AI Literacy in Universities Remains Fragmented and Skewed Toward Tool Skills

A systematic review of 56 peer-reviewed studies published between 2021 and 2025 has found that artificial intelligence literacy in higher education remains conceptually fragmented, with definitions, frameworks and dimensions varying so widely across the literature that no single consensus model has emerged. The review, published in Discover Education, also found that most university initiatives to build AI literacy prioritize operational proficiency and tool use over critical reflection, ethical engagement and longer-term development.

The research team, led by Pingqing Chen with Latha Ravindran, Mansour Amini and Nalini Arumugan, conducted a PRISMA-guided systematic literature review using the ERIC education database and the Web of Science Core Collection. Their search, carried out on 24 October 2025, identified 721 records, of which 685 were retrievable due to an export limitation on ERIC. After deduplication left 666 unique records, title and abstract screening excluded 595, and 71 studies proceeded to full-text retrieval. Ten were inaccessible, and after full-text assessment the final corpus comprised 56 studies: 54 journal articles and two conference papers.

The most common reason for exclusion during screening was a non-higher-education population, accounting for 329 records, followed by non-empirical studies, studies examining AI as an instructional tool rather than AI literacy itself, and specialized clinical or vocational training contexts outside higher education. Methodological quality was appraised using the Mixed Methods Appraisal Tool version 2018, with studies categorized as high, moderate or low quality rather than assigned composite scores. The authors note that formal inter-rater reliability statistics were not computed for data extraction, a limitation they acknowledge.

The corpus was strongly student-centered. Of the 56 studies, 33 focused on students, 9 on pre-service teachers, 7 on in-service teachers, 5 on other higher-education stakeholders, and 2 on mixed or unspecified groups. Publication activity was heavily concentrated in 2024 and 2025, reflecting the surge of interest that followed the launch of ChatGPT in late 2022. Geographically, the literature drew substantially from Asia, North America and Europe, with comparatively limited representation from Africa and South America, a pattern the authors caution does not by itself establish regional differences in policy investment or institutional readiness.

On the conceptual question, the review found that foundational frameworks by Long and Magerko and by Ng and colleagues have shaped the field, but recent work has proliferated into differentiated, context-sensitive models. The review classified the dimensions found in the literature using the affective-behavioral-cognitive-ethical (ABCE) framework and found that cognitive and behavioral dimensions, such as knowledge, application and engagement, were emphasized more frequently than affective and ethical dimensions, such as motivation, self-efficacy and responsible AI use. Frameworks also diverge in how they treat ethics, with some positioning it as one discrete dimension among several and others as a cross-cutting lens applied across all competencies.

Implementation pathways fell into five analytic categories: introductory courses and modules, curriculum integration, quasi-experimental or tool-specific interventions, reflective and mentoring approaches, and professional development programs. The authors report that many initiatives emphasize operational skills such as prompting and verifying AI-generated content, while critical reflection and ethical deliberation receive less consistent attention. Some programs did report notable outcomes: one professional development study reported a 56.0 percent overall AI literacy gain with an 85 percent six-month retention rate, and another documented high satisfaction alongside marked ethical and pedagogical awareness. However, the authors stress that the strength and type of evidence vary across designs and populations, and gains in ethical awareness and long-term retention remain uncertain.

On empirical levels of AI literacy, student-focused studies commonly reported moderate overall AI literacy, though with substantial variation by national context, discipline and prior exposure. Among 1,195 participants, IT students significantly outperformed non-IT students, and male students consistently reported higher readiness than female peers. A study of 664 students identified AI usage frequency and prior training as the two strongest predictors of competency. Among 105 pre-service secondary teachers, overall AI literacy was relatively low, with differences in AI creation and identification linked to prior completion of teaching-methods courses. The authors caution that many studies rely on self-report Likert-scale instruments, which may not map onto actual competence, and that one study assessing students’ actual prompting techniques with ChatGPT revealed technical blind spots not captured by self-report, suggesting that moderate average scores should be read as an upper-bound estimate.

Educator-focused studies, by contrast, indicated readiness concerns in several higher-order, pedagogical and ethical domains, although the authors emphasize that because measures, samples and contexts differ, these findings should not be interpreted as a direct statistical comparison between students and educators. One study of college instructors found overall AI literacy to be weak, with a mean of 3.19 and the greatest deficiencies in the evaluate-and-create and AI ethics dimensions, and specific struggles integrating AI considerations into curriculum design. Another study found that educators’ primary barriers were pedagogical and ethical rather than technical, including uncertainty about addressing AI-assisted academic dishonesty without clear institutional guidelines. Professional development programs for educators also remain skewed toward short formats, from brief intensive workshops to modular training of a few weeks, while multi-semester structured training is comparatively rare relative to offerings for students.

The review also examined relationships among the ABCE dimensions. A quantitative study of 427 university students in Ghana and Nigeria found that affective learning, characterized by intrinsic motivation and self-efficacy, directly strengthened cognitive outcomes, with behavioral engagement and ethical understanding mediating that affective-to-cognitive pathway. The authors note that the similarity of modeled relationships across the Ghanaian and Nigerian samples supports cross-context relevance within those settings but should not be generalized automatically to other populations. Complementary studies found that cognitive AI literacy predicts behavioral engagement, that student engagement enhances AI literacy through anxiety reduction and self-efficacy, and that greater prompting interaction with generative AI correlated with heightened awareness of AI bias.

The authors conclude that the field needs to move beyond narrowly tool-oriented approaches toward theoretically grounded, dimensionally balanced, critically engaged and longitudinally sustained AI-literacy development, with implications for policy, curriculum design and professional development. They acknowledge several limitations: the review covers only 2021 through 2025, restricted inclusion to English-language sources, the literature is heavily skewed toward student populations, and cross-study comparisons are constrained by heterogeneous, largely non-standardized instruments and uneven geographic coverage. Despite these constraints, the synthesis points to a reinforcing relationship among conceptual fragmentation, tool-oriented implementation and dimensional imbalance, in which affective and ethical dimensions, when underdeveloped, may inhibit the transformation of cognitive knowledge into responsible AI engagement.

Subject of Research: A systematic review of artificial intelligence literacy conceptualization, implementation pathways and empirical evidence in higher education from 2021 to 2025.

Article Title: A systematic review examining conceptualizations and implementation pathways alongside empirical evidence of artificial intelligence literacy in higher education from 2021 to 2025

Article References: Chen, P., Ravindran, L., Amini, M., & Arumugan, N. (2026). A systematic review examining conceptualizations and implementation pathways alongside empirical evidence of artificial intelligence literacy in higher education from 2021 to 2025. Discover Education, 5(1), Article 1234. https://doi.org/10.1007/s44217-026-02323-x

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02323-x

Keywords: artificial intelligence literacy, higher education, systematic review, PRISMA, AI ethics, professional development, generative AI, educator readiness, curriculum design, ABCE framework, students, self-report measures