人类对大语言模型的热情和能力感知

he rapid development and application of Large Language Models (LLMs) have significantly enhanced their capabilities, influencing human-machine interactions in profound ways. As LLMs evolve, society is shifting from traditional interpersonal interactions to a multilayered structure integrating human-to-human, human-to-machine, and machine-to-machine interactions. In this context, understanding how humans perceive and evaluate LLMsand whether this follows the Big Two model of warmth and competence in interpersonal perceptionhas become critical. This study examines human perceptions of LLMs through three progressive empirical studies.Participants with prior LLM experience were recruited for the studies. Study 1 comprised two sub-studies: Study 1a (N = 207) used a free-response task, asking participants to describe their impressions of LLMs using at least three words, which were analyzed using the Semi-Automated Dictionary Creation for Analyzing Text to identify key dimensions of perception. Study 1b (N = 219) involved a lexical rating task, in which participants rated the applicability of selected evaluation words to LLMs. Study 2 (N = 178) used a questionnaire, in which participants rated a familiar LLM and provided feedback on their willingness to continue using it and their liking of it. Study 3 (N = 207) employed a questionnaire survey to assess participants ratings of warmth and competence for both humans and LLMs.Study 1 found that humans primarily perceive LLMs through warmth and competence, similar to how they perceive other humans. In general contexts, participants prioritized competence over warmth when evaluating LLMs, showing a significant priority effect (odds ratio = 2.88, z = 9.512, 95% CI [2.32, 3.59], p < 0.001). This contrasts with the typical warmth-priority effect in human-to-human perception. Study 2 investigated the relationship between perceptions of warmth and competence and human attitudes toward LLMs, specifically their emotional (e.g., liking) and functional (e.g., willingness to continue using) attitudes. Results showed that both dimensions positively predicted participants liking and willingness to continue using LLMs. Warmth had a stronger predictive effect on liking (warmth: = 0.41, p < 0.001; competence: = 0.27, p < 0.001), while competence had a stronger predictive effect on willingness to continue using (warmth: = 0.19, p = 0.005; competence: = 0.45, p < 0.001). This outcome suggests that the priority effect of warmth and competence shifts across attitude predictions. Study 3 examined specific LLMs ratings in terms of warmth and competence. Results showed no significant difference in warmth ratings between humans (M = 5.06, SD = 1.09) and LLMs (M = 5.11, SD = 1.23), t(206) = 0.60, p = 0.551. However, LLMs were rated significantly higher on competence (M = 5.16, SD = 1.20) than humans (M = 4.81, SD = 1.23), t(206) = 3.51, p < 0.001, Cohens d = 0.29.This study makes two significant contributions to the field. First, it establishes a preliminary theoretical framework for understanding human perception of LLMs. Second, it offers new insights into human-machine interaction by emphasizing the importance of warmth and competence in shaping user attitudes. The findings have practical implications for AI design and policymaking, providing a framework for improving user acceptance, optimizing LLM design, and promoting responsible human-AI coexistence.