The work aims to explore the aesthetic preferences of users towards AI avatar styles and analyze the differences in different situations. Using the selection experiment method, four styles of AI avatar (2D style, 3D cartoon style, 3D realistic style, human-like style) and three situational factors (education, entertainment, assistant) were used as independent variables. The preference selection of the subjects was used as the dependent variable, and user interviews were conducted to investigate the impact mechanism of image style and situational factors on user aesthetic preferences. The experimental results indicated that there were significant differences in aesthetic preferences for AI visual style among participants, and the preference of the four styles exhibited a "uncanny valley effect" to some extent, with the human-like style being the most preferred, followed by 3D cartoon and 3D realistic styles, and 2D style being the least preferred. This difference was closely related to the triggering of the empathy effect, reflecting the technical aesthetics in the context of the intelligent era. Situational factors could play a moderating role between AI visual style and aesthetic preference. In educational, entertainment, and assistant contexts, the most preferred visual styles of avatars were human-like style, 3D realistic style, and 3D cartoon style, respectively. This moderating effect was closely related to the psychological imagery of participants in specific situational patterns. In conclusion, this study deepens the research on the aesthetic preferences and patterns of AI avatar products among user groups, and can provide empirical theoretical guidance for promoting user research, appearance design, and technological development direction.