AI evidence extraction

Ethical Challenges of Artificial Intelligence in Higher Education: A Four-Pillar Student-Activity Framework for Institutional Governance

Radovan Madleňák, Lucia Madleňáková, Viktória Cvacho, Daniel Gachulinec · Education Sciences · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

1/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

This is an AI analysis. Read the peer-reviewed original at the publisher: https://doi.org/10.3390/educsci16040555

Methodology & findings

Study design

Focused, activity-aligned narrative literature review following PRISMA 2020 reporting standards.

Sample

N = 421, 4 groups

Primary method

This is a narrative literature review, not a quantitative meta-analysis. The review documents heterogeneous methodological designs across included studies, including: surveys, design-based studies, policy analyses, field experiments, mixed-methods comparisons, qualitative interviews, structural equation modeling, regression analysis, inverse probability weighting (IPW), quantitative and qualitative analysis, and case studies. Individual studies used diverse statistical approaches but no meta-analytic pooling or statistical synthesis across studies was performed.

Main result

The study shows that contemporary AI reshapes university life by making student work more iterative, interactive, and documented across learning, research, career preparation, and campus participation. "The clearest benefits are speed, availability of feedback, and expanded opportunities to practice complex tasks; the clearest hazards arise when outputs are accepted uncritically, when data practices outstrip educational purposes, or when access to capable tools is uneven." The 2025 evidence base reveals three major developments: first, "assessment and evaluation have emerged as central concerns, with 41 publications examining how AI reshapes grading, feedback, and validity"; second, "human-AI collaborative frameworks have matured, with 24 publications documenting structured approaches to integrating AI into research, projects, and skill development"; third, "the employability of university students and their career development have seen dramatic growth, with 84 publications in 2025."

Reports effect sizes.

Research paradigm

Interpretivist/critical realist

Author conclusions

The authors conclude that "the promise of AI for higher education is realized when institutions design activities and services so that evidence of learning and responsibility is easy to see, data practices are proportionate to educational aims, and access is broad enough to prevent new divides. When implemented this way, students learn more, research remains credible, career signals stay trustworthy, and campus life retains the human qualities that make universities worth defending." They emphasize that "Responsible and equitable AI use is achieved when governance, pedagogy, and services are designed as a single system: policies define the rules and evidence; teaching and supervision make those rules practicable; and services implement proportional, auditable tools."

Risk of bias

Single-database design (Web of Science only) may introduce selection bias by missing studies in Scopus, ERIC, or other databases; Language restriction (English only) excludes non-English publications; Publication window (2022-2025) excludes earlier foundational work on AI ethics; Selective synthesis (3-4 studies per construct) may introduce reporting bias by prioritizing recency over comprehensive coverage; Single analyst conducting screening and coding increases risk of subjective interpretation bias; Topic-field retrieval strategy may miss relevant contributions using different terminology; Gray literature excluded (no dissertations, reports, working papers); Single database (Web of Science) may introduce indexing bias; English-only corpus excludes non-English publications and may introduce language bias; 2022-2025 publication window excludes earlier relevant work; Selective synthesis (3-4 studies per ethics construct) rather than exhaustive coverage may introduce selection bias toward recent, methodologically clear studies; Coding conducted by single analyst may introduce inter-rater reliability concerns; Topic-field retrieval may miss relevant contributions using different terminology; Geographic bias toward China-Hong Kong studies documented in cross-cultural research; Single database (Web of Science) may miss relevant studies in other databases (Scopus, ERIC); English-only language restriction omits non-English publications; 2022-2025 publication window excludes earlier AI ethics debates that may inform current practice; Selective synthesis (3-4 studies per construct) may introduce cherry-picking bias; Topic-field retrieval methodology may miss relevant contributions through indexing variation

Limitations

This narrative review is limited by several methodological constraints: "This narrative review is limited by its reliance on a single database (Web of Science), an English-only corpus, and the 2022-2025 publication window
These choices may omit relevant studies indexed elsewhere (e.g., Scopus, ERIC), published in other languages, or appearing before 2022 but still conceptually important
The selective synthesis (3-4 studies per ethics construct) prioritizes recency and generative AI relevance over exhaustive coverage." Additionally, "screening and coding were conducted by a single analyst," which raises concerns about consistency and bias in study selection and coding.

Open questions raised

Future work should follow cohorts over time to establish longitudinal evidence
Need for comparative studies of AI-supported and conventional designs with common outcome and governance measures
Reusable artifacts and openly available resources needed for replication
Longitudinal studies particularly urgent given rapid evolution of AI capabilities and institutional policies
Cross-cultural comparative research should expand beyond China-Hong Kong studies documented to include diverse global contexts
Causal evaluations using randomized or quasi-experimental designs (building on Greenspan, 2025) needed to rigorously test which policy and pedagogical interventions most effectively balance AI's benefits with integrity, fairness, and skill development

Extracted from: pdfAgreement 60%

Explore related topics

Institutional Policy AI Governance Literature Review