Artificial intelligence in higher education: Student perceptions, academic integrity and challenges for assessment

Synthesis: Crisol Moya, Romero López and Gámiz Sánchez (2026) surveyed 546 undergraduates at the University of Granada with the five-dimension PEHIA-TA questionnaire and found that academic integrity drew the highest mean of the five dimensions (M = 3.79, SD = 0.746), ahead of required skills (M = 3.64) and AI uses (M = 3.61), with perceived problems last (M = 3.05). Only AI uses (χ² = 16.614, p < .001) and required skills (χ² = 7.822, p = .020) varied significantly by degree program, and both effect sizes were small (ε² = 0.030 and ε² = 0.014), while integrity did not differ at all (p = .595). The distinction worth keeping is that disciplinary difference sits in what students report doing with generative AI and in the skills they believe they need, not in how much they say integrity matters — although the digital-citizenship framework the paper proposes is mapped onto the survey after the fact rather than tested by it.

Key Findings

1. The five means formed a compressed order: integrity, skills, use, benefits, problems. Academic Integrity scored M = 3.79 (SD = 0.746), Required Skills M = 3.64 (SD = 0.809), AI Uses M = 3.61 (SD = 0.906), Perceived Benefits M = 3.49 (SD = 0.723) and Perceived Problems M = 3.05 (SD = 1.110).
2. Reported AI use was treated as established, not emerging. The authors read the 3.61 on AI Uses as evidence that AI is "no longer perceived as an emerging technology but rather as a well-established academic resource."
3. Only two dimensions differed by degree program. Kruskal–Wallis tests returned χ² = 16.614 (p < .001) for AI Uses and χ² = 7.822 (p = .020) for Required Skills. Benefits (χ² = 0.194, p = .907), problems (χ² = 1.562, p = .458) and integrity (χ² = 1.037, p = .595) showed no significant difference.
4. The significant differences were small in magnitude. ε² = 0.030 for AI Uses and ε² = 0.014 for Required Skills, with the remaining dimensions at ε² < 0.003; the authors conclude that degree program "accounts for only a small proportion of the variability" in reported perceptions.
5. The instrument's reported psychometrics were strong but summary-level. The five-factor PEHIA-TA carried KMO = 0.909, RMSEA = 0.0545, CFI = 0.959 and Cronbach's α = 0.902, with no per-dimension reliability or item count printed.
6. The framework is a proposal, not a finding. The AI-mediated University Digital Citizenship Framework maps the five measured dimensions onto four conceptual ones, and the authors describe the correspondence as "interpretative rather than psychometric."

What a five-dimension ranking can and cannot show

The paper's headline is an ordering, and every mean sits between 3.05 and 3.79 — less than a point of separation on a scale whose anchors, item count and per-dimension reliability the article never prints. Its table still assigns "Very High" to 3.79 and "Moderate" to 3.05 without stating the rule behind the labels, so a reader cannot tell whether the gap between integrity and perceived problems is substantively large or an artifact of item wording.

What the ranking does license is narrower and more useful. Students place the competence demand of AI above its perceived usefulness: Required Skills at 3.64 beats Perceived Benefits at 3.49, and both beat Perceived Problems at 3.05. That is a claim about relative salience inside one instrument, consistent with the authors' reading of AI as a resource requiring judgment, not a shortcut. It is not evidence that students behave ethically, and the authors say so twice — the high integrity mean indicates "reported ethical attitudes and perceptions" rather than "direct evidence of their actual academic behaviour." Read against the self-report literature, this is a clean case of an instrument measuring the construct it names while leaving the behavior it implies unobserved.

Why the discipline difference is smaller than the p-values suggest

Two of five dimensions differ significantly by degree program, the kind of result usually summarized as "AI use varies by discipline." The effect sizes undercut that summary. With ε² = 0.030 for AI Uses and ε² = 0.014 for Required Skills, degree program explains a few percent of the variance in each, and the authors themselves warn that the significant differences "should not be interpreted as evidence of substantial differences in students' overall perceptions." With several hundred participants a Kruskal–Wallis test will detect small, orderly differences, so significance here is about detectability, not the size of a disciplinary gap.

The null result on integrity is the more interesting one. Academic Integrity did not differ by program (p = .595), while reported use did, which means the dimension students rate highest is also the one that is most uniform across degree programs. That sits beside, rather than against, the wiki's existing evidence that what counts as cheating varies by discipline: student experience research comparing STEM and humanities students finds markedly different judgments about whether AI use is cheating at all. The two findings are compatible because they measure different things — the importance students attach to integrity versus the boundary they draw around it — and keeping them apart matters for practice. A program with higher reported AI use is not a program with weaker ethical commitment, and an integrity intervention aimed at the heaviest users would be aimed at the wrong variable.

What the design cannot supply is the identity of the groups. The paper reports no post-hoc comparisons, no program-level means and no group sizes, so the significant χ² values establish that some programs differ without saying which. The comparison is a screening result, not a map.

A framework mapped onto the data, not tested against it

The paper's conceptual contribution is the AI-mediated University Digital Citizenship Framework, which aligns AI Uses with Responsible Participation, Required Skills with AI Literacy, Perceived Benefits and Perceived Problems with Critical Awareness, and Academic Integrity with Ethical Responsibility. The mapping is tidy, and the authors are unusually direct that it is a reinterpretation: the correspondence is "interpretative rather than psychometric," the study "did not directly measure the four conceptual dimensions," and validation of the framework is left to future research. Nothing in the survey tests whether the four dimensions relate to one another as the framework claims, so the figure is a lens for reading the descriptive results, not a model of them.

The assessment section inherits the same status. Its recommendations for Assessment — authentic tasks that cannot be reduced to a generic generated response, AI-supported formative feedback under human oversight, process evidence such as drafts and reflective commentary, explicit disclosure of AI use, and consistent institutional guidance — are reasoned from the dimension means and from the literature the authors cite. None of them is tested here: there is no comparison of assessment formats, no outcome measure and no student work. The recommendations converge with what the wiki's assessment pages already argue from studies that did test designs, which makes this paper a student-side warrant for redesign rather than new evidence that redesign works. Its distinctive support is that the population who would have to comply already ranks skill demand above convenience, which is a more favorable starting point for a process-oriented policy than the compliance literature would predict.

What this means for practice

• Target support at use and skills, not at integrity. The only dimensions that varied by program were AI Uses and Required Skills; integrity was uniformly high. Discipline-specific AI literacy support follows from these data; discipline-specific integrity training does not.
• Survey your own cohort before importing the ranking. The five-dimension structure is a usable local diagnostic, but the printed means come from 546 students at one Spanish university and no per-dimension reliability or scale anchors are published for comparison.
• Treat a high integrity mean as climate, not compliance. The authors decline to read 3.79 as evidence of ethical conduct. Treat it as a signal that students will accept an integrity conversation, then verify conduct through work and process evidence.
• Pilot assessment redesign instead of citing this paper as proof. The authentic-task, process-evidence and disclosure recommendations are design arguments here; adopt them as hypotheses with local evaluation, because this study measured perceptions and nothing about assessment formats.
• Do not read small disciplinary differences as a divide. With ε² of 0.030 and 0.014, a program-level AI policy justified by "our students use it differently" rests on a few percent of variance.

Limitations

• The sample is 546 undergraduates at a single institution, the University of Granada, recruited by convenience sampling; the authors name this as a limit on representativeness, and no response rate or program-level breakdown is reported.
• The design is cross-sectional and descriptive, so no causal claim about AI use, skills or integrity is possible; the authors note that longitudinal and experimental designs would be needed.
• All measures are self-reported perceptions; the integrity score cannot stand as evidence of conduct, and the authors call for behavioral indicators, qualitative evidence and assessment artifacts.
• Only Kruskal–Wallis tests and effect sizes are reported for the group comparison: no post-hoc pair comparisons, no adjusted alpha and no group sizes, so the significant results cannot be attributed to named programs.
• The instrument's reporting is thin in the article: no item count, no response scale or anchors, no per-dimension reliability, and no derivation for the "Level of Assessment" labels in the descriptive table.
• The proposed framework was not measured, so its four dimensions and their relationships remain untested; the correspondence to the survey is the authors' interpretation.

Citation

Crisol Moya, E., Romero López, M. A., & Gámiz Sánchez, V. M. (2026). Artificial intelligence in higher education: Student perceptions, academic integrity and challenges for assessment. Italian Journal of Health Education, Sport and Inclusive Didactics, 10(4).