Teaching in the Age of AI: Exploring Faculty Perceptions and Practices

Synthesis: Stalboerger, Marshall, Manning and Hong interviewed 22 full-time faculty members in one college of education at a large public R1 university in the Southeastern United States in Fall 2024, then re-interviewed 12 of them (54.5% retention rate) in Fall 2025 and compared each person's two accounts. The modal movement was from conceptual exploration to practical integration, from cautious curiosity to conditional confidence, and from monitoring student use to guiding responsible use — but two participants became more critical of AI over the year and one reported no meaningful change. The within-person design is what makes the page worth reading: acceptance can move backward as faculty learn more.

Key Findings

1. Twelve matched pairs carry every change claim. The first round drew 22 full-time faculty from a college of education; 12 completed follow-up interviews, a 54.5% retention rate. Responses from all 22 were kept for baseline context, but the longitudinal comparison rests on the 12 interviewed at both time points.
2. Conceptual interest became concrete use. In 2024 one participant called AI planning "still an unexplored area for me"; a year later the same person used NotebookLM to summarize a module's readings for students. Others described generating case studies, requesting topic briefs, and having students compare hand-written research questions with ChatGPT's on the same topic.
3. Comfort increased, but only under faculty control. Most returning participants described conditional confidence: comfort bounded by the need to set the terms of use. One counseling faculty member was comfortable with classroom use only when the instructor said the class was using AI. P5 reported relaxing somewhat while still feeling "not very knowledgeable."
4. The trend had counter-cases. Two participants became more critical of AI over the year, one remained skeptical while beginning to use it more, and one reported no meaningful change. All three teacher preparation participants in the follow-up sample described greater recognition of AI's ethical implications or a more critical stance by 2025, while participants in other program areas more often reported increased comfort.
5. Efficiency was the one stable benefit. Across both rounds, faculty named time savings and reduced workload more consistently than any other benefit: scenario generation, course planning, discussion-question design, and "mindless tasks" that free cognitive time. The authors read that consistency as evidence that efficiency was the entry point for adoption, not its endpoint.
6. Ethical attention moved from policing to process, without fading. In 2024 academic misconduct dominated, with faculty unsure how to detect use and some reverting to paper and pencil. By 2025 more described transparency agreements and guided student use, while concern about academic integrity, ethical implications, and human oversight remained present at both time points.

What the interviews can and cannot attribute to experience

The design holds the person constant. Most faculty AI research captures one snapshot, so a difference between confident users and cautious non-users can always be explained by who answered. Comparing the same twelve people a year apart removes that confound and replaces it with a smaller claim: these accounts changed, and the authors can show in which direction.

What the design cannot do is assign the change to experience. The authors say so: the technology changed between the two rounds, institutional expectations shifted, and public discussion moved, so a year of interviews cannot separate learning from exposure to a moving target. That is the correct reading of a qualitative longitudinal design, and it matters because the most quotable finding — that comfort rose — is the easiest to over-read as an effect of time on task.

The design also keeps the dissent visible. A survey reporting that most faculty grew more comfortable would leave the counter-cases out; here they sit beside the majority pattern in the same table. For anyone building a change plan, that is the more useful object: a campus-level adoption curve can hide individual reversals.

Why learning more made some faculty more cautious

The study used the Technology Acceptance Model as a sensitizing framework without testing it, and its sharpest finding is directional. Acceptance research usually treats experience as a route to ease of use, and ease of use as a route to acceptance. Here, for a subset of faculty, learning more produced the opposite: P2 moved from "curious, but cautious" to "cautious" with less curiosity and "increasingly concerned," and P3 moved from enthusiasm toward a more critical view citing environmental impact and independent thinking.

The authors interpret this as critical thinking rather than resistance, and what those participants asked for supports the reading. They did not ask for AI to be removed. P3 described building shared agreements with students in which use is permitted but disclosed; a research methods faculty member described teaching ethical use in a guided setting. The stance is not refusal but a demand for justification. That distinction matters for competency work, because a program that reads rising concern as a deficit to be corrected pushes away exactly the faculty who have thought hardest about the tool.

The pattern also clusters by program area, and the paper is careful about it. All three teacher preparation participants in the follow-up sample moved toward greater ethical awareness or a more critical stance, while two participants in music education and counseling (P4 and P5) reported increased comfort. With twelve people the authors call this a cautious signal, not a finding, and they name the confound: teacher preparation faculty may receive different professional development or have different conversations about AI during the year.

Why efficiency is the wrong success metric on its own

Efficiency appearing in both rounds is easy to read as progress, but it is the part of AI use least likely to change teaching. Faculty described saving time on scenarios, grading, and curriculum design in 2024 and the same savings, more embedded in workflow, in 2025. What changed alongside it was the pedagogical conversation: whether students may use AI, under what conditions, and how to make that visible.

That gap matters. An institution measuring success by time saved will record adoption while missing whether instruction changed. The paper's sequence — exposure, experimentation, usefulness, conditional confidence, ethical boundary-setting, demand for guidance — is offered as a working framework rather than a causal model, and the authors are explicit that the sample is small and that the counter-cases show not everyone moves through it. As a design hypothesis, it suggests the first year of faculty AI support should expect efficiency gains before pedagogical ones, and should not mistake the first for the second.

What this means for practice

• Sequence professional development around the observed order. Efficiency is the benefit faculty report first; the ethical and curricular questions arrive after use, not before it. Opening with philosophy loses the room; stopping at time savings leaves the harder questions unaddressed. Pair the efficiency workshop with the integrity conversation rather than running separate tracks.
• Write the policy faculty are asking for. In 2024 participants described a vacuum — one said "I made up my own" rules — and by 2025 they wanted guidance on ethical AI use in research and syllabus language they could share with students. Leaving this to individual instructors produces the variation the same faculty then complain about.
• Treat rising skepticism as an outcome of learning. The counter-cases became more critical as they understood the tools better. Give that stance a legitimate place, because a program built only around adoption reads thoughtful hesitation as failure.
• Do not read comfort as competence. Faculty became more at ease while still describing limited knowledge and persistent concern about fabrication and data privacy. Comfort is a precondition for use, not evidence that use is good.
• Make the terms of use explicit in the course. Conditional confidence depends on the instructor deciding when AI enters an activity. A syllabus statement that says when and why sets student expectations and preserves the control that made faculty willing to engage.

Limitations

• One institution, one college. Data came from faculty in a college of education at a single large public R1 university in the Southeastern United States, so faculty elsewhere may differ. Participation was voluntary, which likely over-represents faculty interested in discussing AI and under-represents the hesitant, skewing results toward more favorable views.
• Attrition limits the longitudinal claim. Of the original 22, 12 completed follow-up interviews (54.5%), and non-participation came from non-response rather than withdrawal, so the reasons are unknown. The authors note that less enthusiastic faculty may have been less likely to return, in which case part of the shift toward comfort reflects who stayed rather than change within individuals. The matched sample was not uniformly positive, but those counter-cases cannot rule out that selection.
• Self-report, no demographics, no reliability coefficient. Demographic information was not collected because of Institutional Review Board constraints, so the sample cannot be described by gender, race, career stage, or prior AI experience beyond academic rank. Practice is self-reported: the study observed no courses, syllabi, or student work, so reported integration may exceed enacted integration. And coding used multiple coders with consensus meetings rather than a formal interrater reliability coefficient, defended as a negotiated-agreement model, though reliability is unquantified.
• No comparison group, and no way to isolate experience. The authors state that shifts cannot be attributed solely to increased experience with AI, since the tools, institutional expectations, and public discourse all moved during the same year. Separating learning from context would need a longer series of interviews, or the same protocol where institutional guidance held steady.

Citation

Stalboerger, K. R., Marshall, D. T., Manning, K. Q., & Hong, H. B. (2026). Teaching in the Age of AI: Exploring Faculty Perceptions and Practices. AI Educ., 2(4), 36.