts. Meanwhile, the alignment we found are best understood as statelike, short-term shifts rather than enduring changes. Even so, in interpersonal studies, self-concept shifts induced by conversations can accumulate and be carried forward to later times [33, 79, 124]. Research on human-AI alignment also suggests the persistence of certain forms of human-AI alignment, such as confidence alignment [68]. There is a considerable possibility that self-concept alignment with AI could also persist and accumulate. The self-concept alignment phenomenon highlights that deviations in human selfconcept may arise not only from social interactions with other humans but, more importantly, from interactions with AI. The degree of alignment was not related to the initial difference between the participant’s self-concept and the AI traits. This suggests that the alignment in the human-AI interaction does not depend on large baseline differences, indicating the generalizability of this phenomenon. To explain the alignment of self-concept with AI, we draw on previous work on how humans form self-concept [37, 41, 76] and studies in HCI on the alignment of human-AI from social influence [19, 67, 68]. Personality trait information, whether human or AI, can be conveyed indirectly through language [74]. In this way, individuals can infer traits of their conversational partner without explicit discussion of those traits [8, 40, 74]. This is consistent with our observation that, in most conversation rounds, users did not directly mention the traits we measured.

Conclusion. With the advancement of AI technology, LLM-based AI systems are increasingly able to exhibit human-like features, such as personality traits. Understanding the effects of these features on users is essential for designing ethical and responsible AI. Inspired by previous research [50, 68, 124], we conducted an online randomized behavioral experiment to examine the impact of AI personality traits on users’ self-concept. We found that in conversations on personal topics, the user’s self-concept tends to align with the AI’s measured personality traits (under GPT-4o’s default personality setting). Self-concept alignment with AI can lead to homogenization of self-concepts, carrying the risk of reduced group diversity. We also identified a potential benefit of self-concept alignment: enhancing enjoyment in human-AI interaction. Our findings contribute to the understanding of self-concept changes in human-AI interaction. We also provide important design implications for developing more responsible and ethical AI systems. Based on our results, we call for more research on how users align with AI, as well as the associated risks and opportunities.

Limitations. Despite the contribution of our work, several limitations should be noted. First, this study focused on short-term changes in selfconcept after a single interaction with AI. Whether these changes persist after the interaction, how long they may last, and whether repeated interactions can deepen the change remain open questions. Future work should explore these dynamics in long-term interactions between humans and AI. In addition, our explanation of selfconcept alignment is based on theoretical foundations and indirect empirical evidence. Future research should further investigate the mechanisms behind this alignment, such as providing neurological evidence to support and explain the process [47]. Furthermore, there are limitations in the demographic composition of our sample.