Understanding preferences surrounding artificial intelligence (AI) and human teaching assistants (TAs) helps managers improve AI TAs, effectively deploying AI and human TAs, and providing better services to learners. The literature has explored how AI TAs' characteristics affect students' use intention, neglecting students' comparative behaviors between AI and human TAs, and overlooking the influence of differences between AI and human TAs on student preferences. Based on preference theory, trust theory, and the stimulus-organism-response (SOR) framework, we constructed a mechanism model by which differences between AI and human TAs affect student preferences. We held 26 semi-structured interviews and collected 401 valid questionnaires to validate it. We also examined the influence and moderating effect of social anxiety and problem complexity on student preferences. Differences in response quality and communication ability impacted differences in ability trust; differences in service attitude and psychological safety influenced differences in benevolent trust; differences in response time impacted differences in integrity trust. In turn, differences in trust affected student preferences. Social anxiety positively impacted students' AI TA preferences and negatively moderated the effect of differences in ability trust on students' AI TA preferences. Problem complexity negatively affected students' AI TA preferences and moderated the effects of benevolence and differences in integrity trust on students' AI TA preferences. We propose a theoretical model to clarify the effects of differences between AI and human TAs on student preferences and to identify boundary conditions. Our findings provide new insights into AI TA research and offer suggestions for AI TA developers, managers, human TAs, and learners.