<p>This study aims to investigate the self-positivity bias effect of large language models (LLMs) under different prompting conditions, compare differences in their response patterns when simulating AI versus participant roles, and examine the impact of social comparison on the self-evaluation of large models. Using a self-referential paradigm, large language models were required to generate self-referential descriptions in the role of AI and participant, assessing their self-association with positive and negative words. Additionally, different social comparison scenarios (no comparison, upward comparison, and downward comparison) were set up to observe changes in the self-positivity bias of the models under various contexts. The study found that large language models exhibited self-positivity bias when simulating both AI and humans, assigned higher scores to positive trait words and lower scores to negative trait words in self-referential evaluation. However, when simulating humans, the models assigned higher scores for negative trait words. The results of social comparison further revealed that large models simulating AI were not affected by social comparison, while models in human simulation tended to rate positive words as more self-descriptive and negative words as less self-descriptive, especially after no comparison and downward comparison compared to upward comparison, consistent with human social comparison experiences. Large language models can simulate human self-positivity bias, but there are cognitive differences between simulating AI and humans, and the social comparison effect is only observed when simulating humans. These findings provide important insights into the response characteristics of large models in self-referential tasks and their similarities and differences with human psychology.</p>