AI Teaching Innovation Behavior Among College Teachers: A Structural Equation Modeling Analysis Based on the AI-TPACK Framework, Teaching Self-Efficacy, Professional Identity, and AI Literacy

Synthesis: Bai and Hsieh (2026) set out to explain why some college teachers build genuinely new teaching practice around generative AI while others only adopt the tools. Drawing on the AI-AI-TPACK competence framework, Social Cognitive Theory, and professional identity theory, they surveyed 898 Chinese university teachers and modeled the paths from AI-related competence to AI teaching innovation behavior. Technical AI knowledge on its own did not predict innovation; what mattered was competence embedded in pedagogy and subject matter, working through AI Literacy and professional identity.

Key Findings

• Isolated technical AI knowledge did not predict innovation. Teachers who understood how GenAI tools work, or who could use them functionally, were no more likely to report generating and implementing new teaching practices than those who did not. Competence anchored in pedagogy and subject matter — knowing how to bring AI into one's discipline and how to align it with instructional strategy — is what showed up as innovation.

• Integrative AI-TPACK worked only indirectly. The overarching integration construct had no direct link to innovative behavior at all. Whatever influence it carried ran entirely through AI Literacy, teaching Self-Efficacy, and professional identity.

• AI literacy was the strongest psychological route. Of the three mediators, AI literacy carried the largest share of the association between competence and innovation, and it also fed the other two: teachers with stronger AI literacy reported both greater instructional confidence and a stronger sense of professional identity.

• Professional identity acts as a motivational mechanism. Teachers who saw AI integration as part of their professional role rather than a threat to it were more likely to innovate, and identity mediated the competence-to-innovation link for every AI-TPACK dimension measured.

• Teaching self-efficacy helped, but only partly. Confidence in managing AI-supported instruction mediated the pathway for most competence dimensions — but not for technical AI knowledge, where knowing the technology did nothing to make teachers feel more capable.

• The identity effect was concentrated among less frequent AI users. Among teachers who used GenAI rarely, professional identity predicted innovative behavior nearly twice as strongly as it did among daily users. The authors treat this as borderline and caution against over-reading it.

Study Design & Method

The study surveyed 898 Chinese university teachers from eight comprehensive universities across eastern, central, and western China, recruited by convenience sampling through an online platform. Because the target population included anyone with active teaching responsibilities, the sample spans teaching assistants and junior staff as well as lecturers and professors, and skews young: most respondents held bachelor's or master's degrees and a large share were early in their careers.

Participants completed established scales covering the seven AI-TPACK knowledge dimensions, teaching self-efficacy, professional identity, AI literacy, and self-reported AI teaching innovation behavior, with all instruments translated and contextually adapted for AI-supported teaching. The authors analyzed the data with partial least squares structural equation modeling, testing a multiple-mediation model and then running a multi-group comparison between teachers who used GenAI daily and those who used it monthly, rarely, or never. They controlled for age, educational level, and professional title, and checked the usual measurement properties — reliability, convergent and discriminant validity, multicollinearity, and common method bias — before interpreting the structural paths.

What this means for practice

• Faculty developers. Replace abstract tool sessions with problem-oriented redesign workshops in which teachers work through goal-setting, tool selection, activity design, and assessment for a real course, since faculty development of that kind, not tool operation, is where innovation showed up — isolated technical AI knowledge had no link to innovative teaching.
• Faculty developers. Make AI Literacy the priority mediator: it carried the largest share of the competence-to-innovation association and also raised instructional confidence and professional identity.
• Administrators. Strengthen professional identity explicitly — frame AI integration as an extension of teaching responsibility and clarify how the Teaching changes in AI-mediated instruction — through peer-sharing platforms and visible AI teaching cases, because identity mediated the competence-to-innovation link on every AI-TPACK dimension and mattered most among teachers who used GenAI rarely.
• Designers. Stand up low-risk pilot classrooms and small innovation projects so integrative knowledge can convert into stable practice, giving less experienced users foundational operational guidance while frequent users get advanced collaboration resources.

Limitations

• Convenience sample and self-selection. Voluntary participation likely attracted teachers already interested in AI and more positively disposed toward innovation, and the sample is drawn from eight Chinese universities only, limiting transferability to other higher education systems; The multi-group finding is borderline and the grouping is coarse. The high/low usage split lumps "several times a month" together with "never," which may both weaken real group differences and, where they appear, rest on only one marginally significant path coefficient.
• Cross-sectional design. The data cannot establish causal ordering; the authors note that teachers who already innovate may simply report higher competence, identity, and efficacy, so the paths may run in both directions.
• Self-reported data only. All focal constructs came from the same respondents at the same time, and the innovation scale retained the wording of a general innovation measure, so some responses may reflect innovative teaching in general rather than AI-specific innovation.
• No institutional-level controls. University policy support, digital infrastructure, AI training provision, and organizational climate were not measured, though they plausibly shape whether competence becomes practice, and respondents were nested within universities.

Citation

Bai, X., & Hsieh, T.-S. (2026). AI Teaching Innovation Behavior Among College Teachers: A Structural Equation Modeling Analysis Based on the AI-TPACK Framework, Teaching Self-Efficacy, Professional Identity, and AI Literacy. Frontiers in Psychology, 17, 1834827.