Students reported weekly AI use at higher rates than faculty or administrative staff in a survey at one large university specializing in teacher education. Among respondents with valid answers to the weekly-use question, 44.05% of students reported weekly use, compared with 21.60% of faculty and 15.79% of administrative staff. Because the study was cross-sectional, it describes differences and associations at one point in time, not causes.
Overall, 82.41% of respondents reported using at least one AI service. Administrative staff had the highest any-use prevalence, at 85.48%, even though they had the lowest weekly-use rate. The weekly percentages were calculated only from respondents with valid answers to that item.
The final analytical sample comprised 2,121 respondents: 1,809 students, 250 faculty members and 62 administrative staff. Each role answered a separate 75-item questionnaire, with wording adapted to learning, teaching or administrative activities. The analysis used descriptive summaries, Welch group comparisons, pooled ordinary least squares models, reliability and dimensionality checks, and exploratory K-means clustering limited to students.
Usage was only part of the divide
Students also had higher mean scores for current AI-use intensity, meaning their reported level of present use, and for perceived usefulness. Faculty and administrative staff had higher responsible-use norms and academic-integrity concerns.
In the reported pairwise comparisons, students exceeded both faculty and administrative staff in current-use intensity and perceived usefulness. Faculty and administrative staff exceeded students in academic-integrity concerns, while faculty and administrative staff did not significantly differ from each other on the reported comparisons.
The reported Hedges' g values ranged from 0.54 to 0.98. Hedges' g is a standardized measure of the distance between group averages, putting differences on a common scale.
Policy clarity showed a different pattern. Students reported higher perceived institutional policy clarity than faculty, with Hedges' g = 0.42 and an adjusted p value below 0.001. Neither students nor faculty differed significantly from administrative staff in the adjusted comparisons.
Usefulness tracked with trust and current use
To examine how the measures moved together, the researchers used pooled ordinary least squares, or OLS, models. In the trust model, perceived usefulness had the strongest positive standardized association with trust, with a coefficient of 0.402. Policy clarity had a weaker positive association, with a coefficient of 0.223. The model used 1,254 complete cases and explained 40.8% of the variation in trust.
A second pooled model examined current AI-use intensity. Perceived usefulness had the strongest positive association, with a standardized coefficient of 0.322, followed by competence at 0.268, trust at 0.106 and experience at 0.099. Policy clarity was not associated with current use after adjustment, with a coefficient of -0.002. The model used 1,254 complete cases and explained 42.1% of the variation in current use.
These coefficients describe relationships among self-reported measures, not cause and effect. The cross-sectional design cannot show whether usefulness, competence, trust or experience came first.
Students showed different patterns
Among students, longer self-reported AI-use experience was positively associated with perceived usefulness. Mean usefulness was 2.78 among non-users and 4.06 among students reporting more than two years of experience. Across 1,454 students, a Spearman rank correlation, a measure of whether two rankings move together, was 0.327. The association does not establish that experience caused higher usefulness ratings, because self-selection is possible.
Students also showed different combinations of experience, competence, usefulness, trust and control in an exploratory analysis. A student-only K-means analysis, which groups people with similar patterns across those measures, described four clusters among 961 complete-case students. The largest cluster combined high usefulness and trust with moderate experience.
The clusters should be treated as tentative patterns rather than fixed types. The analysis was exploratory, and the optimality and stability of the four-cluster solution were not established.
The comparison has limits
The main indices generally showed acceptable internal consistency, but preliminary measurement diagnostics did not confirm full measurement invariance. That means the scores may not be perfectly comparable across roles. Tucker's congruence coefficient, a measure of how closely measurement patterns match, was 0.979 for students and faculty, 0.967 for faculty and administrative staff, and 0.934 for students and administrative staff. The lower student-administrative-staff figure limits confidence in direct comparisons between those groups.
The document is an arXiv preprint, version 1, dated 25 August 2026. The authors reported no specific grant funding and declared no competing interests. The anonymized datasets, codebook, role-adapted questionnaires and analysis scripts were reported as available from the corresponding author on reasonable request, subject to institutional safeguards.
Paper data and sources
Original title: The AI Adaptation Gap in Higher Education: Students, Faculty, and Administrative Staff
Authors: Yuriy S. Braun, Salavat M. Khafizov
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text